2000 ST&D Conference Abstracts

Amidst much published confusion about the nature of literacy, a simple distinction between basic literacy and extended literacy is useful. Basic literacy is fundamentally about the acquisition of knowledge of written words, their forms and meanings, and secondarily about the application of language abilities and general knowledge to the printed word. Extended literacy goes beyond these basics to engage students in literacy–based problem solving, writing and argumentation. In considering “the literacy problem”, it is important not to confuse the one with the other. Basic literacy is a prerequisite for extended literacy and different obstacles to competence arise at the two levels. Elsewhere (e.g. Perfetti, 1998), I have focused on the lexical foundations of basic literacy. Here the focus is extended literacy—demanding text environments that require reasoning and problem solving. What processes and skills are required by multiple-text problem-oriented reading? What kinds of learning environments can support the acquisition of these skills?

The study of history offers a distinctive, rich literacy environment. It centers on documents-letters, treaties, notes, official records, diaries--as well as textbooks. Thus, instruction that makes good use of this rich text environment has the potential to support broad-based literacy skills that may extend beyond history classrooms to other cases of text learning, reasoning, and writing. Indeed, in the United States, the National Standards for United States History include thinking skills that "enable students to evaluate evidence, develop comparative and causal analyses, interpret the historical record, and construct sound historical arguments and perspectives on which informed decisions in contemporary life can be based" (National Standards for United States History, 1996, p. 2). The problem is that history instruction in the United States may not do enough to promote these examples of extended literacy: “The typical history classroom is one in which they listen to the teacher explain the day's lesson, use the textbook, and take tests. Occasionally they watch a movie. Sometimes they memorize information or read stories about events and people. They seldom work with other students, use original documents, write term papers, or discuss the significance of what they are studying” (p. 194). This dramatic gap between the literacy potential of history study and the bleak picture of the actual practice of history study (at least in the U.S.) suggests an interesting problem for the application of cognitive science. We–my collaborators, J.F. Rouet, Peter Foltz, Julie Van Dyke, Gareth Gabrys, and especially M. Anne Britt and I–have focused on history learning as a case of extended literacy. In addition to the interesting theoretical issues for text learning that history affords, we hoped to address the gap between history’s potential and practice mentioned above. Our project includes theoretical, empirical, and instructional components, as summarized below: The Documents Model is a framework for describing how readers represent multiple documents that deal with a common problem space (Perfetti, Rouet & Britt, 1999). The framework includes a Documents Space where documents are represented and interlinked by Rhetorical Predicates to form the learner’s Intertext Model. They also are linked to a Situations Model, representing the learner’s understanding of one or more temporal-causal scenarios that are the “story” (or stories) revealed in the texts. Thus, the theoretical formulation represents both causal networks (Trabasso & van den Broek, 1985) and strict text representations, distinct from inferred situations (van Dijk & Kintsch, 1983). It allows a framework for student understanding. For example, many students achieve a well developed situation model with only weak links to one or more texts. A higher standard is seen in students who achieve an articulated intertext model as well as a situation model. Most sophisticated is the student who can entertain two or more situation models linked to specific texts. One assumption of the theory is that learning from multiple documents sharpens the distinction between text and situation for the reader. This implies that multiple documents can promote a more flexible representation of the situation, and, indeed students who write essays based on multiple texts transform the text information more than students with access to a single text (Wiley & Voss, in press).

Our studies have examined sustained learning over time with real texts, experimental manipulations of document characteristics, and other variables. They suggest, among other things, that both younger and college age readers readily acquire basic temporal-causal structures that comprise the narrative heart of many historical texts. And college age readers can construct flexible representations that reflect the integration of information across texts with contrasting perspectives (Perfetti, Britt, & Georgi, 1995). But students at high school and even college levels lack clear perceptions on the uses of documents and the evidential privileges they entail for arguments. We have observed a preference among regular-class American history students for textbooks, not only in use but in trustworthiness. We have observed differences among students, including discipline-background differences (Rouet, Favart, Britt, & Perfetti, 1998) in their appreciation of document privilege. Other studies have examined aspects of the Documents Theory, notably the trade-off between integration of information across documents and the separation of source information.

In light of our studies of students learning history in multiple text environments, we concluded that high school students could benefit from help in the use of multiple documents. We developed the Sourcer’s Apprentice (SA) to assist students in evidence seeking and evaluation skills and to foster an awareness of document type and document privilege (Britt, Perfetti, van Dyke & Gabrys, in press). The SA incorporates instructional design principles and reflects research on obstacles to high school students’ use of documents (Wineburg, 1991). For example, it engages the student directly in text-based problem solving, one of the goals of extended literacy. It displays a bookshelf with documents from which the student can select and take notes while addressing a problem. Students can use the tool as a tutor for direct instruction or as a problem-solving writing environment. Several studies on the effectiveness of the SA as a learning tool have produced positive results in usability and effectiveness. A recent study has suggested more causally coherent essays are written by high school students who have used the SA.

Have we learned something about how to promote a wider realization of the potential of history learning for extending literacy? In a project with many facets, there are many small conclusions: About the privilege of temporal-causal representations as mental models for students; about the obstacles and opportunities this preference for narrativity present to real document study; about the theoretical question of integrating content while separating sources; about the ability of a simple tool, the SA, to support student learning of document use; etc. At a more general level, we believe that history as extended literacy is attainable in a wide range of students who have achieved basic literacy. The special demands on problem-oriented multiple text learning, however, may require some direct and explicit instruction for many students. The Sourcer’s Apprentice is one example of how such instruction might occur.

LSA is a mathematical technique that generates a high-dimensional semantic space from the analysis of a large corpus of written text. The technique was originally developed in the context of information retrieval (Deerwester, Dumais, Furnas, Landauer, & Harshman, 1990) and was adapted for psycholinguistic analyses by Landauer and his colleagues (Landauer & Dumais, 1997; Landauer, Foltz, & Laham, 1998; Landauer, 1999).

LSA must be trained with a large corpus of written text. The raw data LSA are meaningful passages and the set of words each contains. A matrix is constructed whose columns are words and whose rows are documents. The cells of the matrix are the frequencies with which each word occurred in each document. The data upon which the analyses reported below are based consist of a training corpus of about 11 million words (what a typical American school child would read from grade 3 through grade 14), yielding a co-occurrence matrix of more than 92,000 word types and more than 37,000 documents. Note that LSA considers only patterns of word usage; word order, syntax, or rhetorical structure are not taken into account.

Word usage patterns, however, are only the input to LSA which transforms these statistics into something new - a high-dimensional semantic space. LSA does this through dimension reduction. Much of the information in the original pattern of word usage is accidental and inessential. Why did an author choose a particular word in a specific place rather than some other alternative? Why was this particular document included in the corpus rather than some other one? LSA discards all of this excess information and focuses only upon the essential semantic information in the corpus. To tell what is essential and what is distracting information, LSA uses a standard mathematical technique called singular value decomposition, which allows it to select the most important dimensions underlying the original co-occurrence matrix, discarding the rest. The matrix is decomposed into components associated with its singular values, which are ordered according to their importance. The 300 most important components define the semantic space. The dimensionality of the space is chosen empirically: A (roughly) 300-dimensional space usually compares best with human performance.

LSA thus makes the strong psychological claim that word meanings can be represented as vectors in a semantic space of approximately 300 dimensions. But not only word meanings are represented as vectors in this space, documents are similarly represented as well. And new documents - sentences, paragraphs, essays, whole book chapters - can also be represented as vectors in this same space. This is what makes LSA so useful. It allows us to compare arbitrary word and sentence meanings, determine how related or unrelated they are, and what other words or sentences or documents are close to them in the semantic space. A word of caution is necessary here: LSA knows only what it has been taught. If words are used that did not appear in the training corpus, or which are used differently than in the training corpus, LSA, not unlike a person, does not recognize them correctly or at all.

The measure that is used to calculate semantic relatedness is the cosine between two vectors. As a first approximation, readers unfamiliar with this concept may think of cosines as analogous to correlation coefficients. The cosine varies from -1 to +1, +1 denoting identity and 0 denoting unrelatedness. Most cosines between words are positive, though small negative values are common (the average cosine for randomly chosen word pairs is .02, with a standard deviation of .06). The more closely two words are related semantically, the higher their cosine. For instance, the singular and plural forms of a sample of 100 common nouns had a mean cosine of .66, with a standard deviation of .15.

A second measure that is often useful is the length of a vector, which, like the cosine, is defined mathematically. Intuitively, the vector length tells us how much information LSA has about this vector. Thus, the length of sentence vectors is generally greater than the length of word vectors,

and the length of paragraph vectors is even greater. Words that LSA knows a lot about (because they appear frequently in the training corpus, in many different contexts) have greater vector lengths than words LSA does not know well. Thus, horse has a vector length of 2.49, while porch has a vector length of .59. Function words that are used frequently in many different contexts have low vector lengths (the and of have vector lengths of .03 and .06, respectively, and their cosine is .99 - LSA knows nothing about them and cannot tell them apart since they appear in all contexts).

All we can do, however, is compare one vector with another. Inspecting the 300 numbers that compose it tells us little, for the dimensions of the semantic space are not identifiable. The only way we can tell what a given vector means is to find out what other words or sentence vectors are close to it. Thus, we can ask LSA to list the words closest to a given vector in the semantic space. The semantic neighborhood of a word tells us a great deal about the word. Indeed, we shall make considerable use of semantic neighborhoods below.

Often we have some specific expectations about how a vector should be related to particular words or phrases. In such cases it is most informative to compute the cosine between the vector in question and the semantic landmark we have in mind. In most of the examples discussed below when we need to determine what a vector that has been computed really means, it will be compared to such landmarks. Suppose we compute the vectors for horse and porch. To test whether what has been computed is sensible or not, we might compare these vectors to landmarks for which we have clear-cut expectations. For instance, the word gallop should have higher cosine with horse than with porch (the cosines in fact are .75 and .10, respectively), but the word house should have a higher cosine with porch than with horse (the cosines are .08 for horse and .65 for porch). This is not a very powerful test, but it is intuitively compelling and simple. What the particular landmarks are is not terribly important, as long as we have clear shared semantic expectations. Someone else might have chosen race instead of gallop, or door instead of house, or many other similar word pairs, with qualitatively equivalent results.

Readers can make their own computations, or check the ones reported here, by using the web site of the Colorado LSA Research group: http://lsa.colorado.edu. First select the appropriate semantic space and dimensionality. The semantic space used here is the "General Reading through First Year of College" space with 300 dimensions and term-to-term comparisons. To find the semantic neighborhood of horse, one types "horse" into the Nearest-Neighbor-box and chooses "pseudodoc". To find the cosine between horse and gallop, one types "horse" and into one box and "gallop" into the other box of the One-to-Many-Comparison.

LSA has proved to be a powerful tool for the simulation of psycholinguistic phenomena as well as in a number of applications that depend on an effective representation of verbal meaning. Among the former are Landauer and Dumais (1997), who have discussed vocabulary acquisition as the construction of a semantic space, modeled by LSA; Laham's (1997) investigation of the emergence of natural categories from the LSA space; and Foltz, Kintsch and Landauer's (1998) work on textual coherence. To mention just three of the practical applications, there is first, the use of LSA to select instructional texts that are appropriate to a student's level of background knowledge (Wolfe, Schreiner, Rehder, Laham, Foltz, Landauer, & Kintsch, 1998). Second, LSA has been used to provide feedback about their writing to 6th-grade students summarizing science or social science texts (E. Kintsch, Steinhart, Stahl, Matthews, Lamb, & the LSA Research Group, in press). The application of LSA that has aroused the greatest interest is the use of LSA for essay grading. LSA grades the content of certain types of essays as well and as reliably as human professionals (Landauer, Laham, Rehder, & Schreiner, 1997). The human-like performance of LSA in these areas strongly suggests that the way meaning is represented in LSA is closely related to the way humans operate. The present paper describes an LSA-based computational model, which accounts for another aspect of language use, namely, how meaning can be modified contextually in predication. The model is discussed first and illustrated with some simple examples of predication. Then the model is used to simulate several more complex kinds of language processing.

Readers gain a true understanding of a text by building a situation model. The study of situation models thus is central to understand comprehension processes (see the seminal work by van Dijk & Kintsch, 1983). A new development in the study of situation models is the proposal of the Event-Indexing model (Zwaan, Langston, & Graesser, 1995; Zwaan, Magliano, & Graesser, 1995; Zwaan & Radvansky, 1998). This model hypothesizes that readers build situation models by monitoring at least five event dimensions: Protagonist, time, space, causation, and intentionality. By integrating information from these dimensions, readers gradually update their mental representations and build a richly interconnected coherent network or situation model.

Experimental data generally have supported the predictions of the event-indexing model (e.g., Zwaan, Langston, & Graesser, 1995; Zwaan, Magliano, & Graesser, 1995; Zwaan & Radvansky, 1998; Zwaan, Radvansky, Hilliard, & Curiel, 1998). However, converging evidence is necessary to test different aspects of this model. One important feature of a good scientific theory is the display of internal consistency. A rigorous way of validating a model's internal consistency is to examine whether it can be implemented in a computational architecture and still can reasonably account for human performance. Therefore, one purpose of the current study was to examine whether the event-indexing model is detailed and consistent enough to be implemented computationally. A second purpose of this study was to test whether a particular computational model of comprehension, the Landscape model, is able to model online processes of situation models.

The Landscape model is a connectionist model designed to capture the relations between online processes and offline representation of comprehension (Tzeng, 1999; van den Broek, Young, Tzeng, & Linderholm, 1999; van den Broek, Risden, Fletcher, & Thurlow, 1996). The model assumes that readers routinely make causal and referential inferences during the processes of reading and build a coherent representation through a delta learning rule. Furthermore, the Landscape model has a recurrent property, cohort activation, which allows the model to dynamically incorporates readers' activation state from the previous reading cycle into the current mental representation. A cohort is a group of reading elements that are related because they were co-activated at some time during reading. An important property of cohort is that if any part of a cohort is activated the rest of the cohort will be activated as well, to a less degree. The result of reading comprehension by the Landscape model is a representation that emerges from the landscape of fluctuating activations. The landscape itself reflects from various sources of activation. The model has been found to account for a high proportion of readers' recall and to do so across a wide range of texts (Tzeng, 1999; van den Broek, et. al., 1996). Moreover, the model shows potential improvements over the Construction-Integration model (Kintsch, 1988; see Tzeng, 1999).

Accounting for offline recall data only explains the products of reading. The question remains whether the model can predict online comprehension processes. Thus, modeling online reading time data predicted by a theory of situation model becomes a robust challenge for the Landscape model. The event-indexing model was implemented within the Landscape model architecture, with a focus on the idea of continuity between sentences. According to the event-indexing model, sentences vary in their degree of continuity, as determined by the number of dimensions in which consecutive sentences overlap. Specifically, sentence continuities vary along the five dimensions of situation models. The continuity of sentences among these five dimensions will affect readers’ processing speed and, hence, sentence reading times. The event-indexing model predicts that sentences with stronger continuity are easier to process than those with less continuity. As a result, a sentence sharing more dimensions with its previous sentence will have less reading time than another sentence sharing fewer dimensions with its previous one.

Sentence reading time data reported in the second experiment of Zwaan, Radvansky, Hilliard, and Curiel (1998) were adopted for modeling purpose. These reading times were adjusted for the number of syllables. The number of overlap along the five dimensions of situation models was determined for consecutive sentences and used as an indicator of continuity. A sentence can overlap with another by zero to five dimensions therefore the range of continuity was between 0 to 5. These continuity measures were transformed into an input matrix and submitted to the Landscape model. The amount of new activation was computed for each reading cycle by subtracting the total amount of activation of the earlier cycle from the total amount of activation of the current cycle. These activation differences, representing the continuity between sentences, were correlated with sentence reading times by human readers. The resulting correlation coefficients indicate to what extent the Landscape model can model these online reading times.

The observed correlation strengths between activation differences (sentence continuity) and reading times ranged from .46 to .75 with a mean of .57 (all ps < .05). This pattern of results suggests that the Landscape model aptly simulate the online processes predicted by a theory of situation models. It also indicates that the event-indexing model is well specified and it is logically consistent to be captured by a computational model. Thus, this study simultaneously has provided evidence for both the Landscape and the event-indexing theories.

We addressed how readers will detect and process a causal contradiction between the properties of a narrative character and the actions performed by that character. Albrecht and O'Brien (1993) already showed that the contradiction was detected when three sentences intervened between the description of the character's properties and the description of its actions. Thus, Albrecht and O'Brien concluded that readers construct and update a situation model that maps the causal properties of the described situation. Here we looked at what would happen if the contradictory actions were stated immediately after the character's description.

Material. Eighteen experimental texts were written. After an introduction of two sentences, an elaboration part of the text was made of three sentences which expressed specific properties of a main character. These elaboration parts were followed by two sentences which expressed the critical actions performed by the main character. Three coherence conditions were created by varying the character's properties, although the critical actions remained the same. In the Coherence and Incoherence conditions, our knowledge allowed or not to establish a causal relation between the character's properties and the critical actions. In the Neutral condition, there were no causal relation between the properties and the critical actions. Distant and close conditions were also created by adding or not three filler sentences between the elaboration part and the critical actions. In the distant condition, the character's properties were considered as backgrounded by the reading of the three filler sentences and no more present in readers' working memory.

Procedure. Seventy-twelve students read the experimental texts which were randomly mixed with filler texts. The reading was made sentence by sentence on a computer screen and at a free speed. In the experimental design, a coherence factor (Coherent, Incoherent, Neutral) varied within subjects and a distance factor (Short, Long) varied between subjects.

The reading times for the two critical action sentences are presented in Table 1. Globally, the reading times were significantly shorter for the first sentences than for the second sentences (at the critical threshold of p < .05 as for the following results). Separate analyses of variance (ANOVA) were performed for each of these sentences and showed exactly the same pattern. For the first or second sentences, there was a significant interaction between the factors Coherence and Distance. Planned comparisons confirmed that the factor Coherence had no significant effect in the Close condition, but had a significant effect in the Distant condition. In the Distant condition, the reading times were significantly shorter in the Coherent condition than in the Incoherent or Neutral conditions. Our results first confirmed that readers detect the causal discrepancy between the properties previously attributed to a narrative character and an action performed three sentences later by that character. This is attested by readers difficulties to process the contradictory actions and corroborate their use of a situation model to update characters properties. However, our results also showed that when the contradictory action was stated immediately after the character's description, readers had no more apparent difficulties to process the contradictory action. Thus it seems that a situation model first need to be backgrounded by the process of intervening material for a causal contradiction being perceived as such. A plausible explanation is that, when the situation is first introduced in a text and when its model is still activated in reader's working memory, the contradictory actions are interpreted as an interesting paradox (Schank, 1979) or as an initial outcome introduced by the author. Therefore the readers would notice the contradiction, but would adopt a "wait and see strategy", considering that further explanation will be stated in the following of the text. In the distant condition, however, the contradiction was perceived but was now considered as unacceptable. In such case, readers probably considered that a conventional rule of consistency would be violated by the introduction of the contradictory actions at that moment without previous explanation. Therefore readers try to solve themselves the contradiction. They may check their text base representation to find information that they might have neglected while building their situation model or more fully activate their situation model (O'Brien, Rizzella, Albrecht & Hallebran, 1998). They could also try to infer an elaboration that solves the contradiction. Finally, when a situation model has been backgrounded in a fixed state, because of filler information, that state would become a reference point for the reader and thus, without a previous explanation, the content of the model could be updated by coherent completion but not by radical change. On the contrary, the situation model could still be easily modified when it is initially activated in working memory and has not yet been backgrounded. It should however be noticed that this last conclusion is probably limited to characters properties that are described as a usual state. It has been showed to be invalid for unfocused spatial relation (O'Brien & Albrecht, 1992) and for stable knowledge about the world that are incongruent with a character goal ("Swim and sunbathe are practiced in Florida and not in Alaska", Huitema, Dopkins, Klin & Myers, 1993).

Table 1. Mean reading times (ms) for the two critical action sentences in the Coherence and Distance conditions

An example of text: Francoise was living close to many little shops. She preferred to do her shopping in those shops than in the supermarket.

Coherent: Francoise was the best client in a little butchery because she was found of red meat. She specially liked beef and here favorite steaks were the ribs, the sirloin and the roast. She always eat her meat very little cooked and was telling everybody to eat meat underdone.

Incoherent: Since many years Francoise had stopped to go to the butcher because she became a strict vegetarian. She would never eat any beef whatever the pieces. Only the idea of eating ribs, sirloins or roast beef was disgusting her very deeply.

Neutral: Francoise was probably the best client in a little bakery because she was found of their special bread. She was specially found of their bread made with three kinds of cereal: wheat, rye and bran. She always eat her bread as soon as possible when it was still warm.

Filler: Francoise was not living so far from downtown and needed to make a little bit of sport. Evry weekend, Francoise appreciated to walk down town when it was not too cold. It was taking exactly an hour to her to go from her door to the most central building of the town.

Critical 1: That night, Francoise joined a friend in a restaurant and ordered for herself an enormous beef steak.

Critical 2: She eat her meat with pleasure and found it particularly tasteful and juicy.

End: After the diner Francoise proposed a travel plan to her friend. She wanted to cross Africa by foot with her friend. The friend of Francoise couldn't imagine that Francoise was serious.

Current theories of text comprehension assume that readers build situation models of the information described by texts. Situation models constitute the level of text representation associated with "deep" processing, and serve to integrate the information stated in a text with general information supplied by the readers' world knowledge. Many experiments investigated how readers update spatial aspects of their situation model during narrative comprehension, following a paradigm introduced by Morrow and his colleagues (e.g., Morrow, Greenspan, & Bower, 1989). In this paradigm, participants first memorize the spatial layout of a building including its rooms and a number of critical objects, and later read stories about characters moving through the building in pursuit of some goal. These experiments have shown that readers focus their attention on the protagonist of the story. A primary consequence of the readers' focus of attention is that known objects close within the model to the current focus become more activated in memory, so that readers can readily refer to them or retrieveinformation about them to answer questions. The term spatial distance effect was coined for this result: The closer an object is to the protagonist of the story, the more accessible it is in memory.

Despite many replications of the distance effect (e.g., Rinck & Bower, 1995; Rinck, Hdhnel, Bower, & Glowalla, 1997), the type of spatial updating investigated in these studies is actually very limited. Except for the location of the protagonist, all spatial aspects of the situation model remain constant. For instance, neither the rooms of the memorized building nor the objects in them change over the course of the experiment. Therefore, one goal of the two experiments described here was to find out if readers also update more drastic changes of the spatial scenario, namely reconstructions of the previously memorized building. These reconstructions involved the addition and deletion of walls separating rooms of the building. The second goal was to find out if these changes affect the spatial distance effect observed in previous studies. Presumably, the distance effect should be stronger if the protagonist and the probed object are located in different rooms, that is, if a separating wall exists between them (see Rinck et al., 1997). Third, the present experiments aimed at answering these questions with more naturalistic materials than had been previously used. Therefore, a naturalistic long text, similar to a soap opera involving a cast of main characters was read by the participants, and the layout learning procedure was part of reading the text.

Both experiments followed the same basic procedure while participants were reading the text sentence by sentence in a self-paced manner on the computer screen. In the first part of the text, the main characters and the spatial setting were introduced: A group of researchers who had to move into a building which had to be remodeled before it could be used as a research building. In this part, the original remodeling plan was described using both verbal and map information. In the second part of the narrative, participants read that two critical walls had been forgotten during remodeling. Again, this was described by both verbal statements and a map showing all existing and missing walls. After this, a set of episodes described how the protagonists worked and lived in the incomplete building. In the third part, the text described that the building had to be remodeled again to introduce the previously forgotten walls. However, the result of this second remodeling attempt was unsatisfactory again, yielding the four different experimental conditions of "wall status": One critical wall was correct in that it had been put in during the first remodeling and was still there after the second (always existent wall), one wall had been put in correctly first but was accidentally removed during the second remodeling (removed wall), one wall had been missing after the first remodeling and was added correctly during the second (added wall), and one wall was missing both after the first and the second remodeling (always missing wall). In the final part of the text, participants read another set of episodes describing events and actions of the protagonists in the still unsatisfactory building.

As part of these final episodes, three types of experimental sentences were used to explore readers' updating of the spatial changes involved in the two remodeling events. First, "watch sentences" described how the protagonist saw an event in the adjacent room, thereby following the inconsistency paradigm introduced by O'Brien and Albrecht (1992). These sentences made sense only when the wall between the protagonist's location room and the adjacent room was missing. In both experiments, reading times of these sentences were longer if they were inconsistent with the situation model, that is, if the wall existed. If the wall had been removed or had been missing all the time, reading times were significantly shorter, indicating that readers incorporated the existence of the walls into their situation model. Second, "motion sentences" stated that the protagonist walked from his current location room into the adjacent room. Earlier studies have shown that reading times of sentences such as these are correlated with the length of the protagonist's path in the situation model. Indeed, in both experiments described here, reading times were significantly shorter if there was no wall between the location room and the adjacent room, reflecting the fact that the protagonist could walk directly into the adjacent room rather than having to take a route via the hallway. Thus, reading times of these sentences indicate that readers used their knowledge about the existence of walls to represent the protagonist's movements. Third, "anaphoric target sentences" stated that the protagonist thought about or remembered an object located in the location room or in the adjacent room. Given the previously found spatial distance effect, reading times of these sentences should be shorter if they refer to the spatially close objects in the location room compared to the more distant objects in the adjacent room, reflecting the fact that the former are more accessible in memory than the latter. However, no such difference was found in Experiment 1. We reasoned that this might have occurred because - unlike previous experiments - the text did not refer repeatedly to the protagonist's location. Therefore, the text was slightly altered in Experiment 2 to mention the protagonist's location more often. This yielded a weak, but significant distance effect. Thus, it seems that the spatial distance effect might not be so very general after all, and that readers only pay close attention to the protagonist's spatial location if the text or the instructions make it particularly salient.

Morrow, D. G., Greenspan, S. C. & Bower, G H. (1987). Accessibility and situation models in narrative comprehension. Journal of Memory and Language, 26, 165-187.

O'Brien, E. J., & Albrecht, J. E. (1992). Comprehension strategies in the development of a mental model. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 777-784.

Rinck, M. & Bower, G. H. (1995). Anaphora resolution and the focus of attention in situation models. Journal of Memory and Language, 34, 110-131.

Rinck, M., Hdhnel, A., Bower, G. & H., Glowalla, U. (1997). The metrics of spatial situation models. Journal of Experimental Psychology: Learning, Memory, and Cognition, 23, 622-637.

An important process in text comprehension is the updating of mental representations during reading, that is, the process of transforming knowledge already represented in a readers' memory in the light of new, correcting information (van Oostendorp, 1996; Johnson & Seifert, 1999; van Oostendorp & Bonebakker, 1999). For instance, van Oostendorp (1996) had people read a news report about a military intervention in Somalia followed by a second related text. This second text contained transformations of facts mentioned in the first text. Performance on a later inference task showed that little updating occurred. Furthermore, it appeared that changes that were more important to the situation described were less updated. It was suggested that central parts of a situation model may be less easily updated than peripheral parts (cf. Chinn & Brewer, 1993). In this paper we studied whether the central part of a mental representation is less updated when this information is in focus. Focus was manipulated by letting readers compare one text with another along some dimension. The way we examined whether the representation has been updated is based on a method, recently introduced by Ferstl and Kintsch (1999), the Cued Association task. With this task subjects are presented with a word (cue) and they are asked to provide an association to it. The number of times a keyword is given as a response to another keyword is used as a measure of relatedness between these two words. Based on these data an asymmetric proximity matrix is calculated for each subject, which is formally equivalent to a network structure. The keywords represent the nodes and the strengths of links are determined by the response order. In these network structures it is possible to discern concepts that are highly related to each other, and concepts less strongly related. Useful notions for measures of centrality are here the in-degrees and out-degrees of concepts (the summation of all the links going into a node or out of a node, respectively; see Ferstl & Kintsch, 1999). The central part of the mental representation was here defined by the area containing concepts with the highest inner cohesion (that is, having the highest number of links between selected concepts). Subsequently, we examined the similarity of the central parts of the networks of the subjects in the focus group before and after reading the texts and compared that with the similarity in the non-focus group. In line with the preceding study (van Oostendorp, 1996) we hypothesized that the central part of the mental representation is less updated when that information is in focus compared to the updating of this information when it is not in focus. We predicted, thus, that the similarity of the central part of the network structures will be higher in the first case (when it is in focus) than in the latter case (when it is not in focus).

Subjects (40 university students) received two related texts about the Spanish terrorist movement ETA. The second text contained a number of important changes or corrections of facts compared to the first ETA text. Each text could be a typical newspaper article (about 500 words each). Both texts contained two kinds of information. First, information about attacks and violence. Secondly, information about political standpoints and reactions. Focus was introduced by providing the focus group first a text about the Irish terrorist movement IRA, also containing political information. Subjects were requested to compare the political situation in Northern-Ireland and Spain. Subjects in the non-focus (control) group received a neutral text about Spain without any specific reading instruction. Before and after reading the second ETA text the Cued Association task was presented. The subjects read a list of 64 keywords twice and were instructed to write down next to each word the one, two or three words that came first to mind. The list contained 20 words that were important to the domain -based on a pilot study- but were not mentioned in the text, the other 44 keywords came from the text. The first answer to a cue (keyword) got a connection strength of 1, the second one mentioned 1/2, and the third 1/3. Answers that do not come from the list are ignored. We focused on the central parts of the representations, that is, the areas with the highest inner cohesion. The procedure to select these was based on (20) concepts with the highest in-degrees and out-degrees from and to other selected concepts respectively. To assess the amount of updating, that is the degree of change in the networks, we computed the similarity of the networks gathered before and after reading the second ETA text (see Ferstl & Kintsch, 1999). A high degree of similarity corresponds to a low degree of updating. In that case the network-before corresponds to a high degree with the network after reading the second ETA text. Also some textbase-oriented true/false recognition questions were presented at the end of the session.

Because the size of the initial network structure of a reader -influenced by his/her background knowledge- could be a relevant variable to the extent of updating, next to the effect of focus on updating, or even interacting with it, we first determined the network size of each reader on the first cued association task. That is, the sum of the link strengths between concepts in the (first) matrix (see Ferstl & Kintsch, 1999). In other words, the more answers a reader produces consisting of keywords from the (first) list, the greater the network. Based on the mean score we distinguished subjects with an initial small and subjects with an initial large network size. A 2x2 analysis of variance (with focus and initial network size as factors) on the updating scores of the central parts of the networks only showed a significant interaction effect (F(1, 37) = 4.02, p < .05). For subjects with a small network there was less updating in the focus group compared to the non-focus group, while opposed to this, for subjects with a large network, focus showed a significant increase in updating compared to no focus. In a secondary analysis we examined what happened with the political and action concepts that were corrected in the text. We checked whether the embeddedness of political concepts (based on the in-degrees) in relation to the other, action concepts changed more in the focus group than in the non-focus group, since political concepts should be more in focus in the focus group. It appeared that for subjects with a small network the embeddedness of political concepts in other political concepts significantly decreased in the focus group compared to the non-focus group (F(1, 17) = 7.09, p < .05), while opposed to this, for subjects with a large network, focus leads to significantly less embedding of political concepts in action concepts compared to the non-focus group (F(1, 17) = 5.27, p < .05).There were no significant main or interaction effects of focus on the textbase-oriented questions.

Two conclusions can be drawn from these results. First, for subjects with small initial networks there is indeed some resistance to the updating of concepts relevant to the focus of reading. Subjects with initial small networks, reading with an unbiased reading goal, update to a higher degree and they embed the (correcting) political concepts to a greater extent. These results of a low degree of updating with a focus instruction correspond to what was mentioned in the introduction: Changes can be updated less easily with important information than with less important information, that is, information which is not in focus. Second, subjects with a initial large network, probably due to extensive background knowledge, update to a high degree when they read with focus. Furthermore, they show less embedding of political concepts in the action concepts. It seems like they construct two separate knowledge areas, one for political concepts and one for action concepts.

Chinn, C.A. & Brewer, W.F. (1993). The role of anomalous data in knowledge acquisition: A theoretical framework and implications for science instruction. Review of Educational Research, 63(1), 1-49.

Ferstl, E.C. & Kintsch, W. (1999). Learning from text: Structural knowledge assessment in the study of discourse comprehension. In H. van Oostendorp & S.R. Goldman (Eds.), The Construction of Mental Representations During Reading. Mahwah, NJ: Erlbaum.

Johnson, H.M. & Seifert, C.M. (1999). Modifying mental representations: Comprehending corrections. In H. van Oostendorp & S.R. Goldman (Eds.), The Construction of Mental Representations During Reading. Mahwah, NJ: Erlbaum.

van Oostendorp, H. (1996). Updating situation models derived from newspaper articles. Medien psychologie. Zeitschrift fur Individual- and Massenkommunikation, 8, 21-33.

van Oostendorp, H. & Bonebakker, C. (1999). Difficulties in updating mental representations during reading news reports. In H. van Oostendorp & S.R. Goldman (Eds.), The Construction of Mental Representations During Reading. Mahwah, NJ: Erlbaum.

O'Brien et al., (1998) proposed that comprehenders do not update their information according to the here-and-now of the situation, as is claimed by theories of situation-model construction (e.g., Zwaan & Radvansky, 1998). That is, if later information qualifies earlier information in a text, this qualification does not prevent reactivation of the qualified (earlier) information. To test this hypothesis, O'Brien et al., used a clever technique. Their experimental stories came in three versions. In all versions, the beginning of the story described a personality or physical trait of a protagonist (e.g., being a vegetarian or being a junk-food lover). In the Consistent version, a later sentence described an action that was consistent with the trait. For example, a junk-food lover orders a cheeseburger with fries or an athletic young man rescues a little boy. In the inconsistent version, however, the action contradicts what is known about the protagonist, for example, a vegetarian orders a cheeseburger with fries or an octogenarian runs across the street and carries a boy to safety. Finally, and most importantly, in the Qualification condition, the initial description of the protagonist's trait is qualified in a later clause. Thus, the vegetarian is said to occasionally like to eat junk food and the octogenarian is said to be still fit.

O'Brien et al., rightly ascribe the prediction to the updating view that updated information should no longer be re-activated, as it is not part of the current situation. On the other hand, the resonance view proposed by O'Brien et al., predicts that incoming information sends activation to all of long-term memory, such that even information that was supposedly updated can be reactivated. Both views predict that the Inconsistent condition should yield longer reading times than the consistent condition, given that inconsistencies are difficult to integrate with the evolving situation model. However, only the resonance view predicts that the previous information will get reactivated and might thus cause some comprehension problems. However, given that there is not really a contradiction in the Qualification condition, the increase in reading times, relative to the Consistent condition, should be smaller than that produced by the Inconsistent condition. The updating view, on the other hand would claim that the qualified information replaces the previous information in the situation model, so that the Qualified condition should yield a pattern identical to that of the Consistent condition. In five experiments, O'Brien et al., find patterns that appear to be consistent with the resonance view and not with the updating view. In this presentation, we have three goals. We will first argue that some characteristics of O'Brien et al.'s materials do not allow for a proper test of the updating hypothesis. We then report two experiments that support the updating hypothesis, using the same paradigm as O'Brien et al. used, but with different materials. Finally, we will propose an account of our results that integrates the resonance mechanism with our model of situation-model construction (Zwaan & Radvansky, 1998). There are three characteristics of O'Brien et al.'s materials that are likely to have contributed to longer reading times in the Qualified than in the Consistent condition. Therefore, the O'Brien et al., findings cannot be used to properly evaluate the updating hypothesis. The first characteristic has to do with the question of what updating is. We argue that the manipulation used by O'Brien et al., does not involve exhaustive updating, such that the qualified information is still relevant to interpret the current situation. The second characteristic involves a potential confound in the materials which might have been partially responsible for the difference in reading times between the Qualified and Consistent conditions. The third characteristic also introduces a confound in that there was no equal billing for the original information and the qualification. The former was elaborated in much greater detail than the latter, which may have rendered the qualification less effective. In the presentation, we will discuss these characteristics of O'Brien et al's stimulus materials in detail.

We tested the updating and the resonance hypotheses using texts that met the following criteria: (1) the updating and updated information are mutually exclusive, i.e., if one holds, the other does not, (2) the updating information is not inconsistent with world knowledge, (3) the materials are free of a priming bias favoring one of the hypotheses, and (4) the updating information receives equal billing with the updated information. Consider the text in the Appendix. The first sentence states that Bobby's hammer is available for use in the Enablement condition, but not in the Disablement condition. Consequently, the sentence "The hammer was heavy for his young arm." creates a contradiction in the Disablement condition, but not in the Enablement condition. In the Re-enablement condition, the first sentence was identical to that in the Disablement condition, but it was followed by a re-enablement sentence, which, in this case states that Bobby found his hammer. Thus, if readers update their situation models, there should be no contradiction when they read the sentences about Bobby using the hammer.

We conducted two experiments using these materials. In Experiment 1a, the Disablement and Re-enablement texts were identical except for the fact that a re-enablement sentence was inserted in the Re-enablement condition right after the disablement sentence. In Experiment 1b, we tried to control for surface distance by inserting a filler sentence after the first sentence in the Enablement and Disablement conditions. Table 1 shows the results. As can be seen, they clearly support the here-and-now hypothesis. The fact that the instrument had been previously disabled did not elevate the reading times of the Re-enablement condition above those of the Enablement condition, whereas the Disablement sentence did yield reliably longer RTs than the other two conditions. We will provide an integrative discussion of these and O'Brien et al.'s results.

Table 1. Reading times (ms) for the critical sentences in Experiments 1a and 1b (standard deviations in parentheses)

Bobby took out a saw, but then remembered that he lost his hammer (Disable) / Bobby took out a hammer, but then remembered that he lost his saw (Enable) / After some searching, he found it in his father's toolshed (Re-enable) / After some searching, he found his father's tape measure (Filler, Exp.1b) / He also collected the lumber and paint he had bought (S1) / He had already selected an oak tree as the site for the birdhouse (S2) / He marked the boards and cut them out (S3) / Then, Bobby began pounding the boards together (Action) / The hammer was heavy for his young arm (Target sent).

Many studies have demonstrated the effectiveness of tutoring in student learning (Cohen, Kulik, & Kulik, 1982). There are many possible reasons for this benefit, but one of them is undoubtedly the conversational interaction between the tutor and student (Graesser, Person, & Magliano, 1995). Tutorial dialog may allow for the specific remediation of knowledge deficits on the part of the student. In addition, this format engages the student because it is inherently interactive, as opposed to merely reading a text or listening to a lecture.

We have created an intelligent tutoring system, referred to as AutoTutor, that responds to student’s natural language contributions in a manner that mimics the conversational strategies of a normal, unskilled human tutor (Graesser, Wiemer-Hastings, Wiemer-Hastings, Kreuz, & The Tutoring Research Group, 1999). AutoTutor’s contributions are delivered via synthesized speech through an animated talking head with appropriate intonation and facial expressions (Person, Klettke, Link, Kreuz, & the TRG, 1999). The system is designed to help college students learn topics from a course in computer literacy.

Several modules in AutoTutor are used to understand the students’ contributions. These include a word and punctuation segmenter, a part of speech tagger (Olde, Hoeffner, Chipman, Graesser, & the TRG, 1999), and a speech act classifier. In addition, latent semantic analysis (LSA; Landauer, Foltz, & Laham, 1998), a statistical technique used to represent world knowledge, is used by AutoTutor to determine the quality of the students’ contributions (Graesser, Wiemer-Hastings, Wiemer-Hastings, Harter, Person, & the TRG, in press). Included in AutoTutor is a curriculum script that contains primarily concepts, facts, and question-answer pairs related to topics in computer literacy. The dialog move generator component of AutoTutor is composed of 15 fuzzy production rules that determine what, from the curriculum script, AutoTutor will say next. Depending on various aspects of the student’s previous contribution and the preceding dialog in the tutoring session, AutoTutor will respond with one, or a combination of, the following: (1) pump, in which AutoTutor makes a request for the student to provide more information (e.g., "Anything else?"); (2) splice, in which AutoTutor inserts the correct answer to a question in response to a student’s incorrect answer; (3) prompt, in which AutoTutor elicits a specific piece of information from the student (e.g., "ROM is a kind of _____?"); (4) hint, in which AutoTutor presents a fact or leading question, or reformulates the original question; (5) elaboration, in which AutoTutor contributes an important, but overlooked, piece of information; (6) feedback, in which AutoTutor provides an immediate evaluation of the student’s last contribution that is either positive, negative, or neutral; or (7) summary, in which AutoTutor provides a brief synopsis of the answer or solution. For example, if the student’s contributions have been minimal and if little of the topic currently under discussion has been covered, AutoTutor will offer a prompt.

The pedagogical effectiveness of AutoTutor was evaluated in an experiment by having 48 participants interact with the system. The participants were enrolled in an undergraduate computer literacy course and participated for course credit or for pay. The content of AutoTutor’s curriculum script was developed from the textbook for this course. The topics on which participants were tutored in the experiment had been covered previously in their computer literacy course. Participants reread the chapter(s) on one topic and were tutored on a second topic. For a third topic, participants reread the chapter(s) and were also tutored. Participants were then tested on these three topics: Computer hardware, Operating systems, and the Internet.

Two versions of the test were created, each of which consisted of three components: 18 shallow knowledge multiple choice questions, 12 deep reasoning multiple choice questions, and 72 recall questions presented in 18 Cloze passages. The shallow questions were selected from the test bank included with the textbook used in the computer literacy course. The deep questions were designed by the experimenters to get at causal chains (antecedents, consequences), goals, and purposes of procedures (e.g., "How can you best find information on the World Wide Web?"). Finally, the Cloze passages were created by removing key words and phrases from ideal answers to each question covered by AutoTutor in each of the three topics. Each component of the test assessed all three topics covered in the experiment.

On average, the participants interacted with AutoTutor for 63 min (SD = 17), and took an average of 144 turns (SD = 11). The mean performance on the test was 44% (SD = 13), and did not differ according to version (45% vs. 43%). A repeated measures analysis of variance showed that there were significant differences between the conditions (F(2, 94) = 5.68, p < .01). The text only condition resulted in the lowest performance on the test (M = 39%, SD = 15%). In contrast, the test scores on topics for which participants received only tutoring were higher (M = 47%, SD = 17%). The combined effect of tutoring and reading, however, was no better than the tutor only condition (M = 46%, SD = 16%). Paired comparisons showed that performance in the text only condition was significantly lower than the tutor only and the tutor and text conditions.

These results suggest that AutoTutor was an effective pedagogical partner. Specifically, the conditions that included AutoTutor resulted in 8% higher test scores than in the text only condition. Since the content in AutoTutor is functionally the same as the information in the textbook, we can conclude that the interactive nature of tutoring is responsible for this difference. These results are encouraging and suggest that intelligent tutoring systems require careful attention to conversational interaction. Other measures, such as participants‚ computer literacy course grades and self-assessments of computer literacy, will also be analyzed to evaluate the performance of AutoTutor more thoroughly.

Cohen, P. A., Kulik, J. A., & Kulik, C. C. (1982). Educational outcomes of tutoring: A meta-analysis of findings. American Educational Research Journal, 19, 237-248.

Graesser, A. C., Person, N. K., & Magliano, J. P. (1995). Collaborative dialog patterns in naturalistic one-on-one tutoring. Applied Cognitive Psychology, 9, 359-387.

Graesser, A. C., Wiemer-Hastings, P., Wiemer-Hastings, K., Harter, D., Person, N., & The Tutoring Research Group. (in press). Using latent semantic analysis to evaluate the contributions of students in AutoTutor. Interactive Learning Environments.

Graesser, A. C., Wiemer-Hastings, K., Wiemer-Hastings, P., Kreuz, R., and The Tutoring Research Group. (1999). AutoTutor: A simulation of a human tutor. Journal of Cognitive Systems Research, 1, 35-51.

Landauer, T. K., Foltz, P. W., & Laham, D. (1998). An introduction to latent semantic analysis. Discourse Processes, 25, 259-284.

Olde, B. A., Hoeffner, J., Chipman, P., Graesser, A. C., & the Tutoring Research Group (1999). A connectionist model for part of speech tagging. In Proceedings of the 12th International Florida Artificial Intelligence Research Society Conference (pp. 172-176). Menlo Park, CA: AAAI.

Person, N. K., Klettke, B., Link, K., Kreuz, R. J., & the Tutoring Research Group (1999). The integration of affective responses into AutoTutor. In Proceedings of the International Workshop on Affect in Interactions (pp. 167-178). Siena, Italy: Springer.

AutoTutor is an automated computer tutor that serves as a conversational partner with the student. AutoTutor is a working system that responds to Students’ natural language contributions by simulating the dialog moves of normal (not expert) human tutors. Descriptions of how AutoTutor works have been discussed rather extensively in previous publications, and will therefore, not be provided in this brief proposal (see Graesser, Franklin, Wiemer-Hastings, & the TRG, 1998; Graesser, Wiemer-Hastings, Wiemer-Hastings, Harter, Person, & the TRG, in press; Hu, Graesser, and the TRG, 1998; Landauer & Dumais, 1997; McCauley, Gholson, Hu, Graesser, & the TRG, 1998; Olde, Hoeffner, Chipman, Graesser, & the TRG, 1999; Person, Graesser, Kreuz, Pomeroy, & the TRG, 1999; Person, Klettke, Link, Kreuz, & the TRG, 1999; Wiemer-Hastings, Graesser, Harter, & the TRG, 1998; Wiemer-Hastings, Wiemer-Hastings, & Graesser, 1999). The creation of AutoTutor was inspired by numerous studies that have systematically analyzed the collaborative discourse that occurs between human tutors and students (Fox, 1993; Graesser & Person, 1994; Graesser, Person, & Magliano, 1995; Hume, Michael, Rovick, & Evens, 1996; McArthur, Stasz, & Zmuidzinas, 1990; Merrill, Reiser, Ranney, & Trafton, 1992; Moore, 1995; Person & Graesser, 1999; Person, Graesser, Magliano, & Kreuz, 1994; Person, Kreuz, Zwaan, & Graesser, 1995; Putnam, 1987). One reoccurring finding in several of these studies is that human tutors rarely adhere to “ ideal ” tutoring models that are often integrated into intelligent tutoring systems. Instead, human tutors tend to rely on pedagogically effective strategies that are embedded within the conversational turns of the tutorial dialog.

AutoTutor’s overall effectiveness as a tutor is contingent on him being an adequate conversational partner. Hence, many of our efforts have been directed toward implementing mechanisms that will enhance his conversational skills. One such mechanism is the Dialog Advancer Network (DAN). The DAN depicts AutoTutor’s entire dialog move option space for any given student contribution (e.g., Assertion, Yes/No Question, Short Response). The DAN makes AutoTutor a more effective conversational partner in that it enables AutoTutor to: (1) adapt each dialog move to the previous turn of the student, and (2) indicate when the student has the floor to contribute. Both of these DAN functions are elaborated below.

Adapt each dialog move to the previous turn of the student. Coherence emerges in human conversations because participants generally try to adapt the content of their turns so that they are relevant (or linked) to the preceding conversational turn (Clark & Schaefer, 1987; Grice, 1975; Hobbs, 1979; McLaughlin, 1984; Nofsinger, 1991; Sacks, Schegloff, & Jefferson, 1978). This “ turn-adaptation ” process is somewhat problematic for AutoTutor because the content of his dialog moves is predetermined. That is, AutoTutor doesn’t generate the content of his dialog moves on the fly but rather selects each dialog move from a scripted set of moves that is related to the tutoring topic being discussed. The DAN, however, enables AutoTutor to make quasi-adapted dialog moves that are relevant to the learner’s preceding turn. As the DAN illustrates, AutoTutor assigns learner contributions to different speech act categories (e.g., Assertion, WH-Question, and Short Response) and then generates discourse markers, dialog moves, and answers to learner questions, accordingly. Thus, AutoTutor is able to sustain mixed-initiative dialog by acknowledging student Questions (e.g., “ What is RAM? ”) and by responding to Short Responses (e.g., “ I’m lost ” or “ I don’t know ”).

Indicate when the student has the floor to contribute Turn taking is an integral feature of the conversational process. To facilitate the turn-taking process in human-to-human conversations, speakers signal to listeners that they are relinquishing the floor (i.e., it is the Listener’s turn to say something). However, human-to-computer conversations lack many of the subtle signals inherent to human conversations. When conversational agents like AutoTutor lack turn-taking signals, computer users (in our case, learners) often do not know when or if they are supposed to respond. For example, in previous versions of AutoTutor, learners were often confused after AutoTutor’s Hint, Elaboration, and Prompt Response dialog moves. That is, the tutorial conversation would abruptly stop, and students would often ask lab assistants, “ Am I supposed to say something now? ” or “ I’m not sure what to do ”. This is illustrated in the following tutoring excerpt from a session about the CPU. AutoTutor says (Prompt dialog move): “ See if you can get this, when the CPU executes programs, it uses RAM as . . . ? ” Student types: memory. AutoTutor says (Positive feedback + Complete Prompt Response): “ Good! As short-term memory storage. ” Student says to lab assistant: “ Am I supposed to type something in? ”

In earlier versions of AutoTutor, the DAN was instrumental in helping us identify potential dead spots in the conversation (e.g., after Hints, Elaborations, Prompt Responses). We addressed this problem by adding more informative discourse markers and by adjusting the content of some of AutoTutor’s dialog moves. As illustrated in the figure in the attachment, all pathways in AutoTutor’s dialog move option space ultimately lead to the learner’s next turn. In recent tutoring sessions with AutoTutor, learners appear to know when it is their turn to contribute. In the next version of AutoTutor, the DAN will include gestures and other paralinguistic signals (e.g., eye gaze) that facilitate the turn-taking process. In the presentation, we will provide additional excerpts from tutoring sessions with AutoTutor that illustrate how the DAN has enhanced AutoTutor’s overall conversational capacities. We will also present data from a number of evaluative cycles in which AutoTutor’s conversational skills were rated by knowledgeable judges. Some of the evaluative cycles occurred before the DAN was fully implemented.

Clark, H. H., & Schaefer, E. F. (1987). Collaborating on contributions to conversations. Language and Cognitive Processes, 2, 19-41.

Graesser, A. C., & Person, N. K. (1994). Question asking during tutoring. American Educational Research Journal, 31, 104-137.

Graesser, A. C., Person, N. K., & Magliano, J. P. (1995). Collaborative dialog patterns in naturalistic one-on-one tutoring. Applied Cognitive Psychology, 9, 359-387.

Graesser, A.C., Wiemer-Hastings, P., Wiemer-Hastings, K., Harter, D., Person, N., & the TRG (in press). Using latent semantic analysis to evaluate the contributions of students in AutoTutor. Interactive Learning Environments.

Hume, G. D., Michael, J. A., Rovick, A., & Evens, M. W. (1996). Hinting as a tactic in one-on-one tutoring. The Journal of the Learning Sciences, 5, 23-47.

Landauer, T. K., & Dumais, S. T. (1997). A solution to Plato’s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological Review.

McArthur, D., Stasz, C., & Zmuidzinas, M. (1990). Tutoring techniques in algebra. Cognition and Instruction, 7, 197-244.

McLaughlin, M. L. (1984). Conversation: How talk is organized. Beverly Hills, CA: Sage.

Merrill, D. C., Reiser, B. J., Ranney, M., & Trafton, J. G. (1992). Effective tutoring techniques: A comparison of human tutors and intelligent tutoring systems. The Journal of the Learning Sciences, 2, 277-305.

Moore, J.D. (1995). Participating in explanatory dialogues. Cambridge, MA: MIT Press.

Person, N. K., & Graesser, A. C. (1999). Evolution of discourse in cross-age tutoring. In A.M. O’Donnell and A. King (Eds.), Cognitive perspectives on peer learning (pp. 69-86). Mahwah, NJ: Erlbaum.

Person, N. K., Graesser, A. C., Magliano, J. P., & Kreuz, R. J. (1994). Inferring what the student knows in one-to-one tutoring: The role of student questions and answers. Learning and Individual Differences, 6, 205-29.

Person, N. K., Kreuz, R. J., Zwaan, R., & Graesser, A. C. (1995). Pragmatics and pedagogy: Conversational rules and politeness strategies may inhibit effective tutoring. Cognition and Instruction, 13, 161-188.

Putnam, R. T. (1987). Structuring and adjusting content for students: A study of live and simulated tutoring of addition. American Educational Research Journal, 24, 13-48.

Sacks, H., Schegloff, E. A., & Jefferson, G. (1978). A simplest systematics for the organization of turn taking for conversation. In J Schenkein (Ed.), Studies in the organization of conversational interaction. New York: Academic Press.

Wiemer-Hastings, P., Wiemer-Hastings, K., & Graesser, A. C. (1999). Improving an intelligent tutor's comprehension of students with Latent Semantic Analysis. Artificial Intelligence in Education (pp. 535-542). Amsterdam: IOS Press.

Often repetition in oral and written discourse has been viewed as ineffective communication. Teachers admonish students for repetitive and incomplete utterances, which are also viewed as a mark of ineffective thinking. Tannen (1989) introduced a number of features characteristic of high involvement in dinner talk among adults. Repetition was one such characteristic. Weber (1995) hypothesized that repetition in child conversations might signal involvement with text. We will demonstrate, in the research to be presented, how repetition may be an important device in the creation of sustained dialogue and ideas and a sign of cognitive and affective involvement with literary text.

Rationale: There is little understanding about how involvement emerges or how it can be sustained during conversations about literary works. Julie of the Wolves (George, 1972), a winner of the prestigious Newberry Medal, was selected for this research because of its attention to culture and assimilation. This novel was chosen for its potential appeal to individuals in transition--including Mexican-American students who are exploring their cultural identities. It is a novel that immigrant children would find challenging because of the attention it gives to a child's cultural conflicts -- choosing between one's past traditions and the culture(s) of their new world. It addresses issues related to human survival in a markedly different world -- that of the animal kingdom.

Questions posed in the present research were: (1). What are the linguistic features that characterize high involvement in classroom conversations about text? (2). How does repetition function as a linguistic strategy of high involvement in classroom conversations? (3). What forms of repetition do students use? (4). How does the use of repetition evolve and change across four conversations?

Subjects. Mexican-American, 5th graders, 11 years of age, in an urban classroom in South Texas participated in the research. Students were selected on the basis of their school schedule and willingness to participate in the project. They had a mean score of 6.8 ranging from 4.8 to 8.1 on the reading portion of the Metropolitan Achievement Test.

Procedures. The data was collected over a 2-week period in a departmentalized language arts block that met for 2 and 1/2 hours each day for 4 days. Pre-reading sources and activities were designed to build prior knowledge for the text. The classroom teacher videotaped the 4 conversations during the regular class time. The videotapes were transcribed by two teachers, a researcher and a community liaison, with each turn numbered in sequence. Two teachers scored the repetitions. There was ninety percent agreement in the categorizing of repetitions.

Data Analysis. Repetitions used by students were categorized as: 1) self-repetitions of exact word or phrases that occurred within the same turn, 2) self-repetitions that were repeated from what students had read (not from what students had said), 3) self-repetitions of words, phrases or ideas that occurred across turns.

Findings. The children who participated in the study used repetition to serve a number of purposes as they explained and negotiated the meaning of, Julie of the Wolves. To summarize, the fifth graders used repetition to verify, for emphasis, and to persist, in their beliefs and opinions about the text. Functions of Repetition (examples provided in presentation):

1. Verification Repetition is a function of self-repetition. The speaker recalls or uses the text to verify something they have said to capture the listener's interest and show mastery of the novel.

a.) Verification Text Language referred to the speaker's use of the author's language.

b.) Verification Text Search was used more frequently by speakers and refers to a speaker's reference to a particular part of the text, where the speaker stated a page number, location of an idea.

2. Emphasis Repetition occurs when speakers repeat their own words, phrases, or sentences within a single turn for purpose of emphasizing an idea or event. There are two types of emphasis repetition:

b.) Dramatization refers to the speaker's use of language explicitly for the purpose of involving and drawing the audience in a dramatic way. Involvement is building and reflected by dramatization in the 3rd conversation.

3. Persistence Repetition occurs when a speaker repeats a word, phrase, statement or idea over a series of turns. There are two types of persistence repetition:

a.) Turn Signal refers to a speaker's use of repetition to enter a conversation. It is a false start, which occurs during someone else's turn for the purpose of requesting the floor. This sets the audience up to expect something of importance to be said.

b.) Stance occurs when a speaker has strong beliefs, takes a stand about those beliefs and as the conversation moves along, persists at returning to their position.

The turns in self-repetition increased significantly between conversations one and two (p < .02) and between conversations two and three (p < .005). Conversation four had a significantly higher percentage of repetitions than conversation one (p < .005). The turns in verification repetition increased between conversations two and three (p < .02) and decreased between conversations three and four (p < .003). The turns in emphasis repetition decreased significantly between conversations one and four (p < .002). The turns in persistence repetition increased significantly between conversations one and two (p < .002) and between conversations one and four (p < .005).

This study demonstrated that repetition is used by students in teacher-led conversations and functions as a linguistic feature marking and sustaining involvement. The students involved in the study used the functions of repetition to verify, emphasize, and persist while communicating their ideas to the other students. Engagement with the text increased across the four conversations. As the conversations evolved, students had time to develop their beliefs about survival and acculturation. This was evidenced by a significant difference between the number of repetitions in conversations one and four. As the conversations continued, the students moved from a point where they were more text-focused (e.g., use of text verification was the highest in conversation three) to where they were engaged in emphasis and persistence of thinking about their own lives—conversations three and four. Conversation four shows a drop in text repetition, where students were able to step back from the text, and take an interpretive stance.

Most often discussions have been used to study levels of questioning and thinking. In the present research, discussion is viewed as the foundation for interpretation of a literary work. Discussion is a means for conceptualizing a text, engagement with a text, and for the creation of original ideas (Horowitz & Freeman, 1995; Horowitz, 1996, in press; Horowitz & Olson, in press). Our research suggests that extended conversations about text, over time, can help build the momentum that is needed for text involvement. Cultural and social considerations should also be a factor when looking at repetition in conversation. Repetition in conversation not only occurs in children of the United States but is also evident in other cultures of the world (Tannen 1989). The degree and type of repetition varies with the individual and culture of which one is a member. For example, African-American's use of repetition in call/response can be contrasted to the value put on silence by Athabaskan Indians (Tannen, 1989). In this study, the Mexican American student became emotionally attached to the cultural conflict of the main character, Julie, as she entered the world of the wolves. The study of repetition in small group conversations may explain the evolution of involvement and the stance speakers take when they become actively engaged with a text.

Speakers are said to design their utterances to suit the needs of their listeners, insofar as those needs can be known (Ariel, 1990; Fowler & Housum, 1987; Gundel et al., 1993). Yet maintaining an incrementally updated model of what the listener needs to know is a considerable cognitive task. Tests of this claim do not often control the difference between listeners' and speakers' needs (see Keysar, 1997). Where such needs are known (Bard et al., 2000; Brown & Dell, 1987; Dell & Brown, 1991; Keysar, 1997), speakers' behavior seems to be more influenced by their own point of view than by the listener's. Dell and Brown (1987, 1991) propose that speech is initially generated on-line with little regard for the listener, unless the listener explicitly requests adjustments. A recent test of spontaneous speech failed to find even this accommodation. The length and intelligibility of words in referring expressions is sensitive to the status of the referent as New or Given (Fowler & Housum, 1987), with shorter, less clear tokens for the latter. Words from a large corpus of spontaneous dialogue (Anderson et al., 1991) proved to be shorter and less intelligible when referring to entities which were Given in so far as the speaker had heard them previously in reference to the same entity. Whether or not that speaker had produced that word token previously, or the listener had heard it, or either could see the object referred to (Bard et al., 2000) was unimportant. These results accord with Dell's suggestion that control of pronunciation is attributable to a kind of production priming which depends solely on the speaker's experience (Balota et al., 1989; Mitchell & Brown, 1988). The present paper asks whether either proposal holds for form of referring expression, where changes with repeated mention are usually abbreviations [the blacksmith's cottage....it] which can be assigned a place in a scale of referential accessability (Ariel, 1990; Gundel et al., 1993). By using all the references to landmarks in the original corpus, we can compare speakers' treatment of both pronunciation (via duration) and form of referring expression (via accessability) in situations where speaker and listener knowledge differ. In conjunction with current views of the language production process, Dell and Brown's proposals predict that referring expressions, whose structure is set at a planning stage before utterance begins, will tend to be egocentric in nature, like other early processes. Only if listeners provide negative feedback will a later mention be planned with the listener in mind. Bard et al., suggest, however, that is the duration of the process which permits computation of the listener's needs.

We examined all expressions in the same route communication corpus which refer to landmarks, with the exception of those which were ambiguous as to accessability. In the HCRC Map Task Corpus (Anderson et al., 1991) instruction Giver's and Follower's maps matched only in alternate landmarks. Participants knew that their maps differed but not where or how. Each participant served as Instruction Giver for the same route to two different Followers and as Instruction Follower for two different routes. Digital stereo recordings were segmented at word boundaries, and coded for landmarks mentioned. Repeated references to landmarks using the same form of referring expression [the rift valley...the rift valley] or at least part of that expression [the valley] were compared for normalized duration (Bard et al., 2000). Those altering in form [the blacksmith's cottage....it] were classed by accessability on a scale running from 0 (cases with numeral or indefinite article + noun sequences) to 3 (pronoun, personal or relative).

The experimental designs of Bard et al., (2000) were replicated for all suitable referring expressions in the corpus. In each case, changes in duration replicated: As long as the speaker had heard the words used with the same referent earlier, a second use was a shorter token. This finding held when the current listener had not heard the prior token (F2(1, 238) = 12.48; p < 0.0005); whether the current speaker or the current listener had produced the earlier mention (mention: F2(1, 691) = 63.75, p < 0.0000; mention x prior: n.s.) and whether or not the current speaker could see the object introduced by the current listener (mention: F2(1, 224) = 12.37, p < .0005; mention x introducer: n.s.); and even when the listener had denied having the previously named object on his or her map (mention: F2(1, 193) = 9.45, p = .0024; mention x denial n.s.). Changes in accessability differed. Accessability did sometimes take the listener into account: Accessability did not change where the listener had not heard the first mention (mention: F2(1, 115) < 1). Yet accessability change was unaffected by intervening negative feedback (mention: F2(1, 135) = 23.33, p < 0.0001; mention x denial: n.s.). However, accessability was also more sensitive to the speaker's own needs: Accessability increase across successive mentions was more limited when the current speaker lacked the object initially mentioned by the current listener (mention: F2(1, 138) = 24.67, p < 0.0001; mention x sharing: F2 (1, 138) = 6.48, p < .02). Thus the results do not support Dell's prediction that earlier operations will be less sensitive to listener's needs, unless they follow pertinent feedback. Referring expression form was both more sensitive to listeners' needs and more sensitive to speakers'. This outcome is in line with the prediction that pronunciation and phrasal structure are composed by different mechanisms, with the latter available over longer-running processes than the former.

Anderson, A., Bader, M., Bard, E. G., Boyle, E., Doherty, G., et al., (1991). The H.C.R.C. Map Task Corpus. Language and Speech, 34, 351-366.

Ariel, M. (1990). Accessing noun-phrase antecedents. London : Routledge/Croom Helm.

Balota, D. A., Boland, J. E., & Shields, L. W. (1989). Priming in pronunciation: Beyond pattern-recognition and onset latency. Journal of Memory and Language, 28, 14-36.

Bard, E. G., Anderson, A., Sotillo, C., Aylett, M., Doherty-Sneddon, G., & Newlands, A. (2000). Controlling the intelligibility of referring expressions in dialogue. Journal of Memory and Language, 42, 1-22.

Brown, P., & Dell, G. (1987). Adapting production to comprehension -- the explicit mention of instruments. Cognitive Psychology, 19, 441-472.

Dell, G., & Brown, P. (1991). Mechanisms for listener-adaptation in language production: Limiting the role of the "model of the listener". In D. J. Napoli & J. A. Kegl (Eds.), Bridges between psychology and linguistics (pp. 111-222). Hillsdale: Erlbaum.

Fowler, C., & Housum, J. (1987). Talkers signalling of new and old words in speech and listeners perception and use of the distinction. Journal of Memory and Language, 26, 489-504.

Gundel, J.K., Hedberg, N., & Zacharski, R. (1993). Cognitive status and the form of referring expressions in discourse. Language, 69, 274-307.

Keysar, B. (1997). Unconfounding common ground. Discourse Processes, 24, 253-270.

McKoon, G., & Ratcliff, R. (1980). The comprehension processes and memory structures involved in anaphoric reference. Journal of Verbal Learning and Verbal Behavior, 19, 668-682.

Mitchell, D. B., and Brown, A. S. (1988). Persistent repetition priming in picture naming and its dissociation from recognition memory. Journal of Experimental Psychology: Learning Memory and Cognition, 14, 213-222

Wheeldon, L., & Lahiri, A. (1997). Prosodic units in speech production. Journal of Memory and Language, 37, 356-81.

Psycholinguistic studies of spoken discourse have been almost exclusively the study of two person dialogues. Research on computer-mediated interactions has more frequently examined the process of communication among small groups, although this has mainly been focused on turn taking behaviour. There is very little psycholinguistic research which compares the process and content of communication in dyads and small groups although the implications of the research of Herb Clark and his colleagues might suggest that achieving a shared understanding or “ common ground ” would be more difficult where more than two people are involved.

In a recent study (Anderson et al., 1999) we compared two and three party interactions supported by high quality videoconference links and compared these to comparable face-to-face interactions. We found that a similar level of task performance was achieved in the different types of interactions but that three party interactions, both face-to-face and mediated, required participants to expend more interactive effort, i.e. to use more talk to achieve their goal than in two party interactions. Here we explore what underlies these differences between 2 and 3 party interactions. In particular we explore to what extent the process of establishing common ground differs in multiparty interactions and the relative impact of increasing the size of the group from 2 to 3 with the introduction of multimedia communications technologies on these processes.

We used a collaborative problem solving task which we have used in previous research. The Map task has been found to be sensitive to the effects of communicative medium. Most importantly for the purposes of the present investigations it allows the researchers to control the knowledge available to each participant and thus to analyze the way in which participants resolve different kinds of knowledge mismatches and hence establish mutual knowledge during their interactions.

One hundred and forty eight undergraduates were paid £5 for participating as members of a task pair or trio. Participants attempted two versions of the Map Task, either face-to-face or using videoconfernce links. They swapped between the instruction giver (IG) and the instruction follower (IF) role, on trials one and two. The maps which are schematic both/all show a start point and a number of named landmarks, some landmarks being common to all copies of the maps but others are shown on only 1 (or 2) of the maps. The participants’ goal is for the instruction giver(s) to instruct the instruction follower to reproduce the route accurately. All are warned that some features of their maps differ.

In face-to-face conditions, paper maps were used and the three participants sat around a table with low screens preventing them from seeing one another’s maps, although their faces were visible. In the VMC conditions, participants were first shown how the high quality videoconference links operated as they were all new to such technologies. In particular they were shown how images were captured and relayed, and how to use the mouse to draw on the onscreen maps. The experimenters reminded participants that their collaborators were at a location three hundred miles away.

Task outcome was evaluated by comparing the map drawn by the instruction follower with the original being described by the instruction giver(s). Performance did not differ significantly across the conditions. Maps were as accurately reproduced in two and three party VMC conditions and in three party face-to-face interactions.

The communication process was explored in analyses derived from transcriptions of the task interactions. Dialogues differed across conditions, F(2, 54), = 3.27, p < 0.05, with significantly longer interactions in both 3 party conditions. The mean numbers of words per conditions were: VMC 2 party 1199, VMC 3 party 1759, face-to-face 3 party 1649. A second length analysis was conducted using the total number of turns of speech. This showed a similar pattern of results. Two party VMC dialogues contained significantly fewer turns than 3 party VMC interactions, and difference between 2 party VMC and 3 party face-to-face conditions approached significance (p = 0.08). Three party VMC and face-to-face interactions again did not differ in length.

We wished to explore what caused the increased output in multiparty interactions. To explore these issues a new form of communication content coding was devised. This involved an exhaustive categorization of each turn of speech in terms of the speakers’ and listeners’ knowledge states about the topic (map landmark) being discussed.

Categories KS1, KS2, KS3 occur in both 2 and 3 party interactions while KS4, KS5 and KS6 only occur in three party interactions. To test the reliability of the coding scheme a Kappa coefficient was calculated on sample dialogues coded by both judges and showed good interjudge agreement, (kappa = 0.85). By comparing the lengths of interactions which concern directly comparable dialogue states i.e. KS1, 2 and 3 in the interactions we can assess whether there is a general increase in length in multiparty interactions. An ANOVA was conducted on these data which showed that the length of these stretches did not differ according to the number of participants nor communication medium, (Fs < 1).

A separate analysis was carried out on the 3 party data, comparing the mean lengths of stretches of dialogue for all knowledge states. Knowledge state was a significant main effect, F(5, 160) = 33.3, p < 0.001). Post hoc tests showed that significantly lengthier discussions occur around knowledge states (KS4 & KS5) where there is a mismatch of knowledge between the two instructions givers. There was no difference between face-to-face and mediated interactions (F < 1).

Anderson et al., (1999) showed that although two and three party interactions achieve equal levels of task success on a standard problem solving task, three party interactions both face-to-face and videomediated require more interactive work to achieve this outcome. The data clearly indicate however, that it is the complexities of multiparty interaction that seem to produce these lengthier interactions not the addition of technological support. In this paper we explored why three party interactions differed from those involving only two participants. The analyses of knowledge state clearly indicate that a key factor is the extra talk needed to align divergent knowledge states. Longer multiparty interactions arose from specific points in the task, notably from the difficulties in establishing common ground where knowledge is initially distributed among two of the three participants. Establishing common ground is significantly more difficult in multiparty interactions whether face-to-face or mediated by technology. Clark (1996) describes various possible representations for common ground as the basis for joint activities. Clark proposes that people operate in terms of a “ principle of justification ” that is they take a proposition to be common ground in a community only when they believe that they have a proper shared basis for the proposition in that community. Clark points out that the implications of this principle are that individuals expend considerable effort to find a shared basis for their common ground which in turns has impacts upon their language use. The data we have presented in this paper provide support for this position. The map task participants seem to be drawn by something like the principle of justification to extend this common ground throughout their temporary community. As this community enlarges from 2 to 3 participants this requires more collaborative effort. Thecommunication analyses suggest that the multimedia technology did not lead to any less of a sense of a language community. The challenge for future research is to understand more about the subtle characteristics of multiparty interaction, both face-to-face and mediated.

This paper explores the role of ideology in discourse comprehension and production within the more general framework of a multidisciplinary approach to the study of the relations between discourse and ideology. The more specific aims of this paper are to examine the following aspects of this general theory: (a) The general nature of ideology within a more general theory of social cognition; (b) The structure of ideological representations in social memory; (c) The relations between ideologies and knowledge; (d) The ways ideologies influence mental models of events and context; (e) The strategic 'ideological' processing of discourse properties in comprehension and production.

A multidisciplinary theory of ideology assumes that ideologies have both social and cognitive properties. Whereas the social dimensions are defined in terms of groups, group relations, group interests and power, at the macrolevel, and in terms of many forms of ideological practices at the microlevel, the cognitive dimension of ideologies is defined in terms of representations in social memory and its role in the processing of discourse and other forms of social action and interaction. In other words, ideologies are forms of socially shared cognition among others, such as social knowledge, attitudes, norms and values. One of the theoretical problems it shares with these other forms of social cognition is what exactly it means that social representations are "shared" by the members of a group, and whether such sharing is more or less the same as the way sociocultural knowledge or the grammar of language are shared by social collectivities. Does this imply, for instance, that these representations need to be strictly the same, or only more or less comparable? Another theoretical problem is the internal architecture of the social mind, and how exactly ideologies are related to other socially shared representations.

Any cognitive theory of ideology needs to spell out the structure of the mental representations that define an ideology. Traditionally, such an account would be given in terms of some kind of schema, that is, an ordered set of fixed categories used by social members (language users) to organize the ideological representations of the groups they are members of, or with which they identify themselves. Such a schema would explain the efficient acquisition, usage and change of ideologies in everyday social interaction (including discourse) within and between ideological groups. Comparisons with the possible structures of other forms of social representations, such as knowledge or attitudes, do not readily suggest a format for such ideological schemas. For instance, they are not like scripts, or like narrative and argumentative schemas. It is assumed that the categories of ideological schemas need to reflect the basic dimensions of social groups, such as (a) membership criteria (who 'belongs' to Us, and who does not), (b) typical activities (c) goals (d) norms and values (e) social position (relations with other groups), and (f) basic resources.

One of the classical epistemological problems in the (theory of the) social sciences has been the relationship between ideology and knowledge. Indeed, whereas knowledge is traditionally defined as 'true belief', ideological is usually negatively defined in terms of biased, partisan, distorted or otherwise misguided belief. Although this distinction may be further defined both in cognitive and social terms, our approach will not follow this evaluative approach. Indeed, ideologies may be very "true" and group knowledge may very well be "biased", depending on whose position or perspective is taken in their definition and evaluation. In our approach ideologies are not defined in terms of true or false beliefs, but in terms of their hierarchical relationship with respect to other forms of social cognition. Indeed, it is assumed that ideologies form the "axiomatic" basis of all forms of group cognition. This may mean that also the specific (and even specialized) knowledge of specific groups -- such as scholars, journalists, feminists, racists or ecologists -- that is, the shared beliefs held to be true within the group, may well be ideologically based. It will be argued however that this does not mean that *all* knowledge is ideologically based. Contrary to ideological group knowledge, for instance, we also distinguish socially shared (and generally presupposed) "common ground" knowledge, which is shared by a whole culture and hence does not differentiate between different ideological groups within the culture. The present paper will deal at some length with this problem of the cognitive nature of the difference between cultural knowledge and (ideologically based) group knowledge.

Ideologies influence mental models in various ways. First, by direct instantiation ("application") of general ideological propositions in the personal, individual evaluation of specific events. For instance, a general racist belief in the inferiority of Africans may be directly applied in all mental models featuring Africans as participants. This is not only true for mental models of events talked about, but also for mental models of communicative situations (so-called context models) in which members of a group may have ideologically based representations and evaluations of other participants. Secondly, ideologies may impinge on mental models indirectly, that is, through the instantiation of more specific forms of group knowledge and group attitudes, for instance in mental models of specific events related to immigration, abortion, nuclear energy or international policy.

In light of the theoretical framework sketched above, it will be assumed that producing or comprehending discourse in "ideological contexts" (that is, in social situations in which language users primarily participate *as* group members), may have itself a number of rather typical properties. Thus, it is often crucial that language users quickly and efficiently identify ideological group membership, typical activities, aims, norms, values, position and resources of current speakers or writers. That is, they must have strategies that allow them not only to understand a message (make a mental model for it), but also to understand its ideological "point" (make an adequate context model for it), namely as being conservative, liberal, feminist, racist, anti-racist, etc. This paper will examine some of these strategies of ideological discourse comprehension and production.

Studying often involves searching instructional documents for specific information. Locating information in documents requires the student to build appropriate search goals, to select relevant portions of the text, to evaluate content information, and to recycle the search process, updating the original goal structure if needed (Guthrie, 1988; Winne & Hadwin, 1998). It has been argued that information search is a specific cognitive skill distinct from reading comprehension, which may explain why it often proves difficult even for students with normal comprehension ability (Guthrie & Kirsch, 1987). Little is know about the development of search strategies in children. The ability to form goals, select, extract and evaluate information develops gradually between 4 and 10 years of age (Armbruster & Armstrong, 1993) and is related to learning and comprehension abilities (Raphael, 1984; Cataldo & Cornoldi, 1997). We investigated children's ability to search a junior encyclopedia in order to answer factual questions. We hypothesized that the number of search cycles and the amount of inferencing required to provide an answer may both affect children's performance quantitatively (e.g., search time) and qualitatively (e.g., search strategies).

Participants. 36 children from grades 3 to 7 (9 to 13 year-old) participated as part of class activity.

Materials. A commercial paper encyclopedia on the topics of nature and wildlife, intended for a readership of 8 to 12 was selected from the existing educational literature. The book was organized into broad topics (e.g. deserts), which were subdivided in more specific topics, each presented on two side-by-side pages (e.g., the California Desert). Page design emphasized the integration of graphical and textual information, with many color drawings of landscapes, animals and plants. In addition the book included a table of contents and a subject index.

Tasks and procedure. Four search questions were designed. Two "simple" questions could be answered using a single passage in the book. Two "complex" questions required the comparison of two passages on different pages. In each category one question was textually explicit, i.e., the terms in the question could be matched directly with terms in the text. The other question was implicit, i.e., matching the question and the text required the child to generate a simple inference. For each question, relevant information could be found in the table of contents and in the index. The children participated individually in one half-hour session. The book was introduced to the child, and questions were asked one by one, both orally and on a separate piece of cardboard. The child was instructed to use the book so as to find an answer to the question as fast as they could. The child was allowed to re-read the question as many times as needed both before and while searching the text. No search method was explicitly suggested. Whenever the child remained inactive for a while, the experimented prompted comments or questions. In case of a difficulty the experimenter provided help according to a predefined sequence: (a) instruction to re-read the question, (b) explanation of keywords, (c) hints about alternate search procedures. The experimenter also provided feedback on the child's answers. In case of a wrong answer, the experimenter asked to the child to search further. The maximum amount of time devoted to one question was loosely set at 10 minutes. The experimenter transcribed the child's action (turning pages, using table of contents or index, looking at headings, pictures, reading the question etc.) as well as the child's comments or questions.

Hypotheses. We expected that both complex and implicit questions would be harder to answer, especially for younger children. In addition, we expected that spontaneous strategies would evolve from simple page turning to top-down selection through the table of contents, index and page heading.

Search time and correct answers. Search time was analyzed using a three-way ANOVA. Search time decreased with age (F(2, 33) = 12.09, p < .01), and increased for complex and implicit questions (F(1, 33) = 11.83, p < .01 and F(1, 33) = 33.14, p < .01, respectively). Furthermore, we observed an age by impliciteness interaction: Implicit questions took much more time than explicit questions for 9 years old and 11 years old, but not for 13 years old. Complex questions only took more time for 11 years old. The percentage of spontaneous correct answers (i.e., correct answers provided without any help from the experimenter) increased from 46% to 81% and 83% in 9, 11 and 13 years old, respectively. Implicit questions were the most difficult to solve for 9 year-old, whereas complex questions were the most difficult to solve for 11 years old.

Search strategies. Three types of strategies were observed. Flipping through pages, looking for a keyword or a picture was 9 years-old’s most frequent behavior. Searching the table of contents was observed in 11 and 13 years old, while searching the index was the dominant strategy only in 13 years old.

Searching an encyclopedia to answer factual questions was far from a trivial task for most of the participants. Many of them tended to forget the question, to lack an explicit strategy, or to ignore structuring devices such as the table of contents or the index. The ability to search more quickly and efficiently seems to develop gradually between the ages of 9 and 13. However, some 13 years old students still rely on rudimentary strategies such as flipping through pages and reading everything from top to bottom. Our data suggest that 9 years old students are mostly concerned with reading words, accessing word meanings, and confronting words in the question with words in the text. Hence their frequent failure when the text and the question do not exactly match. Thus, search efficiency in 9 years old seems strongly constrained by vocabulary and reading fluency. Remembering the question while searching is another problem, especially in 9 and 11 years old, which suggests that the development of working memory is another prominent cognitive factor of search efficiency in children (and possibly later on). Finally, students' knowledge of structuring devices (headings, page layout, standard superstructures, table of contents...) may also strongly influence their ability to achieve efficient search strategies. Our current research efforts aim at assessing the respective importance of working memory, reading fluency and comprehension ability on various measures of search efficiency in 9 to 13 years old students. Also of interest is the issue of whether and how children in the elementary grades can be taught more efficient search strategies.

Cataldo, M.G. & Cornoldi, C. (1998). Self-monitoring in poor and good reading comprehenders and their use of strategy. British Journal of Developmental Psychology, 16, 155-165.

Guthrie, J.T. (1988). Locating information in documents: examination of a cognitive model. Reading Research Quarterly, 23, 178-199.

Guthrie, J.T. & Kirsch, I. (1987). Distinctions between reading comprehension and locating information in text. Journal of Educational Psychology, 79, 210-228.

Raphael, T.E. (1984). Teaching learners about sources of information for answering comprehension questions. Journal of Reading, January, 303-311.

Winne, P.H. & Hadwin, A.F. (1998). Studying as self-regulated learning. In D. Hacker, J. Dunlowsky & A. Graesser (Eds.) Metacognition in educational theory and practice (pp. 277-304). Mahwah, NJ: Erlbaum.

LSA is a theory of meaning and knowledge representation developed by Landauer and his colleagues (see, for example, Landauer & Dumais, Psychological Review, 1997; Landauer, Foltz, & Laham, [Special Issue], Discourse Processes, 1998) that has rich potential for educational as well as other applications. Our presentation will focus on one of these applications, known as Summary Street, which is intended to help students learn how to summarize instructional materials. We will briefly discuss the educational rationale for this tool and describe the kind of feedback it provides. However, the body of the talk will concern the results of recent classroom trials with the tool.

Our goal in designing a computer tool to support the writing of summaries stems from the curricular needs expressed by two 6th-grade teachers with whom we have been collaborating. Their instruction emphasizes summarization as a means for students to deepen their understanding of complex new material, as well as to improve their writing and revising skills. The problem addressed by Summary Street is the need to provide students with enough practice in extended writing to accomplish these goals without producing an overwhelming amount of work for teachers in evaluating successive drafts of students' writing. Using LSA's knowledge analysis capability, feedback on the content of students' writing can be automatically delivered, allowing students to do much more revising without human guidance than has previously been possible. Summary Street, however, is not intended to replace the advice of human teachers, but rather to enhance their effectiveness. The feedback delivered by Summary Street is mainly in the form of a graphic display indicating to what extent each subtopic section of the source text has been covered by a student's summary. Missing topics are signaled with a link to the appropriate information in the text. The system also indicates whether the summary meets the teacher's prescribed length requirement, and it marks misspellings and flags sentences that may contain irrelevant or redundant content.

We used Summary Street in a series of classroom trials that formed part of the regular instructional curriculum during the spring and fall 1999 academic semesters. The three curricular units that involved summarization activities were about ancient Meso-American cultures, the human circulatory system, and renewable and non-renewable sources of energy. Two classes of 6th-grade students (52 total) participated in these trials during which they wrote and revised their summaries either using the summarization tool or traditional means (pen and paper or word processor). The source texts varied from about two to four pages in length, and summaries were supposed to be about a quarter or less of the original length (c. 150-300 words maximum). The purpose of the two trials that took place in spring 1999 (Meso-American and Heart-Lungs texts) was mainly to try out the system rather than to conduct a formal empirical test. Nevertheless, one intriguing result was consistently obtained in these trials: Namely, even though summary grades did not differ across conditions in these two trials, Summary Street summaries of certain texts received significantly higher scores than traditionally written ones. In both trials these were the texts that both teachers and students rated as more difficult, which was also reflected in the students’ grades. The trial in fall 1999 on energy sources took place under more controlled circumstances, in order to formally evaluate the system. During this trial students revised two summaries either guided by the full content feedback from Summary Street (Feedback condition) or using a similar interface that only provided feedback on length and spelling (No Feedback condition). Conditions were switched for the second summary written a week later. This procedure allowed us to track the time students spent working on their summaries, as well as successive drafts of students' work, and to look for possible transfer effects.

The two classroom teachers, who were blind to students' identity and condition, scored the summaries (a) for adequacy of content coverage of each subtopic section; and (b) for its overall, holistic quality. Our main results are as follows:

Time on task: Students who received feedback on content coverage from Summary Street spent more than twice as long working on their summaries than those did who received no feedback, 72 min. vs. 33 min., respectively.

Content scores: Averaged content scores (based on a scale from 0-2-points) for Summary Street summaries were significantly higher than those in the No Feedback condition: 1.29 points on average vs. 1.01 points, respectively.

Overall Quality: The teachers used a 5-point scale to rate the summaries' holistic quality, taking into account factors such as coherence, style, mechanics in addition to content. These scores were also significantly higher in the case of Summary Street summaries (3.19 points on average) than for summaries written without content feedback (2.87 points).

Text Difficulty: As in the previous trials, the benefits of Summary Street feedback were greatest for the most difficult texts and low to non-existent for the easier ones. This result suggests that the support offered by the tool is most effective when students are struggling with difficult text and task requirements (e.g., tight length constraints).

Individual Differences: We also found that the system was more effective for some students than others. Namely, moderate achievers seemed to benefit most from the feedback they received, whereas students whom the teachers considered high achievers apparently did not need the supportive feedback to write good summaries. Low achieving students, in contrast, could not profit as much from the feedback with texts that may have been too difficult for them. Nevertheless, these students were observed to work much longer and with greater interest using Summary Street than with most writing tasks.

Overall, these results offer strong support for the assumptions underlying the design of Summary Street: By automatically supplying students with feedback on the content adequacy of their writing, they can receive much more writing practice than is normally the case. This is because students can themselves assess this basic component of writing quality, rather than having to rely on a teacher's availability. Educators often complain that students rarely revise their writing, probably because throughout their academic training they have few opportunities to do so. However, we find that the feedback itself, together with the simple, but engaging environment in which it is delivered, is remarkably successful at keeping students fully engaged through repeated cycles of revising over long periods of time.

In recent years the study of argumentation text and of narrative have been of particular interest, both within psychology and in other disciplines. A distinction made about human thinking related to these topics is that one type of thought is of a more "logical" nature and includes the mental activity involved in mathematics and the various forms of logic tasks, and the other may be regarded as narrative, a sequential story telling type of endeavor (cf. Bruner, 1986). While this distinction is reasonable, and may be supported by lines of evidence, the idea that these two realms of study are in fact completely separated, seems to push the issue too far. Indeed, it was Aristotle who pointed out that the narrative could be used in argumentation, especially as it relates to courtroom situations. We have been investigating the issue of how narrative may be used in relation to argument, and what effects using narrative may have upon judgment. In a previously published initial study, participants were presented with an account of a hypothetical murder case, as presented to them by a prosecuting attorney. The attorney's account was a narrative of the committing of the crime, as the attorney interpreted it. We manipulated four narrative conditions, one being a baseline standard narrative, and other three being degraded in relation to the narrative's coherence and chronology (which cannot readily be separated), the narrative's causality, and the narrative's completeness (The completeness condition had a problem and is not discussed further). Narrative condition and four different texts were counterbalanced so that each participant served once in each condition and one time with each text. The basic idea of the study was that when the prosecuting attorney provides the account of the crime, the narrative may be divided into two components, the basic evidence, that is, the fingerprints found on the baseball bat, who was seen with the bat, etc. and the parts of the narrative that are not the hard evidence but consist of other parts of the narrative. What we then did is hold the basic evidence constant, but varied the narrative conditions. Participants then made two types of judgments after hearing and reading the respective narrative. They first judged the guilt of the accused on a 1-10 scale, and then they made judgments about the quality and persuasiveness of the narrative.

The hypothesis was that in those narratives that were degraded, the guilt judgments would be lower than in the baseline narrative. Moreover, the ratings of the quality of those narratives also should be lower than for the baseline narrative. These results were obtained. The primary significance of the results is that the quality of the narrative was shown to produce lower guilt ratings, even though the basic evidence was constant for all narratives. Looking at these results from the perspective of argumentation, the following analysis may be shown. The prosecuting attorney makes the claim that X, the accused, is guilty. To support this claim, the prosecutor provides a supporting narrative. But when the narrative quality is degraded, the attorney's argument is not as well as accepted as when it is not degraded. (Although not mentioned previously, the characteristics of what constitutes a good narrative were obtained from two studies, one in which historians provided attributes of a good narrative and one in which the attributes were obtained from the study of jury decision making.)

The two studies to be reported constitute a continuation of the above research. In one, the coherence/chronolgy condition is investigated in greater detail and in the second the causality degradation is studied. In addition, the possible effect of emotional textual statements is studied, the emotion pertaining to the crime or to the victim. The evidence presented in the first study was strong enough to produce guilt ratings of about 7 on the guilty, not guilty 10-point scale, but is would not be regarded as quite trong. A question then posed by the results of the first study is whether if you had good and poor evidence conditions, a good narrative could compensate for poor guilt ratings. A 2x2 study of quality of evidence and coherence quality was conducted. Coherence was manipulated by having a baseline text, with either good of poor evidence, and a text in which the baseline sentence order was scrambled, the manipulation used in the first study.

The results indicated that good evidence yielded significantly higher guilt ratings than poor evidence, with good evidence also yielding higher judgments of text quality and convincingness. In addition, males made more extreme guilt ratings than females. However, while the narrative manipulation influenced judgments of text quality and convincingness, guilt judgments were not significantly different. Clearly, a good narrative did not compensate for poor evidence.

In the other study causality was manipulated by including statements that made the evidence more probabilistic and less certain as "The fingerprints of the bat were Norman's" versus "The fingerprints on the bat were probably Norman's" In a second condition, the same type of manipulation took place, but if involved subject matter irrelevant to the crime. This was done to determine whether making irrelevant information more probabilistic could have an effect on guilt, a type of general effect. In addition, there was a baseline condition and an emotion condition in which the emotion pertained to the crime and another condition in which it pertained to the victim. Basically, the results indicated that the crime-related causal degrading influenced the guilt judgment and the textual judgments. The irrelevant causal condition has a weak effect. Emotion related to the crime did not influence guilt judgments, but it did influence the text judgments, for women only. The emotion related to the victim influenced both the guilt judgments and text judgments, but only for women.

The results suggest that evidence is critical for both guilt judgments and the judgments of narrative quality, that evidence overrides coherence as long as there is an intelligible level of coherence and the evidence is strong, that emotional contests influence judgments of text quality for women, and for women guilt judgments were influenced by victim-based emotion statements.

The objective of this research is to explain how the choice between different figurative meanings is carried out. To this end, we place the accent on the role of context in everyday situations of metaphor comprehension. We consider that in ecological situations of communication, the context comprises not only knowledge about the world but also new information provided by the situation in which the topic of the metaphor appears. Metaphor disambiguation is based on information furnished by the context which constrains the choice of meanings for the metaphor, and this, even in the case of conventional metaphors. Most of the cognitive models of metaphor understanding have adopted the approach according to which metaphor is an implicit comparison (Gentner & Wolff, 1997 ; Ortony, 1979 ; Tourangeau & Rips, 1991). From this point of view, understanding a metaphor "X (topic) is Y (vehicle)" consists in converting it into a simile "X (the topic) is like Y (the vehicle)" The mechanism posited by the comparison models is a mechanism of property matching. This is why these models are confronted with the problem of measuring the similarity of properties as well as with the problem of calculating the distance between properties which makes a simile literal or metaphoric. More recently, an alternative categorization approach has been proposed by Glucksberg et al., (1997) and Way (1991). According to Glucksberg's class inclusion model, a metaphoric statement of the type "X is Y" is solved by looking for the category, represented by the term Y, which furnishes properties that are potentially relevant for the topic X. The disadvantage of this model is that there is no mechanism for explaining how the relevant properties are selected. From our point of view, this aspect must be adressed if we are to explain how metaphors are disambiguated, i.e. to explain why one property is chosen and not others.

Our general hypothesis is that metaphor understanding consists of including the topic in the category of the vehicle and attributing to it the properties of that category that are compatible with what we already know about the topic. Legros, Tijus et Pudelko (1998) have shown that when the context provides specific information about the topic, the selection of vehicle features is compatible with this information. We are assuming that interpretation is constructed on-line and that knowledge about the topic intervenes at an early stage in processing by constraining the selection of features.

The experiment we report here aims at studying the way in which different contexts providing information about the topic activate features of the vehicle. We opted for presenting 14 metaphors in mini-scenarios in order to make the laboratory situation more like usual conversational situations in which animal metaphors are used. The scenarios were of three types: designed (i) to be compatible with conventional interpretations, (ii) to be compatible with one of the actual features of the animal or (iii) to be neutral, i.e. not particularly compatible with conventional or with novel interpretations (Table 1).

26 participants were first presented with the context sentence, then with the metaphor, and finally with an interpretation sentence. We tested property transfer by alternately presenting three different interpretation sentences of a given metaphor ("Does this mean that Guy is... funny? agile? likes music?). Participants had to say whether or not they agreed. We predicted that (i) when the interpretation sentence is not compatible with the context, the agreement score would be low; (ii) when the interpretation sentence is both compatible with the context and with a real or conventional property of the animal in question, the agreement score would be high; and, (iii) when the context is neutral, the rates of agreement would be higher in this case for the conventional interpretation which predominates.

Table 1. Percentages of "Yes" responses and Reaction Times of "Yes" responses. The 3x3 experimental conditions where, for each metaphor 3 context phrases describing a person in terms of (i) an actual animal trait (e.g. "likes to climb"), (ii) a conventional animal trait ("likes to joke") or (iii) an unassociated trait (e.g. "often goes to concerts") were crossed with 3 target phrases in which a trait is attributed to a person (i) an actual animal trait (e.g. "agile"), (ii) a conventional animal trait (e.g. "funny"), or (iii) a trait not associated with an animal (e.g. "likes music"). Each cell represents 364 data; 26 participants responded to the 9 conditions for each metaphor.

Our results show that the context in which the topic appears constrains the selection of the relevant trait to be transferred from the vehicle. Indeed, context can determine the selection of either conventional or novel traits. No effect was observed for kind of trait as long as both context sentences and interpretive sentences were compatible, in either case the metaphor was disambiguated quickly. Neither was kind of trait effect observed when the context sentence was incompatible with the target sentence.

How can the effect of initial knowledge about the topic in the construction of meaning for metaphor be explained, and in particular, how are elements of this knowledge selected? The CADS-T model proposes that the mental representation of situations described using language is composed of category networks which are formed by factorising the properties activated by the context as well as through a property adjustment process. CADS-T enables attributing a meaning to the topic through selecting a vehicle property as a function of the constraints provided by the topic's context. As we see it, the comparison and the categorization approaches are not antagonistic but complementary: understanding a metaphor consists both in (i) including the topic in the vehicle category and (ii) in selecting the properties that are compatible with the context.

Gentner, D., & Wolff, P. (1997). Alignement in the processing of metaphor. Journal of Memory and Language. 37, 331-355.

Glucksberg, S., McGlone M.S., & Manfredi, D. (1997). Property attribution in metaphor comprehension. Journal of Memory and Language, 36, 50-67.

Legros, D., Tijus, C., & Pudelko, B. (1998). The effect of knowledge about topic on metaphorical meaning. In Proceedings of the 20th Annual Conference of the Cognitive Science Society. p.1238. Mahwah, NJ: Lawrence Erlbaum Associates.

Ortony, A. (1979). Beyond literal similarity. Psychological Rewiev, 86 (3), 161-179.

Tijus, C., & Moulin, F. (1997). L'assignation de signification étudiée à partir de textes d'histoires drôles. L'Année Psychologique, 97, 33-75.

Tourangeau, R., & Rips, L., (1991). Interpreting and evaluating metaphors. Journal of Memory and Language , 30, 452-472.

Way, E.C. (1991). Knowledge representation and metaphor. Dordrecht: Kluver Academic Publishers.

In this paper, we briefly describe Relevance Theory (Sperber & Wilson, 1995) and a cost/benefit explanation of metaphors that the theory inspires. We argue that apt, unanticipated metaphors require extra cognitive effort to process. We present two developmental studies that buttress our claims based on a paradigm from Gibbs (1990). Our ultimate aim is to show how Relevance Theory can do a good job of accounting for psycholinguistic findings on metaphor.

Relevance Theory views inference-making as a constant feature of communication geared towards gathering in (and sharing) one's intentions. Essential to Relevance Theory is the claim that, in processing any utterance, people endeavor to draw out as many cognitive effects (i.e. benefits) as possible for the least effort (i.e. cost). Two features of Relevance Theory are crucial when describing metaphors (see Sperber & Wilson, 1986): A) Utterances need not be literally true in order for a listener to draw implications effectively and; B) A metaphoric utterance is likely to carry more information than a more-literal equivalent. To illustrate both features, consider a scenario in which a swimming instructor says "you are a tadpole" to a 5-year-old. The utterance a) is not literally true while effectively conveying information from teacher to student and it; b) goes further than a literal equivalent (something like "you are a young child doing a frog kick"); at least, the instructor's metaphor is arguably endearing whereas the literal equivalent is not.

The second feature is of particular interest. Relevance Theory essentially argues that the metaphor prompts multi-tasking. In the tadpole example, the metaphor is making reference to the swimming student plus it is describing something about him and transmitting affection; a neutral expression like "you are a child" in the same context would be doing the reference portion only (and would not seem terribly informative). This analysis shows that there are (at minimum) two components to a full appreciation of metaphor: Understanding what the metaphor is referring to and understanding the interlocutor's intention in using it. Being that it is difficult to imagine that multiple cognitive effects associated with metaphor come cost-free, we make two claims. Our primary claim is that it should not be surprising to find that -- all other things being equal -- metaphors are costly to process when compared to literal controls. Our auxiliary claim is that the extra costs that come with an apt metaphor ought to be commensurate with extra benefits. That is, one should be able to detect that apt metaphors have benefits when compared to less-costly, non-figurative equivalents.

To address these issues more directly, we investigate the development of metaphor comprehension because differences in effort ought to be even clearer among less proficient readers. We adopted a paradigm from Gibbs (1990), in which he presented participants with seven lines of a story before presenting one of three different concluding sentences that varied their form of reference (metaphoric, synonymous, and new). We prepared stories in French having metaphoric or synonymous references in the penultimate line and a follow-up question. For example, one story told of second-graders who went off to a swimming class. At one point during the class the teacher was interrupted by a phone call and upon returning he ordered all the children to the side of the pool, at which time they were sent to their lockers. The second to last line of this story was expressed either as:

In two experiments, stories were distributed so that each participant would see half the stories with a metaphoric reference and half with a synonymous reference.

Two hundred and thirty children between the ages of 8 and 12 were presented 16 stories on paper. All the stories provided questions that directly asked about the referent and all the questions required a "yes" or "no" response. For example, with respect to the story above the question (presented while the text was available) was "Were the pupils the ones who went to the side of the pool?" Regardless of formulation (metaphoric or synonymous), the correct response was "yes." We found two revealing effects. First, rates of correct responses show that referential ability in general improves with age. Second, formulations with synonymous references are linked to rates of correct responses that are consistently about 7% higher than those with metaphoric ones until around 12 years of age. Among 12 years old, one sees the gap close (to 2.6%). This indicates that metaphors do come with a small risk of leading young readers astray but the risk appears to diminish with age.

Fifty nine-years-old, 48 eleven-years-old, and 51 fourteen-years-old were presented 12 of the sixteen stories from Experiment 1. Forty adults were presented all sixteen stories from Experiment 1 plus filler items. In this experiment, each story was accompanied by one of three kinds of follow-up questions that justifies either a "yes" or "no" response: a) a question about a detail of the story, b) a general comprehension question, or c) a question like the one in the first experiment concerning the referent. Only one kind of question was presented after each story. The dependent measures were the reading times of the penultimate (metaphoric or synonymous) line and responses to the follow-up questions. Analyses concerned only those stories that were seen by both the children and adults. The reading times revealed that, at each age, sentences containing the metaphoric reference are read consistently more slowly than those containing the synonymous reference and that the gap closes with age, though never completely. The adult data confirm that a metaphoric reference in Gibbs's paradigm prompts a significant slow-down when compared to a synonymous control (a result that Gibbs, 1990, tended to disregard). The rates of correct responses confirm that the youngest children pay a small price in comprehension when they encounter a metaphoric reference. Adults' responses reveal that the metaphoric reference actually aids comprehension slightly. We remain cautious about our second finding because only the adults saw filler stories. Nevertheless, the pattern of results is consistent with the cost/benefit analysis that led to our initial predictions.

In conclusion, we argue that metaphors can be analyzed in terms of costs and benefits. Our findings show that a) compared to controls, metaphoric references consistently prompt longer reading times and b) in terms of comprehension, metaphoric references are sources of difficulty for younger children and sources of potential benefit for adults. We thus hope to show that a metaphoric reference is an imposition on a reader but that its potential for impact is linked with an ability to appreciate its intended meaning.

Gibbs, R. (1990). Comprehending figurative referential descriptions. Journal of Experimental Psychology: Learning, Memory and Cognition, 16(1), 56-66.

Gibbs, R. (1994). The Poetics of Mind: Figurative Thought, Language, and Understanding. New York: Cambridge University Press.

Sperber, D. & Wilson, D. (1986). Loose Talk. Proceedings of the Aristotlean Society, 86 (1985-1986), 153-171.

Sperber, D. and Wilson, D. (1995). Relevance: Communication and Cognition. Blackwell, Oxford.

Authors use typographical means such as sentence and paragraph boundaries in order to structure their message. Our research deals with spoken texts. The main question in this research is whether prosody has a comparable function in conveying information about text structure in spoken texts as typography has in written texts. More concretely, we address the question whether and how prosodic features are correlated with text structural features. Ultimately we search for a model for mapping the text structure onto prosodic structure. The research presented in the current paper consists of two parts. The first part deals with the question whether a theory designed for the linguistic analysis of texts can be applied reliably for annotating text structure and the second part investigates the relation between the linguistic analyses of texts and their prosodic features.

In order to find prosodic characteristics of discourse structures, we need a description of the structures of texts. To avoid circularity, the identification of discourse units and the relations between them should be independently motivated and not depend on prosodic considerations. There are several ways to obtain discourse structures. One possibility is an empirical approach by asking a number of judges to determine paragraph boundaries in the text. The result of this approach is a text segmented into several units, between which the cohesion differs depending on the number of judges who determined a break as a boundary. Although this is an objective and scientifically sound method, it leaves people's intuitions about the grounds for segmentation implicit.

An alternative method is to employ linguistic theories that describe the structure of texts. The advantages of such a theoretical approach are twofold. First, it does not require a large number of judges. Second, the resulting text structures are not based on implicit intuitions, but are theoretically motivated. However, before such theories can be applied for our purpose, we need to assess whether they are reliable.

The theory we used in our research is Rhetorical Structure Theory (Mann & Thompson, 1988, henceforth RST). RST analyzes a text into a hierarchical structure with labeled relations between discourse segments and clusters of segments by defining explicitly the conditions to be fulfilled for each relation. To find out whether RST can be applied in a reliable way, we conducted a study to test the inter-judge reliability. We used two news reports, representing informative texts, and two commentaries, representing argumentative texts, all of which had been broadcast on Dutch radio. The texts were segmented into elementary units according to RST’s definition of a clause. All texts had a length of about 30 segments.

Six expert RST analysts were asked to analyze the texts segmented into elementary units in terms of RST as reported in Mann and Thompson (1988). Although the texts had been spoken originally, the analysts had access to the written transcripts only. The output consisted of one hierarchical text structure per text per analyst. In order to compute the inter-judge reliability, quantitative representations of the hierarchical text structures were obtained in the following way. Each boundary in the hierarchical text structure of each analyst obtained a level number, starting at 1 from the topmost branching node downwards to the lowest branching nodes. The reliability was determined by computing Kendall’s measure of concordance W between the scores of the individual analysts for each text. The agreement between the analysts was .72, .73, .51 and .79 for the texts respectively. For all texts these reliability results were significant.

In addition, we compared the reliability for RST with that for the theory based on Grosz and Sidner (1986, henceforth G&S). Like RST, G&S analyze a text into an hierarchical structure. The basic idea of G&S is that a speaker or writer has a particular intention for uttering each constituent in the hierarchical representation of the discourse. The task of a discourse analyst is to recognize and to define these intentions, and to make decisions about dominance and satisfaction-precedence relations between them. Practical instructions for analyzing texts based on this theory are formulated by Nakatani, Grosz, Ahn and Hirschberg (1995). In our research three G&S experts were asked to analyze the same texts in terms of this practical guide. The analysts had access to the written transcripts only. The agreement between the three G&S experts was .73, .67, .68 and .68 for the texts respectively. These reliability results were significant too.

The main difference between RST analyses and G&S analyses is that the RST analyses are very detailed, generating very deep discourse structures, while the G&S analyses are less detailed, generating rather flat structures. With regard to prosody the question is whether the detailed RST annotations are reflected fully by the prosodic variation. If not, we might confine ourselves to more flat discourse analyses, like the one of G&S.

The second part of the research is about the correlation between the RST text structures of the four texts and their prosodic features. In a later stage we will use the G&S analyses to investigate their correlations with the prosodic features of the texts. We used one RST text structure per text to which the prosodic features were related. The spoken material was analyzed with a speech processing program. We measured pause durations at all segment boundaries and the maximum pitch of all segments (in Herz).

Our hypotheses about the relation between textual and prosodic features were as follows. (1) A higher segment boundary is associated with a longer pause. (2) A higher segment boundary is associated with a higher pitch maximum in the following segment.

There are several possible ways to test these hypotheses. For example, we may apply a relative method taking all adjacent segment boundaries in the text and look whether a boundary that is higher in the hierarchy is associated with a longer pause and a higher pitch maximum than a boundary that is lower in the hierarchy. Or we may apply a relative method taking segment boundaries that are adjacent in the hierarchy in terms of superordinate-subordinate relations and look whether a given (superordinate) boundary is associated with a longer pause and a higher pitch maximum than a subordinate boundary. Alternatively we may apply an absolute method and look for a direct relation between levels of text structure and prosodic features. We used these three methods.

According to the method of adjacent pairs, we inspected the prosodic features of all adjacent segment boundaries. On average higher segment boundaries are associated with a pause duration of .69 seconds and lower segment boundaries with a pause duration of .45 seconds (t(103) = -5.62, p < .001). In addition, segments following higher segment boundaries are associated with a mean maximum pitch of 215 Herz and segments following lower segment boundaries with a mean maximum pitch of 210 Herz (t(103) = -.72, n.s.). According to the other relative method, we inspected the prosodic characteristics of only those boundaries in the hierarchy that have a relation of subordination and superordination with each other. Based on 151 cases, Wilcoxon Signed Ranked tests confirmed both hypotheses (p < .001 for both pause duration and pitch maximum). According to the absolute method we correlated the levels of the segment boundaries in the hierarchies with the pause durations associated with these boundaries and the pitch maxima of the segments following these boundaries. The correlation between level in the hierarchy and pause duration was significant. However, pause duration showed a clear pattern for levels 1, 2 and 3 only. In other words, very long pauses occurred at level 1, shorter pauses occurred at level 2, and still shorter ones occurred at level 3. At lower levels in the hierarchical structure the pauses were more or less equally short. The correlation between level in the hierarchy and pitch maximum was not significant. The behavior of the pitch maximum showed the expected pattern for levels 1 and 2 only. These findings confirm the hypothesis concerning the relation between text structure and pause duration. The relation between text structure and pitch maximum is significant as far as subordinate and superordinate pairs in the hierarchy are concerned. At the same time the results of the absolute method suggest that there is no direct mapping between the fine-grained structure as obtained with RST and prosodic structure.

Grosz B. & Sidner, C. (1986). Attentions, intentions, and the structure of discourse. Computational Linguistics, 12, 175-204.

Mann, A.C. & Thompson, S.A. (1988). Rhetorical Structure Theory: Toward a functional theory of text organization. Text, 8, 243-281.

Nakatani, C., Grosz, B., Ahn, D. & Hirschberg, J. (1995). Instructions for Annotating Discourses. Center for Research in Computing Technology, Harvard University, TR-21-95.

History offers a literacy environment as rich as any a student is likely to encounter prior to college. The study of history centers on documents--letters, treaties, notes, official records, diaries--as well as textbooks. Instruction that makes good use of this rich text environment has the potential to support broad-based literacy skills that may extend beyond history classrooms to other cases of text learning, reasoning, and writing. It is not enough then merely to make documents available and to advocate their use. We must demonstrate, in standard classrooms, the value of document-supported history learning in a context that will promote actual student use of the documents. Both students and teachers must find document access to be useful and interesting. Toward that end, we have focused not just on developing an instructional tool that will encourage document use in the classroom, but also on the deployment of our tool in classroom situations where its effectiveness can be evaluated.

Sourcer's Apprentice is a Java application to promote an attitude of evidence seeking and evidence evaluation in students while developing a sense of document type and document privilege in history. It was designed to provide high school history students with opportunities to practice the kind of document-based reasoning exhibited by expert historians (Wineburg, 1991). Our goal was to create a simple coached-apprenticeship system (Gabrys, Weiner, & Lesgold, 1994) that would provide students the supports they need to interact with documents in a more authentic way. Although such systems are usually quite technically complicated (Lesgold, Lajoie, Bunzo, & Eggan, 1992), Sourcer's Apprentice is an example of how a coached apprenticeship system can be implemented in a very simple way.

This presentation will provide an on-line demonstration of the Sourcer's Apprentice, illustrating how the system can structure a student's inquiry into a historical problem. The Sourcer's Apprentice was built on a foundation of six principles which Cognitive Psychology has shown to be important for successful tutoring: Learn by problem solving, support expert representations, task decomposition, support transfer, provide explicit instruction, motivate engagement. We will demonstrate how each of these principles is realized in the Sourcer's Apprentice interface. In addition, a primary goal of our project has been not just to develop an interactive tool for history instruction, but to evaluate this tool in actual teaching situations. To date, the Sourcer's Apprentice has been used by lower-level "mainstream" students, intermediate level students described by their school as "average", and upper-level advanced placement students (Britt, Perfetti, Van Dyke & Gabrys, 1999).

We will present the results of a new study that examines how Sourcer's Apprentice should be integrated into the classroom. This study was designed so as to manipulate students' prior knowledge of the historical controversy to be explored. We found that prior knowledge did not affect the number of factual propositions students include in their essays, but it does affect the amount of original material in their essays. A corresponding increase in the number of connectives used by students with prior knowledge suggests that this new material represents more complex reasoning about the historical situation, including inferences about underlying causes and the motives for the actors involved. In addition, we have shown that even limited interaction with Sourcer's Apprentice increases the number of propositions and the number of connectives students use in their essays. We conclude by suggesting that the Sourcer's Apprentice can be useful either as a way to augment pre-existing instruction or as a way to present new material in a novel and engaging format. In addition, this study implies that learning from texts is hierarchically structured. Basic knowledge of the situation and actors is important not only as a context for text comprehension, but must be acquired in order to facilitate complex reasoning about causal relationships.

Research conducted over the past decade reveals that learners use a wide range of strategies to comprehend what they read. Think-aloud methods have revealed that adults and children who explain and elaborate what they are reading to themselves and who have a flexible approach to solving comprehension problems (i.e., use a variety of strategies) remember text and solve problems better than those who do not (e.g., Chi, deLeeuw, Chiu, & LaVancher, 1994; Coté, Goldman, & Saul 1998). Such activities reflect a “knowledge building” approach to creating a mental representation of the information conveyed in text. Successful knowledge builders engage in processing that transforms the text through self-explanations that rely on prior knowledge and active efforts to interconnect information from different parts of the text (Coté & Goldman, 1999). Knowledge building results in mental representations that are coherent, integrated with prior knowledge, and reflect the situation described by the text (Coté, & Goldman, 1999; Coté, et al., 1998). The vast majority of these studies have examined comprehension of individual texts.

However, in recent years, there has been growing interest in reading comprehension and learning across multiple texts (Bloome, et al., 1999; Goldman, et al., 1999; Perfetti, Rouet, & Britt, 1999). The multiple text focus is stimulated, in part, by research on disciplinary expertise that reveals that experts routinely engage multiple sources of information in the course of their work (e.g., Leinhardt & Young, 1996; Wineburg, 1994) whereas nonexperts do not. Experts integrate across documents, compare, contrast, and evaluate knowledge claims, and pay attention to the sociocultural context of the text (e.g., author, time, and place). Specifically in history, Wineburg (1994) reported that experts compare documents (corroborate), place them in their spatial and temporal contexts (contextualize), and consider the bias and intent of the authors (source). However, although university undergraduates noted the source of documents (Rouet, Britt, Mason & Perfetti, 1996), they tended not to create intertextual links among documents (Greene, 1994). Studies of student learning from social studies textbooks in primary and middle schools show that students typically learn lists of facts and acquire minimal understanding of causal principles and relationships among aspects of social systems (Beck & McKeown, 1994).

The present research examined the processing and knowledge building efforts of adolescents who read explicitly contradictory texts on the same historical event. Primary interest was in whether the patterns of processing and knowledge building were similar to those observed for the processing of single texts (Coté et al., 1998). Specifically would students process the texts as sequential, independent texts or would they draw connections between them and make reference to authors. In previous research we had found that students in this age range knew that a text reflected the perspective of the author but they did not use this kind of information in interpreting texts (Goldman et al., 1999).

A total of 44, 12 and 13 year old students participated. They represented a range of reading comprehension skill levels. All students had been exposed to similar instructional units on ancient Rome within the three months prior to their participation. Students read two, short (approximately 200 words) explanations for the Fall of Rome, explicitly indicated as having been written by two different historians. Each text explicitly stated disagreement with the other author’s explanation, made a claim about their own explanation and provided evidence for their claim. Students provided think-aloud protocols as they processed each text. After reading, students generated their own explanations for the Fall of Rome. The think aloud protocols were parsed into event types and coded using a system similar to our previous scheme (Coté et al., 1998). The coding categories distinguish among paraphrases, monitoring statements, evaluations and various kinds of elaborative reasoning about the texts, including self-explanations and connections among text segments.

The distribution of processing events for the sample as a whole indicated 23% paraphrases, 58% elaborations, 13% evaluations, 3% monitoring, and 3% problems. This is a higher rate of elaboration, a lower rate of monitoring, and about the same rate of paraphrasing and evaluating as we obtained with single text processing. Of the evaluations, 9% were of the author, indicating that students did not generally pay a great deal of attention to the author. Consistent with our previous work, about half of the elaborations were self-explanations that used prior knowledge. The other elaborations reinstated information from elsewhere in the text, from the first text that had been read, or prior knowledge that had been stated earlier in the protocol. Interestingly, 35% of the reinstatements were information from the first text that had been read, illustrating that students were connecting across the multiple sources. Furthermore, when students reinstated information from the previous or current text, 63% of the time they explained the connection. One interesting difference between the present and previous studies was in the percentage of prior knowledge elaborations that were categorized as irrelevant to the meaning of the text: 68% of prior knowledge elaborations were associations to personal experience. In past work, the topics of the texts made it more likely that associations to personal experiences would be relevant to the text interpretation. For historical accounts, personal experiences were less likely to be germane to the authors’ knowledge claims. The findings for the group as a whole mask some of the interesting differences in knowledge building approaches taken by individual and clusters of students. In ongoing analyses we are looking at processing profiles to better understand these differences.

Adolescents generally approached the processing of multiple historical explanations for the fall of Rome in a manner similar to single text comprehension. The differences, however, suggest that students were engaged in some critical evaluation of the claims being made in each text. Students did little reasoning about the authors of the texts, but did make attempts to integrate the content of the information across the two texts. These comparisons provide a basis for developing students’ skills in comprehending, evaluating and subsequently constructing evidence-based accounts of historical events.

Beck, I. L., & McKeown, M. G. (1994). Outcomes of history instruction: Paste-up accounts. In M. Carretero & J. F. Voss (Eds.), Cognitive and Instructional Processes in History and the Social Sciences (pp. 237-256). Hillsdale, NJ: Erlbaum.

Bloome, D., Goldman, S. R., Meyerson, P., Coté, N., Wolfe, M. B. W., & Mayfield, C. (1999, April). Reconceptualizing Reading as Intertextual Practice. Paper presented at the meeting of the American Educational Research Association, Montreal.

Chi, M. T. H, deLeeuw, N., Chiu, M., & LaVancher, C. (1994). Eliciting self-explanations improves understanding. Cognitive Science, 18, 439-477.

Coté, N., & Goldman, S. R. (1999). Building representations of informational text: Evidence from children’s think-aloud protocols. In H. Van Oostendorp & S. R. Goldman (Eds.), The Construction of mental representations during reading (pp. 169-193). Mahwah, NJ: Erlbaum.

Coté, N., Goldman, S. R., & Saul, E.U. (1998). Students making sense of informational text: Relations between processing and representation. Discourse Processes, 25, 1-53.

Goldman, S. R., Meyerson, P., Wolfe, M. B. W., Mayfield, C., Coté, N., & Bloome, D. (1999, April). "If it says so in the text book, it must be true:" Multiple Sources in the Middle School Social Studies Classroom. Paper presented at the meeting of the American Educational Research Association, Montreal.

Greene, S. (1994). Students as authors in the study of history. In G. Leinhardt, I. Beck, & K. Stainton (Eds.) Teaching and learning history (pp. 133-168). Hillsdale, NJ: Lawrence Erlbaum.

Leinhardt, G., & Young, K. M. (1996). Two texts, three readers: Distance and expertise in reading history. Cognition and Instruction, 14 (4), 441-486.

Perfetti, C. A., Rouet, J-R., & Britt, M. A. (1999). Toward a theory of document representation. In H. Van Oostendorp & S. R. Goldman (Eds.), The Construction of Mental Representations During Reading. Mahwah, NJ: Erlbaum.

Rouet, J. F., Britt, M. A., Mason, R. A., & Perfetti, C. A. (1996). Using multiple sources of evidence to reason about history. Journal of Educational Psychology, 88 (3), 478-493.

Wineburg, S. S. (1994). The cognitive representation of historical texts. In G. Leinhardt, I. L. Beck, & C. Stainton (Eds.), Teaching and learning in history (pp. 85-135). Hillsdale, NJ: Lawrence Erlbaum Associates.

When I asked undergraduates to explain to a partner how deadbolt locks work, they spontaneously spoke while gesturing, drawing diagrams, or demonstrating with a lock. In this paper, I claim that these multimodal explanations supported the construction of multiple analogical models of locks, each one associated with distinct regions within the physical space shared by interactants. Almost all the gestures, diagrams, and lock demonstrations contributing to a given model were located in the same region of shared space. In contrast, visible signals contributing to different models were usually performed in different regions. Systematically locating each visible signal in a particular region of shared space may have helped participants collaboratively ground the explanations (Clark & Schaefer, 1989) by helping addressees identify which model each multimodal utterance was updating.

Six English-speaking undergraduates examined a sample lock and studied a written explanation of how locks work. Each undergraduate then explained locks to a series of three individual undergraduates who did not know how they work. For one explanation, the sample lock was available; for another, paper and pen were; for a third, no props were available. Order of prop availability was counterbalanced. Having different props available was included to mimic the variation in real-world explanations. All 18 explanations were videotaped by three cameras, two horizontal on each participant, and one vertical focused down on the table where participants gestured, drew diagrams, and manipulated the lock.

Transcription and sampling: Speech was coded for pauses, false starts, and intonational contours. All hand movements including gestures, manipulations of the lock, and drawing or writing actions were described and then timed to the nearest syllable of ongoing speech. One episode from each explanation was intensively analyzed. Each episode was a coherent piece of the explanation that had been collaboratively grounded as a "conversational contribution" (Clark & Schaefer 1989). On average episodes contained 7.1 intonation units and 6.2 visible signals, a total sample of 127 intonation units and 111 visible signals.

Coding models: Each multimodal utterance (i.e., a visible signal and its associated intonation unit) was coded as updating a particular analogical model. This was done sequentially starting at the beginning of each episode. For each utterance, I determined whether any of the existing models could be updated with the information from the utterance while still forming a coherent model. To be coherent, a model had to: (a) include a connected set of objects, (b) use a single spatial perspective (e.g., side or top view, but not both), (c) be relevant to no more than one continuous time span (e.g., the lock's layout before or after the key is inserted, but not both), and (d) be specified at the same level of abstraction. An utterance was considered to establish a new model if a coherent model that satisfied all 4 conditions could not be formed by updating any existing model.

Identifying spatial locations: Using horizontal and vertical video displays, the outer boundaries of each visible signal were traced onto transparencies. Transparencies for different visible signals were superimposed over each other to determine whether pairs of visible signals were located in the same region or a different one. For each visible signal it was determined: 1) whether it was located in the same region as previous visible signals contributing the same model, and 2) whether it was performed in a different region from previous visible signals used to construct different models. To count as a different region, the locations could not intersect; to count as the same region they had to.

Using the same region for each model: Did explainers consistently use the same region to locate each visible signal that updated the same model? They did. Of the 87 visible signals updating an already existing model, 94% were located in a region that intersected the region used by previous visible signals contributing to that model. In fact, the locations of most visible signals updating the same model almost completely overlapped each other. There was no difference between conditions in how often the same region was used for the same model's visible signals (F(2, 84) = 2.13, p = .13)

Using different regions for different models: If models are located in shared space, then visible signals contributing to different models should be located in different regions of the space. In the sample, there were 132 pairs of visible signals that contributed to different models. Of these, an average of 81% were performed in completely non-intersecting regions. (Most exceptions only intersected slightly, suggesting that explainers at least made sure that different models were centered around distinct loci.)

More consistently using different regions in the paper and lock conditions: Visible signals for different models were more likely to be kept in non-intersecting regions in the lock (95%) and paper (97%) conditions than in the no props condition (63%, F(2, 129) = 13.95, p < .0001). Having the lock or paper provided new places for locating models which included permanent artifacts useful for keeping track of model locations. Without the lock or paper in the no props condition, participants could only locate their models with gestures, which also generally leave no trace after being performed.

In a large majority of cases, the spatial location in which explainers performed a visible signal was informative about which model was being updated by that visible signal and its associated speech. Different models were almost always located in different regions of lock, paper, and gesture space. These results extend previous findings on the use of space in gesture (McNeill & Levy, 1993) to include more kinds of visible signals (i.e., also diagrams and demonstrations) that establish more complex referents with speech (updates to analogical models). This use of space may also be similar to ASL and other signed language practices of distinguishing referents using spatial location (Emmorey, Corina & Bellugi, 1995; Liddell, 1995). Systematically locating visible signals in shared space may help interactants keep track of which model is being updated so they can coordinate their interpretations. In future work, I plan to test this hypothesis by seeing whether addressees' choices about which model to update are indeed sensitive to where speakers locate their visible signals. If they are, then locating visible signals in particular spatial locations is another mechanism people use to establish shared understandings in discourse.

Clark, H.H., & Schaefer, E.F. (1989). Contributing to discourse. Cognitive Science, 13, 259-294.

Emmorey, K., Corina, D. & Bellugi (1995). Differential processing of topographic and referential functions of space. In K. Emmorey & J.S. Reilly (Eds.), Language, gesture, and space (pp. 43-62). Hillsdale, NJ: Erlbaum.

Liddell, S. (1995). Real, surrogate, and token space: Grammatical consequences in ASL. In K. Emmorey & J.S. Reilly (Eds.), Language, gesture, and space (pp. 19-41). Hillsdale, NJ: Erlbaum.

McNeill, D. & Levy, E.T. (1993). Cohesion and gesture. Discourse Processes, 16, 363-386.

I will discuss local structure in written texts, taking a linguistic approach. I argue that we should recognize a level of structure, the discourse mode, which is intermediate between genre units, the units of discourse coherence relations, and sentences. Discourse modes are realized by stretches of text with characteristic semantic and pragmatic meanings. The meanings are licensed by composites of linguistic forms.

On encountering a discourse, one uses a great deal of world knowledge to understand it. Each discourse represents an activity of a certain kind with its own function and structure (Levinson 1979). But genre is not sufficient to account for what one knows about text structure. There are stretches of text which are intuitively of different types, e.g. Narrative, Description, Commentary, Argument. They tend to have a particular force which is due to cluster of linguistic forms and their interpretation. Stretches of text with different clusters realize different discourse modes. The discourse modes cut across genre lines. I will show that discourse modes can be identified through close analysis of a text. Linguistic forms give rise to semantic and pragmatic interpretation. I will use the term 'interpreted linguistic features' to include linguistic forms and the interpretations which they trigger, or license. There are key interpreted linguistic features which characterize each mode. I will investigate the five modes of Narrative, Description, Report, Expository-informative, Argument-commentary. In order to ensure a relatively small set I will allow for considerable variation within a mode. The modes vary in point of view, or authorial stance, and in register, or level of formality.

The notion of discourse mode accounts for the fact that actual discourses are rarely monolithic: They tend to contain passages of different kinds. Narrative sequences, for instance, appear in fiction, history, newspaper editorials, letters, etc. And in a narrative fictional text one finds descriptive passages, and perhaps commentary, as well as event sequences. Narrative, description and commentary make different contributions to a text; they also have different linguistic features. The discourse modes can be characterized according to three key interpreted linguistic features: (i) the types of entities they introduce; (ii) temporality; and (iii) principles of local organization.

(i) Types of entities. Type of entity is intended in the sense of Discourse Representation Theory: The sentences of a text licenses the introduction of various entities - including individuals, events, and times - into the developing semantic structure that represents the meaning of the text (Kamp & Reyle 1993). The entities that figure in this analysis are situations in the world, or events and states; and abstract entities such as complex statives, facts, propositions and perhaps others (depending on level of detail). The linguistic forms that license recognition of these entities are complex, usually involving a verb and its arguments. Each type of entity constitutes a covert linguistic category with unique properties, distributional and semantic (Vendler 1967; Smith 1991; Asher 1993; Peterson 1997). Very generally, texts in the modes of Narrative and Description primarily involve situations; Report and Exposition primarily involve situations and facts; Argumentation has facts, propositions, and often situations.

(ii) The temporality of a text depends to some extent on the entities it introduces. Events and stage-level states are located in time, whereas abstract entities (individual-level states, facts, propositions) are atemporal. If located in time, situations are static or dynamic. Temporally located entities appear mainly in the Narrative and Report modes, both dynamic; and Description, which is static. The atemporal discourse modes are Exposition and Argumentation. Relevant linguistic forms are tense, adverbials, constellations of verbs and their arguments (Smith 1980, 1991, 1998).

(iii) The third interpreted feature is the principle of local organization. Principles differ among the modes. Narrative is organized in terms of narrative advancement, which is fairly well understood. Narrative events are sequentially ordered; narrative time advances with each successive closed event, and does not advance otherwise. Thus narrative is organized in terms of temporal location, assuming the pervasive analogy between space and time. States involve metaphorical location and motion, focusing on what is predicated of the primary figure and how the components of the situation pertain to the figure (following Talmy, 1985). This idea provides a basis for understanding of certain non-narrative modes.

Texts in the Descriptive mode advance from one part of the scene to another; more globally, from scene to scene. Texts in the Report mode are organized either sequentially according to the reporter's perspective, by changes of time or space, or in terms of the domain talked about. Reports tend not to present events sequentially but are organized by the reporter, at Speech Time (Caenepeel, 1995).

Extending the approach to abstract entities, I assume metaphorical location and motion within a domain. The principle of advancement for modes with abstract entities involves metaphorical change of location, motion from one part of a given domain to another. In the Expository-Informative mode, the domain is often organized by the material talked about; in the Argument mode, the form of the argument usually organizes the material. Thus a stretch of text in this mode may move from claim to evidence or, more globally, from claim to claim.

Metaphorical movement from one part of a domain to another is conveyed by linguistic cues of different types. For texts with abstract entities the linguistic forms that realize dynamism of narrative time are not available. Cues include shifts from one type of entity to another, e.g. from Fact or Proposition to Event. Other cues involve information structure: change of sentence topic is an important cue. Pragmatic inference is often required to understand such shifts. Position is only one of several kinds of cues to sentence topic in English texts. Topic in English is determined by a confluence of factors, including the structure of adjacent sentences. Cues include word order, lexical repetition, syntactic structures that defocus, and patterning with other sentences. The points are illustrated with discussion of real examples.

I assume that receivers assemble and interpret the different cues in a text, e.g. lexical choices, syntactic and information structure, patterns within and across sentences. Pragmatic inference is used frequently as one arrives at an interpretation. The preferred interpretation is most compatible with all the information and relevant pragmatic knowledge. I do not attempt to state the processes involved.

Smith, C. S. (1980). Temporal Structures in Discourse. In N. Rohrer, (ed), Time, Tense, and Quantifiers.

In the 1980s, Rhetorical Structure Theory (RST) was developed as a way to explore the structure of written monologues. Typically an RST analysis starts by dividing a text into some minimal units of interest, such as independent clauses or orthographic paragraphs. It assigns a role in the text to each such unit, primarily by linking parts of the text together using relations, and aggregating related parts into spans. Spans can be linked to other units or spans, and so that the text is connected together into a hierarchic structure. Most of the relations are asymmetric, linking a nucleus to a satellite. The largest span created in this manner encompasses the whole text. One use of RST has been as an account for the coherence of written texts. Coherence of a text can be seen as the text having an evident role for every part. Negatively, it is the absence of non sequiturs. RST analyses are created by people (observers) who make a structured set of judgments about a written text. An RST diagram is equivalent to affirmation of a particular set of judgments, which can be identified without consulting the observer. Thus each analysis has a definite interpretation. RST has been used more or less directly in a wide range of work. In addition to studies involving coherence, it has been used to study inter-clausal relations, conjunctions, clause combining, implicit communication, writing style and genre. In computation it has been used in text generation, summarization, text indexing and modeling of comprehension. The scientific status of RST has been found to be unclear, which has led to confusion and inappropriate expectations. A better understanding of this point would clarify notions of success, progress, appropriate rigor, completeness of accounts, degree of demonstration, prediction, confirmation, significance. This paper is intended to contribute to the understanding of these matters.

The paper presents two views of RST. In one view, RST is simply a way of identifying and making precise certain phenomena of language use. It collects data, and the data collection activity encounters patterns which help us to understand how language functions. It is a partially subjective process, involving reading and producing a diagram of text structure. RST in this view does not make claims or build hypotheses about how language works. We will call this the data gathering view. In the second view, RST is a partial model of how people structure written texts, or how they seek structure in written texts. Such a partial model could in principle be combined with a collection of other partial models to form theories of language use. Such theories would be subject to the normal scientific processes of formulation, test, confirmation, refutation and revision. We will call this the theory component view.

RST was defined with principal concern for the data gathering view. The developers were particularly concerned that each RST analysis has a definite interpretation. This was carried out partly by using definitions for relations, so that the process of text analysis is essentially application and affirmation of definitions. In order to make RST analysis more reliable, and also to increase the tendency toward agreement between observers, requiring observers to make difficult judgments has been avoided. As a result, the relations and structures of RST are more irregular than they otherwise would be. For the data gathering view, we examine the scientific interpretation of some of the patterns that have been identified during extensive RST analysis of texts. These include the very high proportion of texts that have analyses, tendencies for analysts to agree, low perceived ambiguity of discourse structures, relationships between coherence of a text and the existence of an analyses, the wide diversity of text genres for which RST is effective, patterns of explicit and implicit indication of discourse structure, and ways that authors communicate implicitly using the discourse structure. These are not simply issues of methodology. Rather, the fact that the analysis process works tells something about what it is working on, i.e. about the nature of text. For RST as data gathering, it should be as "theory neutral" as possible, but complete neutrality is not possible. There are assumptions built into any text analysis method. A collection of assumptions inherent to the data gathering view are examined, including assumptions about language, text, observers and authors. The other view, the theory component view, takes a different approach to text structure.

Text structure has been seen in many ways, sometimes in isolation and sometimes as part of a larger linguistic scope such as semantic structure, pragmatic structure or communication. The underlying abstract nature of text structure -- its theory, has been considered by some to be a well-formedness notion, akin to a text grammar. Another suggestion is to see text structure as aggregations of cohesive links. Compositions of speech acts have also been suggested, as well as expression of particular genre organizations, the semantics of discourse markers, or an outworking of necessary rhetorical features. RST does not follow any of these suggestions. Another group of suggestions, which are closer to RST's methods, says that readers or hearers seek to perceive the goals, intentions or desired effects of the producers of text, that text structure expresses some of those intentions, and that writers rely on this intention-seeking activity of readers.

The theory component view of RST is probably most easily conceived as a small portion of a model of text creation. This follows the first major use of RST in computational linguistics, which was to organize computer-based text generation. Partly because of this use, RST has sometimes been expected to be more regular, objective and formal than it really is as a data gathering approach. RST cannot be a whole model of text creation, of course, since it does not deal with the roles of words, grammar or semantics, and it does not touch many parts of pragmatics. For these reasons, RST cannot function in isolation in this role. In an overall model of text creation, RST will be accompanied by other components. Furthermore, RST does not specify how it would be used in text creation. It merely suggests guidelines for ways that coherence might be achieved and how intentions of the author could partially shape text construction activity. Because RST does not specify how it would interact with other formative forces, it needs to be extended and reconciled to other bodies of knowledge. When such extensions are made, there is naturally a need within RST for regularization and augmentation to meet the demands of the framework adopted for the model as a whole. Some of the projects that have used RST have clearly seen a need for this sort of development.

In summary, scientifically, RST is a way of producing particularly examinable cases of text affecting people, yielding phenomena and structured data rather than a model of how such cases arise. The work of analyzing texts produces insights and surprises about text structure and function, enriching the phenomena to be accounted for. RST contributes ideas about what might be included in theories (or models or explanations) of language creation and reception, but it is at most an incomplete fragment of such a theory.

Since the 70’s, research on inferences in text comprehension aroused great interest, and this, more particularly, since the experiments of Bransford and Franks (1972) who showed that in addition to information elements present in the text, elements of the readers‚ knowledge are activated during reading. Indeed, successful reading comprehension involves not only the representation of the textbase in memory, but at the same time, the activation of readers’ background knowledge, i.e. implicit elements that are inferred in order to obtain a coherent representation of the text. These inferences are a part of the relatively stable cognitive representation of what the text is about: The situational model (van Dijk & Kintsch, 1983, p. 337).

In our work, we are interested in the activation and maintenance of activation of knowledge about human emotions and the use of the activated knowledge to build mental representations of fictional characters’ emotional states. If research on inferences are numerous today, the particular type of emotional inferences received only little study. These inferences correspond to the clarification of the emotions felt by the protagonists in response to the events or actions reported by the text (Graesser, Singer, & Trabasso, 1994). Three characteristics of emotional inferences seem to emerge from literature today. Emotional inferences would be generated during reading (on-line) and not once the reading accomplished (off-line). Indeed, in a succession of three experiments, Gernsbacher, Goldsmith and Robertson (1992) showed that the reader activates his knowledge about human emotions when reading a narrative, in order to infer the protagonist’s emotional state. In extension to Gernsbacher et al.’s (1992) research, de Vega, Léon and Diaz (1996) specify that, not only emotional inferences are produced during reading, but they remain activated in working memory throughout the reading of the text. Emotional inferences would thus make it possible to ensure establishment of coherence when the text presents ambiguities or inconsistencies such as the introduction of new characters, changes of places or temporal discontinuities. De Vega et al., (1996) postulate that the interest that the reader carries to characters’ emotional states supports integration of different portions of the text. From this point of view, emotional inferences would contribute to the development of a global coherent text representation. Results of another experiment of de Vega et al., (1996), permits to advance that emotional inferences are dynamic, that is, once produced, they are updated after a change in the protagonist’s emotional situation in the continuation of the story.

We have examined the course of emotional inferences activation, within the framework Landscape Model (van den Broek, Risden, Fletcher & Thurlow, 1996). Indeed, this model of text comprehension permits us to study the inferential processes that take place during reading in considering inferences‚ activation in terms of a continuum of activation, and the way in which these processes lead to the construction of a relatively stable memory representation of a text. So we have been able to explore at the same time, the moment of emotional inferences’ production and the maintenance of their activation in memory. A conventional way of representing a successful comprehended text in memory is in terms of a coherent structure. This structure can be represented as a network of nodes (corresponding to concepts) that are linked by anaphoric and causal connections (Trabasso & Suh, 1993; van den Broek & Lorch, 1993). On the force and the number of connections associated with a particular concept depend the probability of its recall (Graesser & Clark, 1985; Miller & Kintsch, 1980; Trabasso et al., 1984; Trabasso & Sperry, 1985). Moreover, considerable research has demonstrated that concepts that are activated simultaneously are connected in memory representation (Fletcher & Bloom, 1988; Fletcher, Chrysler, van den Broek, Deaton, & Bloom, 1995; Kintsch & van Dijk, 1978). Thus, activation and co-activation of the different concepts conveyed by a text, can be determining in the recall of a text.

In considering the factors which determine the recall of concepts of a text according to van den Broek and al., on the one hand, de Vega and al.’s considerations (i.e. emotional inferences‚ activation maintenance throughout the text), on the other hand, we can advance that emotional inferences will have a better score of recall than do more neutral inferences. Indeed, emotional inferences are maintained activated longer, which according to authors of the Landscape Model would generate an important rate of total activation and, consequently, a high rate of recall. Moreover, if emotional inferences are maintained activated in memory during a more significant number of cycles, they will be probably strongly connected to other concepts, increasing their probability of recall. In sum, we expect that emotional inferences occupy an important place in the representation of the text in memory. We proceeded in two stages. First, we have employed the procedure developed by the authors of the Landscape Model. That is, we have submitted four emotional and four neutral stories to two groups of ten participants that noted activation scores for each word and several inferences on an activation matrix. The first group of participants (called "expert") received preliminary training of the contents and the application of the activation rules, the second group (called "naive") was not informed of these rules. Results show a) that the Landscape Model efficiently predicts activation of concepts during reading (i.e, results of the two groups correlated at r = .94), and b) emotional inferences are maintained activated longer than predicted the model.

Second, we proceeded to the recall tasks. We solicited 120 volunteer psychology students to whom we submitted the four emotional and four neutral stories (each participant reads one story). In order to measure the memorizing of the texts, we replicated the step of van den Broek et al. We submitted the narrative texts by giving the following instruction: " Read this story very attentively during one minute ". Immediately after reading the text, we gave the participants this second instruction: " Mark on a sheet of paper, all you remember of the previously presented story ". Then, after a one hour’s delay, the same instruction was given. And finally, a supplementary task of recall was effectuated after a week’s delay. We also proceeded at a recognition task at that moment. We presented six sentences to the participants: One sentence containing the emotional and neutral inferences, one sentence that made part of the story, two sentences that did not made part of the story, and two paraphrases. Participants had to answer if yes or no these sentences made part of the story they read a week before. And, in addition they had to indicate on a five point scale how sure they were of their answer. By effectuating this recognition task we were able to distinguish the accessibility and availability of our inferences. Results indicate a general superiority of emotional inferences recall in comparison with more neutral inferences. On the contrary, there is no difference for the two types of inferences concerning the recognition task.

This research is part of the studies investigating the processes and conditional production of causal inferences during reading, leading to a causal network representation of the text: Namely the ones on the Causal Inference Maker model (van den Broek, 1990; van den Broek, Fletcher & Risden, 1993). In this model the causal strength of a relation and the distance between two statements in the text surface structure determine the type of causal inference produced during reading, mainly forward and backward inferences. Our purpose was to deeper investigate the influence of sufficiency and distance between two statements on the production of forward causal inferences during narratives reading. Indeed, van den Broek and his collegues showed that the more sufficient a relation is, the more likely it is for a predictive forward inference to be made, which means activation of a highly probable consequence. If the reading of a cause is due to lower sufficient relation but with high necessity, the reader can infer the importance of the event and maintain it activated until its consequence appears, which is another type of forward inference. Therefore the cause and/or the consequence can be maintained activated during the production of forward inferences. The questions that arise are (1) the possibility for a forward inference to become backward under the influence of the distance between the cause and the effective appearance of the consequence in the text surface structure. Is it a decay of the cause activation or of the probable consequence one? (2) What are the events activated and maintained during the production of predictions: The cause, the consequence, or both? Varying the distance between the cause and the appearance of its consequence in the text surface structure, and varying the sufficiency of those relations should allow us to determine the conditions under which the cause and/or the consequence are activated and maintained.

To investigate this theoretical questioning, we used an on-line probing task on words refering either to the cause event or to the consequence event. The reading was interrupt just before the reading of the consequence. Two causal relations were tested for each narrative (two interruptions per narrative). The causal relations could be adjacent (one sentence), semi-adjacent (five sentences) or non-adjacent (ten sentences) in the text surface structure. Moreover they could satisfy to high sufficiency (physical causality) or to medium sufficiency (psychological causality). Necessity was always satisfied. Therefore we constructed 12 simple narratives of 22 sentences in 3 versions so as to have the same statements in the three distances. Then, these three versions of narratives were transformed so as to present less expected consequences, but satisfying to the same type of causal relation. We had then 6 versions of the 12 narratives. To select the most " fitting-to-naïve readers " consequences, we did a pre-test in which we asked naïve readers to complete small texts presented, with either a plausible consequence or a less plausible consequence but always true according to what they read before. As expected the answers for the physical causality relations were mostly the same. Differences appeared for the psychological relations, as sufficiency is lower, several probable consequences are possible. We used in our material the most often given concepts. Then for the probing task, we selected words refering to either the cause or the consequence, created non-words, and selected other words from the french language. A participant tested on cause words for a given relation, could not be tested on consequence words for that same relation. Therefore at each interruption, a participant had to judge whether the randomly presented words were words or non-words. The reading times and the reaction times of the probe task were recorded. Our main predictions on reaction times were that: (i) For consequence words refering to high sufficient statements, we expected no difference between the words cause and consequence, i.e., forward prediction. (ii) For medium sufficiency, we expected reaction times on consequence words to be lower than for cause words. And this difference would increase with distance. This decay of activation should be the result of the change of type of causal inference, i.e., from forward to backward. The distance semi-adjacent being the central point of activation decay.

Our results confirmed our main hypothesis. Probe words were rated faster than words. We found no significant effect of the interaction Cause * Consequence. But within reaction times on probe consequence words, high sufficient causal relations showed that consequences are maintained and integrated in memory as the distance increased. Reaction times decreased (but not significatively) according to the increase of distance. Therefore, forward inferences highly sufficient are maintained and strengthened through distance (assumption i). High sufficiency leads to the production of purely forward inferences (predictions). For medium sufficient causal relations, we found an increase in reaction times on probe consequence words under the influence of distance. Therefore, consequences activated are progressively deactivated as the reader proceeds through the text without encountering the effective consequence. A forward inference can become backward under the influence of distance (assumption ii). The distance semi-adjacent appeared to be the central point of activation decay/integration since we did not find any difference in reaction times for medium and highly sufficient consequences activation. Within reaction times on probe words cause, results showed a different pattern of activation for cause statements. It seems that probe words refering to highly sufficient causes are rated faster than the ones refering to medium sufficiency statements. Both strengths showed the same pattern of activation according to distance, i.e., progressive deactivation as the distance increases (assumption ii). With highly sufficient causal relations, a forward predictive inference was made and maintained regardless of distance (assumption i).

As a conclusion, we can assume that forward inferences consist in maintaining both cause and consequence events, i.e., the relation between the two events. Readers then, used their causal knowledge to predict future events. As long as they proceed through the text, the predicted relation is more deeply integrated in the network under construction, therefore when an effective unexpected consequence appeared in the text, they had to create a new connection, which takes more time. The lower the sufficiency is, the more different probable consequences can be infered. Nevertheless, our results on reading times, confirmed that the activation of the consequence is less important in forward inferences with lower sufficient relations. Cause and consequence being equally activated in the production of predictions with (notably) highly sufficient relations, we can expect that in some particular cases, a symetrical connection could be represented in memory. The same causal relations could be used in a priming task, according to the three distances, and presented in a related and reversed way, so as to investigate this possibility.

van den Broek, P. W. (1990). The causal inference maker: Towards a process model of inference generation in text comprehension. In D.A. Balota, G.B. Flores d'Arcais, & K. Rayner (Eds.), Comprehension processes in reading (pp.423-446). Hillsdale, NJ: Erlbaum.

The expression of a causal relationship in a text can either be implicit and suggested by juxtaposed clauses or on the opposite, explicitly signaled by causal connectives such as because or and. Several studies using sentence pairs have shown that the presence of an interclausal connective facilitated the construction of a coherent local relationship between the two sentences. For Millis &Just (1994), the causal connective because triggered the reactivation of the information read in the first clause and allowed the integration of the two clauses in a single representation. However, during the reading of a real text, several lines long, the readers normally attempt to achieve coherence at both local and global levels. Therefore, it is sometimes necessary when reading a text, to connect the current sentence to previous knowledge in long term memory as in this exemple from McNamara et al., (1996): “ The blood cannot get rid of enough carbon dioxyd through the lungs. Therefore, the blood becomes purplish. ” In spite of the presence of the connective therefore, the readers are not able to construct a coherent representation for these two sentences because they failed to integrate the information currently read to background knowledge as for exemple “ the blood that flows from the heart to the lungs is low in oxygen and looks blue whereas the blood that flows out of the lungs is high in oxygen and looks bright red ” (McNamara et al., 1996, p. 5). Thus, the role of connectives on scientifc texts appears to be quite specific because in narratives, it is easier for a reader to infer interclausal relationships from the content of the texts whereas in popular scientific texts, the construction of these relations is only based on cohesion cues and not on pragmatic coherence (Mac Cutchen, 1987). Millis, Golding and Barker (1995) described a facilitative effect for causal inference generation during the reading of pairs of narrative statements connected with because. However, one can argue that the sentences to be read included a coherence break that has possibly contributed to the inference generation. So, the aim of this study is twofold : We determined first whether the presence of a causal connective increase inference generation in scientific texts too; then, we raised the question of the durability of this activation effect by comparing the results of an on-line completion test and an unexpected cued recall test. In other words, we wondered whether causal inferences activated during reading are encoded and integrated in the representation of what the text is about.

Subjects. A total of 135 undergraduate subjects took part to this experiment. They were split in four experimental conditions "because condition","therefore condition", "and condition" and "a constrained word stem completion condition (without reading any text)"

Material. Nine experimental popular scientific texts and 6 filler texts were displayed. Attention was payed that none of the filler texts described similar technological or natural mechanisms as those evocated in the experimental texts. At the beginning of each text, there is no title and no superordinate goal explicitly mentionned in the introductory sentences. Specific background knowledge are not necessary to understand the texts because they rely on basic principles of physics or chemistry learned at secondary school. The target word occurred after the fourth sentence in the experimental texts and after the first or the second sentence in the filler texts.

Procedure. Subjects were instructed to read carefully the texts in order to answer final comprehension questions (not relevant to the details of each story). The sentences were either conjoined by a connective (therefore, and, because) or separated by a period. Subjects read the texts at their own pace in a self paced fashion. After the last segment of the sentence had been presented, the string "XXX" appeared on the screen and then replaced by the first two letters of a target verb. Subjects were instructed to give orally the first verb that came to their mind and fit the first capital letters of the target verb. Then, after the whole reading, subjects were asked to answer comprehension questions on the filler texts as soon as possible. Finally, an unexpected final cued recall test was proposed to the subjects in the three experimental conditions. They were instructed to write down the sentence in the text that immediately followed versus preceded (for half the subjects of each experimental condition) the clause displayed on the center of the screen. They were told that if they could not remember the exact words, they should write down any words or ideas that they could remember for the first versus second sentence.

We developped here only the relevant results to the verb completion test and the recall data. Verb completion scores were submitted to a mixed analysis of variance with one between subjects factor (the experimental condition: "because", "and", "therefore" and "a constrained word stem completion condition ") and one within subjects factor (with or without connective). The results showed an effect of the experimental condition [F(3, 104) = 5.88, p = .01] with the highest score of completion in the "because condition" and the lowest in the "constrained word stem completion condition". The connective "therefore" lead to a score of completion significantly different from the "because condition" [F(1, 104) = 6.62, p =.01]. However, the difference between the "and" condition" and the "because condition" did not reach significance [F(1, 104) = 2.65, NS]. This finding suggests that causal inferences are activated during reading. The with or without connective effect is not significant [F(2, 85) = 2.51, NS] although the verb completion scores are higher in the without connective condition obtained in the three experimental conditions. In addition, the interaction did not reach significance [p > 1]. This result fits the connective interference-based interpretation proposed by Millis, Graesser and Haberlandt (1993) about connectives and memory for text.

Recall data: To be categorized as inferential recall, the subject had to write the inferential verb or expression corresponding to the cue rather than the sentence that was actually read.

Table 1. Mean inferential recall as a function of the three experimental conditions.

For exemple, Table 1 indicates the mean number of inferences that the subjects recalled. It is worth to notice that in the "because condition", the first clause refers to the consequence of the action while the second clause refers to the cause. The opposite pattern happens in the "and & therefore conditions".

The results showed that inferential recall score is highest for the second clause whatever the experimental condition. We observed that in the "because condition", the subjects failed to recall the inferential consequence of the action. Relationship between Comprehension and Memory for Text A very few number of inferences activated during reading are recalled in the unexpected cued recall test. However, we noticed that a great proportion of inferences mentioned in the recall test were not activated during reading. This finding shoud suggest that causal inferences are briefly activated during reading rather than integrated in the representation of what the text is about.

The present study examined the relevance of the distinction between two processes involved in the generation of predictive inferences. First, the activation processes which activate meaning that are associated to text information in readers semantic memory. Second, the integration processes by which the activated information becomes a component of readers' episodic text representation in long term memory. That distinction is central to the Construction-Integration Model (Kintsch, 1988) and to the Landscape-Model (van den Broek, Young, Tzeng & Linderholm, 1999). In four experiments, subjects read the texts, sentence by sentence, at their own pace. We first collected activation measures of predictive inferences meaning in a lexical decision task. Second, we collected plausibility evaluation of the predictive inferences with a judgment task: Subjects had to judge if the predictive inference expressed in a sentence corresponded to an event that will probably happen next according to the text they read.

In experiment 1 (cf. Table 1), the activation measures were obtained by comparison with a condition which controlled the effect of lexical priming by text words. According to the data, the predictive inferences were activated to a same level when they were only associatively related to the content of a contextual sentence that preceded a predictive sentence (Associative version) or when they were causally related to the content of that contextual sentence (Causal version). On the contrary, the predictive inferences were not activated when the predictive sentences were preceded by a contextual sentence with words unrelated to the predictive inference (Simple version). Thus the predictive inferences appeared equally activated when supported by the same amount of semantic associations. However, the judgment task, performed after the reading of all the texts, indicated that the plausibility of the predictive inferences was evaluated as higher for the Causal than for the Associative versions. Thus, the amount of causal relations between the text events and the predictive inference might determined the integration of these inferences in readers episodic representation of the text.

In Experiment 2 (cf. Table 2), subjects performed the judgment task after the reading of each text plus three neutral sentences which followed the predictive sentence. The latencies of correct responses were longer for the Associative versions than for the Causal versions. That result again suggests that the predictive inferences were not strongly integrated to readers causal representation of the text events, contrarily to what had already happened when the Causal versions were read.

Experiment 3 (cf. Table 3) and Experiment 4 (cf. Table 4) respectively replicated the results of Experiment 1 and 2 and also showed that the predictive inferences were activated when only the contextual sentences of the former Associative and Causal versions were read (C-Associative and C-Causal versions). However, the judgment data indicate that the predictive inferences were not considered as plausible continuation of the C-Associative versions. Thus, when the contextual sentence were processed alone and were associatively but not causally related to the predictive inferences, these inferences were activated but presumably not integrated in readers' situation model (C-Associative versions). On the contrary, the predictive inferences were not only activated but presumably integrated in readers' situation model when they were causally related to text information (Causal Contextual version). Thus, applied to predictive inference generation, the distinction between activation and integration processes gives a coherent explanation of the data. Although clearly distinct from a lexical priming effect, a predictive inference would be transitorily activated via semantic associations (Keefe & Mc Daniels 1993) and integrated to readers representation of the text content (Klin, Guzman & Levine 1999). Our result allowed to precise that the activation level of predictive inferences mainly depends on the strength of semantic associations with the inference that are provided by text words. Moreover, the causal relation that can be establish between the associated text information and the inference are not crucial for this activation processes. When these relation are missing, the predictive inference is probably activated as a general expectation. However, our judgment data suggest that the predictive inferences can be integrated at various degree in readers episodic representation of the text content. They also suggest that the integration processes are governed by the causal relation that related the text events to the inference in the situation described by the text.

Table 1. Mean latencies (ms) of positive responses in the lexical decision task (L. D.) and mean percentages of positives responses in the judgment task collected for each text version.

Table 2. Mean percentages and mean latencies (ms) for positive responses in the judgment task collected for each text version.

Table 3. Mean latencies (ms) of positive responses in the lexical decision task (L. D.) and mean percentages of positives responses in the judgment task collected for each text version.

Table 4. Mean percentages and mean latencies (ms) for positive responses in the judgment task collected for each text version.

It becomes apparent that the comprehension of written directions must be conceived of as involving a different processing according to the reading goal. For example, cognitive activities brought into play during comprehension involve a preferential processing of the semantic structure whereas those developed during reasoning require the construction of a specific situation model (Kintsch, 1994; Perrig & Kintsch, 1985; Schmalhofer & Glavanov 1985). In the latter case, information provided in the text is matched with background knowledge and integrated with it (van Dijk & Kintsch, 1983). One of the special features of written directions is that the instructions are conveyed in a general and condensed way, even for a particular task. This is based on the fact that written directions generally define macro-actions or general categories supposed to cover all the cases that could be encountered. This is especially apparent with pharmaceutical instructions. Indeed, they often express rules that can be applied to categories of sick persons. Theses rules correspond to a dosage of the drug and to frequencies of application. Generally, they are expressed in a form of intervals. In order to elaborate a particular situation model, one needs to translate intervals given in the text into the form of individual-dose amounts of drug. This translation, which requires the generation of many inferences, leads to the definition of the result to reach. According to Richard (1990), this stage of information processing corresponds to the elaboration of a declarative representation. This is the basis on which one constructs a procedural representation of actions to bring into play. However, the conversion of the representation of the result to attain into "effective procedures" is not direct (Georges, 1988): It is frequently necessary to produce inferences in order to understand and to act on the situation, to define plans, and to elaborate action sequences. Our aim was double : a) To analyse the process underlying the construction of a declarative representation during comprehension instructions in order to use a drug ; b) to study the nature of the required representation for the planification of actions which are necessary to accomplish the task. In order to reach these two aims, we manipulated the reading goal and asked subjects to think-aloud.

All subjects (N = 26) received the same experimental text: The dosage and the directions for the use of a drug. The only difference between the groups laid in the reading goal given in the experimental instructions. In a first group (n = 8), subjects were asked to read and to understand in order to relate what they had read. In a second group (n = 8), subjects were asked to read and to understand in order to prescribe the drug to a depressively ill patient. Furthermore, we asked these two groups to think aloud, that is, to verbalise anything that came to mind when reading a sentence. In a third group (n = 10), we asked a) to read and to understand, b) to carry out the instructions and, simultaneously, c) to verbalise every action brought into play. The verbalizations achieved by the subjects were recorded on a tape recorder. This allowed us to collect verbal protocols from which we elaborated a classification of inferences EFFECTIVELY generated according to the reading goal and supposed to play a part in the construction of each representation.

When it is question to read in order to relate back or to read in order to prescribe, subjects expressed verbalizations which testify the generation of bridging-inferences. We have also found verbalizations which attested that all subjects of these two groups generated micropropositional inferences. These sorts of inferences, elaborated from world knowledge, enabled the specification of textual information and are inherent in the nature of the text itself. Moreover, subjects explained constraints given in the instructions through the generation of causal inferences. Only subjects who read in order to prescribe generated superordinate goal inferences. This kind of inferences is closely linked to the goal that motivates reading: To prescribe the drug (in order to treat). All subjects who read in order to prescribe or to carry out instructions generated particularization inferences. For example, from an information like "4 or 6 tabloids 2 or 3 times daily", subjects inferred that the appropriate dosage for the patient was "2 tabloids in the morning, 2 tabloids at noon, 2 tabloids in the evening". Constructing these inferences implies that relevant information would insert into pre-existent slots of pragmatic knowledge possessed by the reader. Subjects of these two groups generated pragmatic inferences. These are also generated from schemata, but, unlike particularization inferences, they are not directly linked to textual information. Therefore, they seem to depend on an implicit reading goal: To give the application method of the drug. Like particularization inferences, new information indicates the intended result, but it provides no indication about actions to be brought into play in order to satisfy this result. Only subjects who read in order to carry out instructions verbalized inferences related to the planification of actions, the role of which is to specify the overall actions to perform in order to attain the goal. The structure and the content of verbalizations attested to the programming of actions required for the fulfilment of pre-requisites (i.e., actions that must be carried out first in order that the actual procedure could take place) and the construction of sub-goals in working memory. Also, as soon as a sub-goal was defined, subjects planed the sequence of actions to perform in order to satisfy it. We noted that particularization inferences and inferences about the planification of actions come from the same linguistic data; this allows us to make the following two observations: (a) subjects who read in order to carry out instructions elaborated two distinct inferences from the same information; (b) depending on whether the goal was to prescribe or to carry out instructions, the inference constructed from a same textual data depends upon a different point of view.

The originality of this experiment is that it permits to test experienced inferences, namely, inferences that are not created by the experimenter, but inferences really identified in subjects verbalizations. It allows us to specify differences in inferential production and in the mode of construction of the representation, which depends either especially on semantic structure or on a situation model (Kintsch, 1994). More precisely, comprehension / memorisation activities involve a preferential processing of the semantic structure of the text. Here, inferences required for textual information comprehension and for the establishment of coherence are primarily generated. When the goal is to prescribe a drug, subjects construct a situation model in the form of a declarative representation. Inferences enable to derive the result to attain from information given in the text. When the goal is to perform actions, subjects construct a procedural representation on the basis of a declarative one. Also, they elaborate inferences about goals to reach and action sequences to bring into play in order to satisfy these goals. Comprehension of a procedural text can be conceived as a variable depth process. In this sense, the procedural representation of actions to perform is based on a declarative representation of the result to attain, which is itself based on the semantic structure of the text.

Recent research has shown that negation can automatically suppress the activation of concepts in a reader’s mental representation of a text (e.g., MacDonald & Just, 1989; Gernsbacher & Jescheniak, 1995). But how does negation affect inference making? In three experiments, we investigated the role that negation plays in the making of deductive inferences during reading. According to at least one account of logical competence, people are able to draw certain propositional-logic inferences easily and accurately (Braine et al. 1995). This claim has held for performance on elementary logic problems, and recently it has been extended to the domain of discourse comprehension (Lea et al. 1990; Lea, 1995). The present work examined two logical inferences: One that yields a positive conclusion (or-elimination: a or b; not a; therefore b) and one that leads to a negated conclusion (not-both elimination: not both a and b; a; therefore not b). According to data collected by Braine et al. (1984) these two inference forms were made equally easily when participants were working with abstract content, but the question arises as to whether the effect that negation has on propositions goes beyond explicit negation. That is, does negation inhibit or suppress the production of negative information?

In Experiment 1 we presented texts in which participants could make the or-elimination inference or the not-both inference. In addition, four control conditions were included in which no inference could be made. Word pronunciation (naming) response times were collected for items related to the inference concept. For example, in a passage about a student who is deciding whether to take History or English, and then later decides not to take English, the inference concept HISTORY was presented as the naming probe. In the not-both version of this text the decision about which course to take is presented in the form of someone who cannot take both History and English, and who later decides to take English. In this case the inference is a negated proposition - that she did not take History. Verification times to comprehension questions about the protagonist’s decision also were collected to provide converging evidence for inference making. Both naming and verification results indicated that readers were making the or-elimination inference, though these measures were mixed for the not-both inference. That is, the naming times showed no priming for the inference concept, while the verification times indicated that the inference had been made during reading of the text. This pattern is consistent with the possibility that readers were making the not-both inference while reading, and that the negated inference concept was suppressed, as MacDonald and Just (1989) had found in a naming task of explicitly negated concepts.

An alternative explanation for the results of Experiment 1 is that a change in discourse focus - not inference making - produced the naming time results. In other words, when the naming probe appeared, the discourse focus was shifting toward the probe concept in the or-elimination case but not in the not-both case; Perhaps the faster naming times in the “or” stories simply reflects the discourse focus at the time of test. So in Experiment 2 we presented the premises in reverse order so that no change in focus occurred at the time when the inferences could be made. For example, the passage about a student deciding which course to take was changed so that she began knowing that she had to take English, and then later discovered that she couldn’t take both English and History. The inference is that she wouldn’t take History. The pattern of results was the same as Experiment 1, thus eliminating discourse focus as the critical factor. As in Experiment 1, the naming and verification latencies were mixed with regard to the not-both inference. Specifically, the naming times for not-both inference concepts were not different from controls (suggesting that no inference had been made), but the pattern of verification latencies indicated that the participants had indeed drawn the inference. Again, this paradoxical result is consistent with the hypothesis that the inference (e.g., not History) was made less accessible because it was negated.

In Experiment 3 we examined reading times for sentences that contained information that was inconsistent with the inference concept. Reading times were a logical choice because previous research has shown that negation does not seem to influence reading time (e.g., MacDonald & Just, 1989). In Experiment 3, we composed passages in which certain target sentences would make sense only if the inference in question was NOT made, but would appear contradictory if the inference was made. This technique has been exploited to investigate a variety of text processing phenomena (e.g., Albrecht & Myers, 1995; Lea, Mason, Albrecht, Birch, & Myers, 1998). For example, in a passage about two brothers making Chili, the recipe states that "they could not use both fresh jalapenos and ground pepper in the chili." When they decide to use fresh jalapenos, the reader can infer that they will not use ground pepper. The target sentence reads "I'll get the ground pepper said Joseph", which is inconsistent with the inference. The only difference between the inference version and the no-inference control version of the passage is the omission of the word "not" in the second sentence. The local coherence of the target sentence is identical (and reasonably good) in both the inference and control versions. The target sentence is only inconsistent when the not-both inference is made. Thus, if readers are in fact making the not-both inference, their reading times should be slower on the target sentence compared to the no-inference control. The reading time data from Experiment 3 showed a very large slowdown on inference passages, which clearly indicates that participants were making the not-both inference. Reading times on the pre-target sentence ruled out a spill-over explanation of the results.

Together, these three experiments present a complex and potentially interesting picture of the role that negation might play in inference making. First, negation does not seem to inhibit the inference making process- at least not in a way relevant to the current inferences; Experiment 3 established that readers were making the negated not-both inference. Second, it appears that negation does affect the inference after it is made. Specifically, the present data indicate that inference concepts that are negated are less accessible than comparable inferences that are not negated. This result is parallel to MacDonald and Just’s (1989) finding that explicitly negated concepts (as opposed to inferred negated concepts) were less accessible from the readers mental representation of the text. MacDonald and Just explained their effects in terms of discourse focus. However, there appears to be more going on in the present passages than just focus shifts. By separating focus shift from presentation of the second premise, Experiment 2 established that the priming of the or-elimination inference targets was not due to a shift in discourse focus. In addition, Experiment 3's demonstration that readers were in fact making the not-both inference shows that the inference mechanism takes precedence (at least chronologically) over whatever mechanism drives focus shift. In other words, the inference is made first, and then the focus shift, or suppression mechanism, is activated. Thus, the inference is made even though it yields information that is not particularly relevant to the discourse focus. These results fit well with discourse comprehension theories such as Gernsbacher’s Structure Building Framework (e.g., Gernsbacher, 1990), and Kintsch’s Construction-Integration model (Kintsch, 1988). The present work also further establishes logical deductions as part of the common repertory of inferences that people make in everyday cognitive processing.

It is traditionally accepted that text comprehension is a constructive process which results in the elaboration of a global representation of what the text is about. The construction of this representation requires the activation of both the concepts explicitly mentioned in the clauses and implicit information derived from general world knowledge. This implicit integration during reading is based on inferential processes. In our study we focused on inference activation and especially on predictive inferences. They refer to the on-line activation of consequences of events or actions described in the text. Predictive inferences are usually not thought to be activated on-line because they are not required to make the text coherent and to understand it. However, recent studies showed that they can be activated on-line in certain circumstances (Keefe & McDaniel, 1993; Whitney et al., 1992). Several authors (Graesser et al., 1994; McKoon & Ratcliff, 1992) proposed that their activation can be viewed as a continuum: Predictive inferences are activated at different degrees depending on whether the context is strong or weak. In this view, we wanted to show that text characteristics can lead to different degrees of activation. We worked with two types of context based on two types of causality: Intentional causality vs physical causality. Type of causality is seldom controlled in studies. We thought that this lack of control can explain divergent results. First of all, we hypothesized that a stronger context provided by the intentional causality can lead to reinforced the probability of occurrence of consequent events: In this way, the inference degree of activation would be greater than in a weaker context provided by simple physical causality. In fact, the consequent event is caused and expected by an acting agent in the first case, and only caused by a physical agent in the second case. Our aim was also to confront different methodologies involved in predictive inferences study. We first tested inference activation in two ways: On-line and off-line. We also used two different on-line tasks: A word completion task and a lexical decision task.

Two types of text are used for the whole experiments. They describe scientific mechanisms about physic phenomenon (physical causality) and objects fabrication (intentional causality). In this way we have a human agent specified in the texts concerning the making of objects (paper, plastic, glass) with the intentional idea compared with texts without any intention (i.e. any human agent) about physical, natural mechanisms (lightning, erosion, volcano). Seven experimental texts are constructed in the same way: 3 introductory sentences and 2 sentences with the same agent as the grammatical subject (human agent: chemist, artisan, workman vs natural agent: water, lightning, rocks, waves) - the last sentence leading to a predictive inference. The 9 inferences are first tested off-line in a protocol task in which 70 subjects are told to read the texts and to answer the question "What happened next ?" which was placed after each sentence leading to a predictive inference. In order to test inference activation during reading two different tasks were used. In the first experiment, we employed a completion task which involves providing the first two letters of the inference target word. Readers were told to complete those items with the first verb that comes to mind and begins with those letters. If the predictive inference is activated on-line, the complete verb would be the one that corresponds to the inference.

The second experiment used the same exact method as the first, but a lexical decision task is used. The same verbs are presented to readers who were told to decide if they are French verbs or not. Three versions of each passage were created in order to evaluate inference degree of activation. Explicit and implicit versions were identical except for the last sentence verb: The inference verbs were provided in the explicit passage. A third control version was created with the same words and context but without leading to the inference. This control version was used for the lexical decision task. For the first task, baseline rate for completing inference verbs were integrated as control condition: 20 other subjects were asked to complete the items without reading the texts.

Results from off-line and on-line experiments are presented and compared in order to debate on our interest: Predictive inferences. Predictive inferences were activated in the off-line protocol. A qualitative effect of type of context was obtained: Subjects answers reflected the two different causalities (intentional and physical) through goal markers as connectives in one hand and probability markers on the other hand. The type of context had however no effect on predictive inferences on-line activation: The expected facilitation effect of intentional contexts was not obtained both in the completion task (F (1, 60) = 1.43, p < .23) and in the lexical decision task (F (1, 56) < 1). Completion scores were too low to conclude that predictive inferences were activated on-line, and the difference between implicit and control versions only tend to be significant (F (1, 60) = 2.8, p < .09). In the same way, the difference between lexical decision times in the implicit and control versions was not significant (F(1, 56) < 1). Predictive inferences did not seem to be activated on-line. The comparison between results from off-line (protocol) and on-line (completion and lexical decision tasks) methods of study led us to enhance several problems: First, compared with completion task, our lexical decision results seemed to support the existence of post access effects involved by this task and already criticized by many authors (Potts et al., 1988; Whitney et al., 1992). We saw that predictive inferences can be made off-line when they are explicitly incited: In general reader's answers at the question "What happen next ?" provided different words to express a same consequence. We selected the more frequently used word as the target-verb for the two on-line tasks. But we wonder whether the inference can be reduced to a single word. This question leads to a larger debate concerning the nature of inferences and text representation. First of all, with reference to McKoon & Ratcliff (1986) position, we wonder if predictive inferences can be activated on-line but in a minimal way without corresponding to the exact presented word-item. In the same way, Kintsch (1988, 1998) suggests that reading leads to the build of a propositional representation of text: So we have to wonder if the proposition corresponding to the inference which could be activated on-line is really reflected by a single verb. Fincher-Kiefer (1993) also suggested that predictive inferences are integrated during reading at a more abstract level of representation: The situation model. Because of the variety of words used by subjects to express the inferences, our results from verbal protocol support the fact that they could be made in such a more abstract format. Kintsch (1998) also proposed that the situation model might be an image in certain circumstancies. In this case, the presented verb could not exactly and completely represented the created image during reading. By consequent, the inference could be made on-line without facilitating the treatment of the verb presented for the tasks.

During the last two decades, there has been significant research on inference generation during reading. Whereas many investigators have used speeded techniques, in which one element is probed for activation during reading (e.g. Graesser, Singer & Trabasso, 1994; see Haberland, 1994 or Singer, 1994 for a review), others have chosen verbal protocol methods such as question answering (e.g. Graesser, 1981; Graesser & Clark, 1985; Long, Oppy & Seely, 1997) and think-aloud tasks (e.g. Trabasso & Suh, 1993; Trabasso & Magliano, 1996; Zwaan & Brown, 1996; see Pressley & Afflerbach, 1995, for a review). In think-aloud research, readers have been found to engage in a variety of cognitive processes during reading, including strategic inference-making, text integration, and text interpretation (Pressley & Afflerbach, 1995). In particular, think-aloud methods have revealed inferences and mental processes that occur during comprehension (Olson, Duffy, & Mack, 1980; Fletcher, 1986; Suh & Trabasso, 1993; Trabasso & Magliano, 1995, 1996; Narvaez, van den Broek & Barrón, 1999). In addition, think-aloud tasks show the information available in working memory as well as the information that can be used strategically and is codable in language (Ericsson & Simon, 1994).

The purpose of this study is twofold. On the one hand, we analyze the effects of text type (expository in contrast to narratives and newspaper articles) on the expectations and inferences activated by university students during a think-aloud task. Traditionally, research has focused almost exclusively on two types of text - narrative and expository. The results of this research have been mixed. Some authors suggest that there are similarities in the processing of narratives and expository texts (e.g. Goldman & Varma, 1996) and literary texts (van den Broek, Rohleder, & Narvaez, 1996), whereas others propose that readers spontaneously set in motion different patterns of activation or inferences according to the type of text (Brewer, 1980; Harris et al., 1998; Mc Daniel, Einstein, Cunay & Cobb, 1986; Mc Daniel, Owen & Cote, 1990). For example, some researchers have suggested that narratives are more likely to generate predictive inferences and that expository texts lead to more text integration (Olson, Mack & Duffy, 1981). A second objective is to determine if there is a universal cross-cultural inference-making process or if linguistic differences produce varying inference patterns. There exists very little research on this issue. A few researchers have attempted to analyze cultural differences in the context of telling stories (Chafe, 1980; Tannen, 1980, 1983, 1984, 1988). Other authors investigated memory for narratives by North-Americans and Liberian (children and adults, literate and non-literate) observing similar behaviors (e.g. Mandler, Scribner, Cole & De Forest, 1980).

Most constructionist frameworks adopt the idea that a finite set of inference categories can be produced during comprehension (e.g. Graesser et al., 1994; van den Broek, Fletcher, & Risden, 1993). Trabasso and Magliano (1996) proposed three types of inferences that could be activated during the comprehension of a sentence: (1) Explanations, (2) predictions, and (3) associations. In addition, they attempted to relate the various types of inferences to two other important aspects of comprehension: Sources of information and memory operations. With respect to the first, an inference can be based on the text or on relevant world knowledge. With respect to the second, inferences are related to three memory operations that make information available for inferences. Thus, information can be a) maintained in working memory across story sentences (e.g. for explanation inferences), b) retrieved from a long-term memory representation of the text (e.g. for predictive inferences), or c) activated from general knowledge (e.g. for association inferences). We hypothesize that each of these memory operations is not necessarily linked to one type of inference but rather than memory operations and inferences are dynamically interconnected.

A total of 40 subjects (20 North American and 20 Spanish university students) took part in this research. We used 20 short texts of 4 different types: 3 fairytales, 6 complex stories, 4 newspaper articles, and 7 expository texts. The texts had been taken from newspapers, novels, encyclopedias, handbooks, and stories used in previous experiments by other researchers. The texts were presented to North-Americans and Spanish students in English and Spanish respectively. Subjects performed a think-aloud task as they read the texts. To categorized the subjects´ responses we used the categories proposed by Zwaan & Brown (1996) and Narvaez, van den Broek and Barrón (1999). All the texts were calibrated in length and causal structure according to the model proposed by Trabasso and Magliano (1996). No explanation was given regarding the kind of texts that the subjects were to read. Texts were presented in random order.

Genre was found to be an important factor in the determining the kinds of inference produced (p < .001). The main findings are the following. First, it influenced the generation of explanatory inferences: Expository texts required a greater number of explanations than did newspaper articles (p < .05), whereas complex stories required fewer explanations than did fairytales (p < .05). No other differences in the number of explanatory inferences generated in the various text types were significant. Second, text generation influenced the generation of predictions: Complex narrative texts gave rise to a significantly greater number of predictions than did newspaper articles and expository texts (p < .001), with expository texts eliciting fewer than newspaper articles (p < .05). No other differences in predictive inferences were significant. Third, newspaper articles and expository texts produced a greater number of comprehension problems that did narrative texts (p < .05). No significant differences amount text type were observed for other inferential activities. Also, no differences were observed between North-American and Spanish students.

In summary, expository texts evoked more explanatory inferences and comprehension problems than other text types, whereas narratives gave rise to a larger number of predictive inferences. Newspaper articles elicited a pattern of inferences that held the middle between those for expository texts and the narratives. Thus, different types of text evoke different working memory operations (Trabasso & Magliano, 1996). Expository texts require more activation of relevant knowledge, more maintenance of information in working memory, and perhaps when this is incomplete, more comprehension problems could be produced. These patterns can be explained in terms of differences in reader's prior knowledge, in structure of the various text types, and in the nature of the contents of the text itself. These results also support the idea that the subjects implicitly realize the kind of text that they are reading. In fact, no explanation was given regarding the kinds of texts that the subjects were to read. This represented a significant departure from previous experiments (e.g. Narvaez, van den Broek & Barron, 1999). The fact that these pattern were observed in both North-Americans and Spanish students indicates that the inference-making process is not restricted by linguistic and/or cultural differences.

Whether the instrument inference was generated during comprehension has been the controversial topic among discourse researchers. Although Kim, Lee, Lee, and Lee (1998) provided a convergent evidence for the on-line generation of instrument inference, recent studies showed that individual differences in working memory span resulted in differential inference processing (Kim, Lee, Kim, & Yoo, 1999). This study was conducted to investigate the individual differences in instrument inference using lexical decision task (Experiment 1) and picture naming task (Experiment 2).

Participants. 24 undergraduate students (10 low span readers and 14 high span readers) in Experiment 1 and 30 undergraduate students (14 low span readers and 16 high span readers) in Experiment 2 were selected on the basis of their scores on the reading span test. Low span readers were defined as those whose reading span score was 2.0 or less and high span readers 4.0 or more.

Design. 2 (high span, low span) x 4 (explicit, implicit, alternative, control) mixed design was used.

Materials. Reading span test: The materials for the reading span test were based on Daneman and Carpenter (1980). The total of 175 unrelated sentences were divided into 7 sets, with five trials at each span size. Each sentence ranged in length from 7 to 10 words. The reading span score was defined as the highest set size for which the participant correctly recalled all of the words from three out of five sets.

Experimental Materials. A set of 32 experimental sentences was constructed with four versions each. In order to select appropriate instrument in each sentence context, forty-seven students not involved with the experiments were asked to generate the most plausible instrument given particular sentential context. From a list of 61 sentences, 32 instrument-sentential context pairs were selected. The average generation rate of the selected instruments with particular sentential context was 92%. In explicit condition, the instrument was explicitly mentioned in the sentence whereas the instrument was not mentioned in implicit condition. In alternative condition, the improbable instrument was mentioned instead of probable instrument. The control sentence was constructed in order to equate the effect of word-based priming based on semantic association among target words and predicates or other words.

Procedure. Participants read sentences that were presented one word at a time on the center of a computer monitor with RSVP rate of 350 ms plus 15ms per character. In Experiment 1, the participant's task was to decide as rapidly and accurately as possible whether the target word was a Korean word or not. In Experiment 2, participant's task was to say the name of the picture out loud as soon as the picture appeared.

In Experiment 1, a 4 x 2 ANOVA revealed a significant main effect for experimental condition, F1(3,126) = 23.65, p < .001, F2(3,93) = 10.66, p < .001. Although the main effect for reading span was not significant, the interaction effect for reading span by experimental conditions was significant, F1(3,126) = 3.80, p < .05, F2(3,93) = 2.59, p = .058. For high span readers, the lexical decision times for implicit condition were significantly faster than alternative and control condition whereas there was no significant difference among the three conditions for low span readers. In Experiment 2, it was found that significant main effects for both experimental condition, F1(3,90) = 44.52, p < .001, F2(3,93) = 34.33, p < .001, and reading span, F1(1,30) = 7.19, p < .05, F2(1,31) = 47.40, p < .001. The naming times for explicit and implicit conditions were faster than alternative and control conditions for both high and low reading span readers and low span readers were significantly faster than high span readers. There was no significant interaction effect between reading span and experimental conditions.

The results of Experiment 1 indicated that low span readers did not generate instrument inference on-line whereas high span readers did. However the results of Experiment 2 showed that both high span readers and low span readers generated instrument inference on-line. These contradictory findings can be resolved by differential sensitivity between lexical decision task and picture naming task. It is suggested that lexical decision task is an appropriate task for detecting linguistic inferential processing while picture naming task might be appropriate for reflecting perceptual aspect of inference. Since reading span reflects some kind of verbal ability or processing of phonological loop in working memory, low reading span readers might have some difficulty to generate verbal instrument inference. On the other hand, since picture naming task taps perceptual aspect of instrument inference, which is not directly related to the reading span, low span readers generate perceptual inference on the instrument.

Table 1. Mean Response Times in Lexical Decision Task and Picture Naming Task (ms)

Daneman, M. A., & Carpenter, P. A. (1980). Individual differences in working memory and reading. Journal of Verbal learning and Verbal Behavior, 19, 450-466.

Kim, S., Lee, J. M., Kim, S. Y., & Yoo, H. J. (1999). Individual differences in discourse comprehension: The interaction between reading span and text structure. Society for Text and Discourse: Ninth Annual Meeting of the Society for Text and Discourse, Vancouver, B.C., Canada.

Kim, S., Lee, J. M., Lee, J. H., & Lee, K. (1998). Does instrument inference occur on-line during reading? Society for Text and Discourse: Eighth Annual Meeting of the Society for Text and Discourse, Madison, WI.

The results of several studies suggest that the right and left hemispheres may process inferences differently. First, studies with right hemisphere damaged (RHD) patients have shown that patients with lesions in the right hemisphere have deficits in inference generation. Some RHD patients have longer response times to inference-related words than to unrelated words when compared to controls (Beeman, 1993). Second, split-visual field studies with normal participants show priming of specific types of information in the right and left hemispheres. The right hemisphere shows priming effects for more distant semantically related words (Beeman, et al., 1994). In contrast, the left hemisphere shows priming effects for more closely related associates. These two studies suggest that the right hemisphere may have an important role in making certain type of inferences during text comprehension. One such type is that of a predictive inference, in which readers generate expectations about what will happen next in the text. The reason that one might expect this type of inference to be processed differentially by the two hemispheres is that a critical parameter in the generation of predictive inferences is the amount of semantic constraint for it. In particular, the degree of causal constraint determines the likelihood of a predictive inference (van den Broek, 1994). For example, predictive inferences will be made more often if the text includes sufficient antecedents. This is also reflected in theoretical models of text comprehension (e.g., the Causal Inference Maker, van den Broek, 1990). Thus, the level of textual constraint may influence the production of predictive inferences.

The aim of the present study is to examine the interaction of causal constraint and the hemispheric location of processing during predictive inferences. Findings from brain damaged studies, split visual field studies, and theories of causal inference generation lead to interesting ideas about how the brain may process inferences. These findings suggest that there could be differences in the right and left hemispheres in the way they process predictive inferences, especially when the level of constraint is manipulated. If this is true, then we hypothesize that the right visual field - left hemisphere (rvf-LH) should be faster in responding to highly constrained predictive inferences. In contrast, the left visual field - right hemisphere (lvf-RH) should be faster in responding to low constraint predictive inferences.

Forty-eight inference items were constructed that consisted of a high constraint and low constraint version. The items consisted of 4 sentences that were constructed so that the predictive inference might be generated after reading the last sentence in the item. For both the high and low constraint items, the target words presented to the participants in the lexical decision task were both words associated with the inference. Forty-eight neutral items, that did not promote these inferences, were constructed as a baseline condition for each hemisphere. The neutral items were paired with the same target words as those presented in the inference items. In addition, 48 filler items that did not promote any inferences and were paired with non-word targets were constructed.

After several practice trials, participants read each item on the computer one sentence at a time. Then, a fixation plus was presented in the center of the screen (for 750 milliseconds) to focus the participants' attention. Next, a lexical decision task was presented to the participants. The probe word or non-word was presented either to the participants' left or right visual field for 176 milliseconds. All words appeared approximately 3.5 degrees of visual angle to the left or the right of fixation. The center of each probe was 3.4 cm from the fixation center, and the inner edge of each word was at least 1.5 cm from the outer edge of the fixation point. Participants were instructed to press the index finger on the middle button of the button box for a "YES" response, and press their thumb or first button for a "NO" response. Also, participants were asked to recall the main idea of some of the items they read in 1-2 sentence in order to assure comprehension.

The main variables were visual field (left or right) and constraint (low or high). The dependent variable was response time to a lexical decision task. Only items that were correctly answered in the lexical decision task were included in the analysis. A main effect was found on the constraint variable. As expected, high constraint stimuli evoked faster responses than low constraint stimuli. Importantly, an interaction effect was found between visual field and constraint. A priori contrasts established the location of this interaction. Similar levels of priming were found for the highly constrained items in both the lvf-RH and the rvf-LH. In contrast, the low constraint items showed a greater level of priming in the lvf-RH than in the rvf-LH.

The results of this study suggest that the role of constraints play an important role when readers make predictive inferences. Also, the results show that the hemispheres work together to process different types of predictive inferences. Although the left hemisphere in general is faster at processing predictive inferences, this hemisphere appears particularly efficient at processing predictive inferences that are highly constrained by the text. In contrast, the right hemisphere is more efficient at producing predictive inferences that are less constrained by the text. These results indicate that the two hemispheres may play rather different, and complementary, roles in the comprehension of texts.

Beeman, M. (1993). Semantic processing in the right hemisphere may contribute to drawing inferences during comprehension. Brain and Language, 44, 80-120.

Beeman, M., Friedman, R. B., Grafman, J., Perez, E., Diamond, S., & Lindsay, M. B. (1994). Summation priming and coarse coding in the right hemisphere. Journal of Cognitive Neuroscience, 6, 26-45.

van den Broek, P. (1990). Causal inferences and the comprehension of narrative texts. The Psychology of Learning and Motivation, 25, 175-196.

van den Broek, P. (1994). The causal inference maker: Towards a process model of inference generation in text comprehension. In D. A. Balota, G. B. Flores d'Arcais, and K. Rayner (Eds.) Comprehension Processes in Reading. Hillsdale, NY: Lawrence Erlbaum Associates, 423-446.

This study deals with the multidimensional content of mental representation readers constructed from text. It is now well established that the model readers built reflects the multiple dimensions the situation implied: It has been demonstrated that events reported in a text are connected in memory through space, time, causation, intentionality and protagonists (Zwaan, Langston & Graesser, 1995; Zwaan, Magliano & Graesser, 1995; Zwaan & Radvansky, 1998; Zwaan, Radvansky, Hilliard & Curiel, 1998). In a recent study (Blanc & Tapiero, 1999), we have demonstrated that readers monitor the situational dimensions as a function of their prior knowledge, the task they have to perform and the comprehension phase of the model (i.e., construction and updating). In addition, we have underlined the complexity of the protagonists dimension. We showed that the characters personality traits and emotional reactions have to be considered as separate situational dimensions since readers differentially monitor these two dimensions as they proceed through the text. Issued from these results, we carried out another experiment not only to confirm what we previously observed concerning the role of readers' prior knowledge, task demands and comprehension phase on the monitoring of the situational dimensions, but also to test whether some links can be established between specific situational dimensions.

The material we used consisted in a text that described a play. We modified the experimental text we previously used (Blanc & Tapiero, 1999) to strengthen the characters personality traits and emotional reactions by providing more detailled information. The text was composed of two types of sentence, those that referred to the situational framework (i.e., the scenery of the stage) and those that dealt with the situational dimensions (i.e., space, intention, personality traits and emotion), assuming a differential monitoring of these two types of information (Zwaan & Radvansky, 1998). In addition, we divided the text into two parts to study how the situation model evolves all along the text, that is from the construction to the updating phase. Moreover, we constructed two versions for the second part of the text, the first one was consistent with the situation described in the first part of the text whereas the second one was inconsistent. Indeed, emotion and intention inconsistencies were introduced. We postulated that the processing of emotional inconsistencies should disrupt the representation of characters personality traits information being contradicted. We also expected that the processing of intention inconsistencies disrupt the representation of the characters movements since, in our study, the intention information corresponded to the description of a movement required to accomplish a specific goal. The reading of each part of the text was immediately followed by an inferences judgment task. Statements referred either to the situational dimensions or to the framework. Finally, to investigate whether readers' prior knowledge determine the monitoring of situational dimensions, subjects were provided with prior information on the situation via an introductory text. Before reading the experimental text, participants had to read a text that described either the scenery of the stage or the characters personality traits. Then, half of the subjects of these two conditions (i.e., scenery vs protagonists) had to draw a map of the scenery of the stage by precisely locating all the objects mentioned in the experimental text, whereas the other half had to summarize in writing the part of the text they just read. These two tasks were used to investigate whether the task demands could determine the situational dimension foregrounded. Then, all subjects performed an inferences judgment task. They had to judge whether the statement that appeared on the screen was true or false in relation to the part of the text previously read. The accuracy and the speed of inference judgments were recorded. Then, subjects had to read the second part of the text (either consistent or inconsistent) and performed the second testing phase similar (except for the material) to the one used after the reading of the first part of the text.

The results of this study contribute to a deep understanding of which factors determine the monitoring of the dimensions implied by the situation, and enable us to demonstrate that some links can be established between specific dimensions. First, our findings are compatible with those of previous studies, the multiple dimensions taking part of the situation model readers construct (Zwaan et al., 1995, 1998). Specifically, we obtained evidence for the fact that the monitoring of the situational dimensions evolves all along the text. In the first part, readers foregrounded information relative to characters personality traits and spatial location, whereas in the second part they focused on characters emotion and intention. Moreover, we observed that readers differentially process and integrate information relative to the situational dimensions as a function of their prior knowledge and the task they had to perform. Consistent with our previous study (Blanc & Tapiero, 1999), participants who already knew the scenery more accurately represented the spatial dimension whereas those in the protagonists condition were more accurate for the intention dimension. Also, we showed that the summary task promotes the monitoring of the emotion dimension whereas the drawing task facilitates the monitoring of the spatial dimension. More importantly, our findings support our assumption that the personality traits dimension shares links with the emotion dimension. Indeed, we noticed that the introduction of emotional inconsistencies in the second part of the text leads to a decrease in performance for the personality traits dimension. Consequently, readers integrated inconsistent emotional information and then encountered difficulties to retrieve the contradicted personality traits information. Moreover, we underlined that intention information is considered as important since participants detected inconsistencies relative to this dimension. Finally, our study confirm that the situational framework should be distinguished from the situational dimensions since readers differentially process these two types of information. First, it appeared that participants focused on the framework information during the construction phase whereas they shift their attention on the situational dimensions during the updating phase. Second, we demonstrated that, unsurprisingly, participants in the scenery condition constructed a more accurate and a more available representation of the situational framework than those in the protagonists condition. Finally, the situational framework was more accurately represented and more rapidly accessible when participants had to perform a drawing than a summary task.

Blanc, N., & Tapiero, I. (1999). Multidimensional situation models. Paper presented at the 11th Conference of the European society for Cognitive Psychology, Ghent (Belgique), 1-4 Septembre.

Zwaan, R. A., Langston, M. C., & Graesser, A. C. (1995). The construction of situation models in narrative comprehension: An event-indexing model. Psychological Science, 6(5), 292-297.

Zwaan, R. A., & Radvansky, G. A. (1998). Situation models in language comprehension and memory. Psychological Bulletin, 123(2), 162-185.

Zwaan, R. A., Radvansky, G. A., Hilliard, A. E., & Curiel, J. M. (1998). Constructing multidimensional situation models during reading. Scientific Studies of Reading, 2(3), 199-220.

The main purpose of our study was to investigate the multidimensional aspects of the situation model constructed from text (Zwaan, Magliano, & Graesser, 1995). Several studies have shown that readers represent more particularly information about spatiality, causality, intentionality, temporality and protagonist (Zwaan, Langston, & Graesser, 1995; Zwaan, Radvansky, Hilliard & Curiel, 1998; Tapiero & Blanc, 1998). In our experience, we were more specifically focused on the spatial and emotional dimensions implied in the situation described. Indeed, several researches have demonstrated that spatial dimension is difficult to monitor but, at the same time, appears to be necessary to represent the situational framework. Moreover, although Zwaan et al., (1995) did not take into account the monitoring of emotional dimension in the event-indexing model, Gernsbacher, Goldsmith and Robertson (1992) have demonstrated that readers are able to represent this kind of information, even when the characters’ emotional states are not explicitely mentioned. As the situation model results from the interaction between textual information and readers’ prior knowledge (Tapiero & Blanc, 1998; Zwaan et al., 1998), we also investigated whether the nature of readers’ prior knowledge (spatial vs. emotional) determine the monitoring of situational dimensions. First, we supposed that the situational dimension foregrounded in the model readers construct evolves from the construction to the updating phase. Furthermore, we made the assumption that the processing of the multiple dimensions varies as a function of readers’ prior knowledge.

To pursue these issues, we constructed a descriptive text, composed of two parts, the construction phase and the updating phase. Each part contained eighteen sentences among which we distinguished emotional and spatial sentences. In the construction phase, the spatial sentences described the location of characters, the furniture and the accessory, and the emotional sentences described protagonist nature traits. In the updating phase, the spatial sentences involved accessory and characters’ movement, and the emotional sentences involved the emotional reactions. To test whether readers monitor these two dimensions during the reading, the study of each part of the text was immediately followed by an inference judgment task. Participants had to judge six spatial and six emotional statements according to the part of the text they just read. In addition, to study the effects of prior knowledge on the processing of emotional and spatial dimensions, an introductory text, composed of nine sentences, was presented before the reading of the experimental text. One dealt with the spatial layout, presenting only the location of furniture (spatial knowledge group) and the other one with the characters’ emotional reactions (emotional knowledge group). According to the fact that spatial information is necessary to represent the situational framework (Zwaan & Radvansky, 1998), we predicted that the spatial dimension should be foregrounded during the construction phase whereas readers should focus on the emotional dimension during the updating phase. Moreover, we assumed that the monitoring of the dimensions should depend on readers’ prior knowledge (Blanc & Tapiero, 1998). Subjects with spatial prior knowledge should focus on the emotional dimension whereas those who already know the characters’ emotional reactions should focus on the spatial dimension. In addition, we expected that participants with spatial knowledge should be more accurate and more rapid to judge spatial inferences whereas those with emotional knowledge should obtain higher performances to emotional inferences for the first inferencejudgment task. On the other hand, the number of correct reponses should be more important for emotional inferences than spatial inferences for the second inferences judgment task, whatever the nature of prior knowledge given to participants. More importantly, we predicted that, during the construction phase, the monitoring of the emotional dimension is easier when readers have spatial prior knowledge than when they do not have prior knowledge on the situational framework implied.

Our results indicated first, that the emotional sentences were longer processed during the updating phase than during the construction phase whereas no difference between these two phases occurred for the spatial dimension. In addition, we observed that the emotional dimension was more accurately represented and more rapidly available than the spatial dimension, whatever the nature of prior knowledge provided (emotional vs. spatial). These findings are consistent with previous researches which have shown that spatial dimension is difficult to monitor compared to others dimensions (Zwaan et al., 1995, 1998) and confirmed that the situation model readers construct from text should be considered as a multidimensional representation. Moreover, two important points go beyond previous studies: First, unexpectively, the spatial dimension is relatively well processed during the updating phase and second, the emotional dimension appears to be a part of the multidimensional representation constructed by readers. Second, we noticed that subjects with emotional knowledge have shorter response times than subjects with spatial knowledge. Thus, emotional prior knowledge appeared to facilitate access to situational information. Finally, our results revealed that the situation model availability varies as a function of readers’ prior knowledge and of comprehension phase. Indeed, participants having spatial knowledge showed longer response times during the construction phase than during the updating phase, whereas those having emotional knowledge presented the reverse pattern of response. In others words, spatial prior knowledge seemed to facilite the access to situational information during the updating phase, whereas this facilitation appeared during the construction phase for subjects with emotional prior knowledge. This finding indicated that the retrieval of situational information depends not only on the nature of prior knowledge but also on the comprehension phase.

Our study gives arguments in favour of a multidimensional model of situation constructed from text. Futhermore, it shows the importance of the spatial dimension and especially of the emotional one, in the mental representation elaborated by readers. Therefore, it would be interesting to investigate this emotional dimension in interaction with other ones proposed by Zwaan et al., (1995) or to study the effects of the emotional valence of information (positive versus negative) on the reading processes. At the same time, the processing of the spatial dimension observed during the updating phase leads to rise a question about other factors that could influence this reading process, even if we can already set (following our results) the main influence of prior knowledge on information processing. Finally, as expected, the elaboration of the situational dimensions depends on the nature of prior knowledge (spatial versus emotional) given to readers. Moreover, this effect appears to be modulated by the comprehension phase (construction versus updating).

In the 70s and 80s, research on text comprehension has mainly focused on the nature of knowledge representation and its role in reading processes and memory organization. In the last 15 years the research field has widened. On the one hand, increasing attention has been paid to other cognitive aspects of comprehension, such as the reader's inference generation and search for coherence (e.g. Graesser, Singer, & Trabasso, 1994; Graesser, Gernsbacher, & Goldman, 1997; van den Broek, 1997). On the other, several scholars have underlined the limitations of a purely cognitive approach to comprehension and the need for considering the interactions of affective and cognitive components of reading (Miall, 1989; Gernsbacher, Goldsmith, & Robertson, 1992; van Oostendorp & Zwaan, 1994).

Contributions to a wider conceptualisation of text comprehension have come from various approaches, and particularly from research on interest and response to literature. Studies on interest (Hidi, 1990; Hidi & Baird, 1986; Renninger, Hidi & Krapp, 1992; Schraw, Bruning, & Svoboda, 1995) have highlighted the features which make a text interesting (novelty, curiosity, suspense, life themes, etc.) and the effects of interest on learning and memory. Research on response to literature (Beach & Hynds, 1991; Halàsz, 1996; Kreutz & McNealy, 1996;a Rosenblatt, 1983) analyses the reader's interaction with a text, that is his/her involvement, emotional responses, mental representation of characters' intentions and emotions, empathy with characters and the author. By focusing on the reader-text interaction, rather than on the separate aspects of text structure, on the one hand, and reader's processes on the other, this approach truly represents a new frontier in the psychological study of reading. In fact, in order to understand in depth the experience of reading, and particularly of reading literary narrative, the reader's subjective experience and emotions have to be considered and analyzed.

The aim of this study was to analyse the pattern of cognitive and affective-emotional components of adult readers' response to literary narrative reading. A basic assumption of this study was that these components are not separable in reading experience. Emotional states arise from text processing and affect the reader's coding processes and search for information in the text. Thus, they facilitate the comprehension process and make information easier to remember. The hypothesis of a close relationship between a reader's emotional response reading and his/her cognitive operations in reading was synthesised in Brewer's (1996) structural-affective model. Brewer has introduced a major distinction between event structure (the organization and sequence of events in the story) and discourse or text structure (the sequence of events narrated in the text). According to this model different reader's responses (interest, curiosity, surprise, suspense, etc.) may arise from different text structures. Thus, different versions (text structures) of a same narrative text should elicit different patterns of readers' cognitive and affective responses.

This study was aimed at investigating the ways in which the discrepancy between text and event structure affect: (1) Reader's cognitive-affective responses, such as interest, curiosity, surprise, suspense; (2) Readers' evaluation of text comprehensibility and memorability, as well as post-dictability and imagery; (3) Affective responses such as liking, emotionality and involvement. Moreover, a relevant research question concerns the interaction between these three aspects of response to literary reading.

Participants. One hundred thirty-five psychology students participated in the study.

Materials and procedure. Three versions of E. A. Poe's "The oblong coffer", a classic of gothic genre, were prepared. One was the original story, the second was the story with a flash-back which anticipated some crucial events, the third was the original one, with the conclusive sentence, which describes the intense emotional reaction of the narrator, moved up to the beginning. Each version, 12 pages long, was submitted to a group of readers (N = 45), who were asked to read the story and rate it on various measures which are typical of a reader’s main response types to a narrative text: (1) Their emotional states (involvement, empathy, etc.) at each critical point ; (2) Their cognitive states (predictability of the ending, imagery, importance of various elements of the story, etc.) ; (3) their affective-cognitive states, that is the states induced by a discrepancy between the available and expected information (interest, curiosity, surprise, etc.). Each of the answers was obtained through a six point scale (1: non at all interesting, comprehensible, etc. 6: very interesting, comprehensible, etc.)

As far as the answers given to the six point scale, a main result obtained through MANOVAs, was a confirmation of Brewer’s model, since some significant differences between the responses elicited by the three versions were obtained. However, this result, should be confirmed by an analysis of the reader’s response during the reading and non only after. A second result regarded the factor analysis of the cognitive, affective and cognitive-affective ratings, from which a complex pattern of results emerged, showing that the three aspects are highly interconnected.

Beach, R., & Hynds, S. (1991). Research on response to literature. In R. Barr, M. L. Kamil, P. B. Mosenthal, & P. D. Pearson (Eds.), Handbook of reading research (Vol. 2, pp. 453-489). New York: Longman.

Brewer, W. F. (1996). Good and bad story ending and story completeness. In R. J. Kreutz and M. S. McNealy (Eds.), Empirical approaches to literature and aesthetics (pp. 261-274). Norwood, NJ: Ablex.

Gernsbacher, M. A., Goldsmith, H. H., & Robertson, R. R. (1992). Do readers mentally represent character's emotional states? Cognition and emotion, 6, 89-112.

Graesser, A. C., Gernsbacher, M. A., & Goldman, S. R. (1997). Cognition. In T. A. van Dijk (Ed.), Discourse as structure and process (pp. 292-319). London: Sage.

Graesser, A. C., Singer, M., & Trabasso, T. (1994). Constructing inferences during narrative text comprehension. Psychological Review, 101, 371-395.

Halàsz, L. (1996). General and personal meaning in literary reading. In R. J. Kreutz and M. S. McNealy (Eds.), Empirical approaches to literature and aesthetics. Norwood, NJ: Ablex.

Hidi, S. (1990). Interest and its contribution as a mental resource for learning. Review of Educational Research, 60, 549-571.

Hidi, S., & Baird, W. (1986). Interestingness: A neglected variable in discourse processing. Cognitive Science, 10, 179-194.

Kreutz, R. A., & McNealy, M. S. (1996)(Eds.). Empirical approaches to literature and aesthetics. Norwood, NJ: Ablex.

Miall, D. S. (1989). Beyond the schema given: Affective comprehension of literary narrative. Cognition and Emotion, 3, 55-78.

Renninger, K. A., Hidi, S., & Krapp, A. (1992). The role of interest in learning and development. Hillsdale, NJ: Erlbaum.

Rosenblatt, L. (1983). Literature as exploration (4th edition). New York: The Modern language Association of America.

Schraw, G., Bruning, R., & Svoboda, C. (1995). Sources of situational interest. Journal of Reading Behavior, 27, 1-17.

van den Broek, P. (1997). Discovering the cement of the universe: The development of event comprehension from childhood to adulthood. In P. van den Broek, P. J. Bauer, & T. Bourg (Eds.), Developmental spans in event comprehension and representation (pp. 321-342). Mahwah, NJ: Erlbaum.

van Oostendorp, H., & Zwaan, R. A. (1994). Naturalistic text comprehension. Norwood, NJ: Ablex.

Recently, O'Brien, Rizzella, Albrecht, and Halleran (1998) examined the process by which readers "update" their knowledge of a story's protagonist. In their experiment, participants were timed in reading a sentence (e.g., "Mary ordered a cheeseburger and fries.") that described the protagonist executing a behavior that was either consistent with what was initially stated about the protagonist (e.g., Mary was described as a junk food addict), inconsistent with what was initially stated (Mary was described as a strict vegetarian), or was consistent with a qualification of the inconsistent information (Mary was a strict vegetarian, but occasionally ate greasy food when dining out.). The target sentence was read more quickly in the consistent condition than the inconsistent condition, replicating earlier research (Albrecht & O'Brien, 1993) showing that readers are sensitive to breaks in global coherence. Of primary interest was reading time in the qualification condition, where the target line was read more slowly than in the consistent condition but more quickly than in the inconsistent condition. The researchers interpreted this result as indicating that readers consulted both the inconsistent and the qualification information when interpreting the target sentence.

The focus of the present research was to explore the extent to which backgrounded information that has been updated is accessible later in a passage when protagonist-relevant information is presented. Specifically, is the backgrounded information automatically activated, as would be predicted by a passive, resonance process (Myers & O'Brien, 1998)? Or is it reinstated via a slower, more strategic process? Stated with respect to the example, the present study tested the extent to which "vegetarian" was accessible after reading "Mary ordered a cheeseburger and fries.". This is of primary interest in the case where readers' initial knowledge is subsequently updated (e.g., Mary is described as a strict vegetarian but one who occasionally eats greasy food). Although O'Brien et al.'s results indicate that this information was consulted, it is unclear from their results whether the original information was accessed automatically or strategically.

The present research repeated and extended O'Brien et al.'s experiment. Participants read passages one line at a time and were also timed in naming a word congruent with the inconsistent information (e.g., "vegetarian") immediately after reading the target sentence. With respect to reading times on the target sentence, a replication of O'Brien et al.'s results was predicted: the slowest target sentence reading time should be in the inconsistent condition, followed by the qualification condition, which, in turn, should be slower than the consistent condition. With the naming times, it is predicted that naming times will be faster in the inconsistent condition than in the consistent condition (I < C). This prediction is based on the fact that readers will have had exposure to the inconsistent information only in the inconsistent condition. Of primary interest, however, is the accessibility of the inconsistent information in the qualification condition. It was reasoned that if this information is automatically retrieved when the target sentence is encountered, then naming times in the qualification condition should be faster than in the consistent condition, and similar to the inconsistent condition [(I = Q) < C]. On the other hand, if the inconsistent information is not automatically activated but is consulted through a slower, more strategic process, then naming times in the qualification condition would be slower than the inconsistent condition, and similar to the consistent condition [I < (C = Q)]. A third possibility is that the naming data will reflect the original reading time data, where the inconsistent, qualification, and consistent conditions are all statistically different from one another (I < Q < C).

Undergraduates enrolled in Introductory Psychology (n = 110) participated in exchange for extra credit. Participants read a total of 24 passages, one line at a time, on a computer display, preceded by three practice passages. The eighteen experimental passages were borrowed from O'Brien et al. (1998). Each passage had three versions, each differing in the relatedness of the target sentence to the backgrounded information pertaining to the protagonist. In the consistent condition, the target sentence was consistent with the backgrounded information. In the inconsistent condition, the target sentence was inconsistent with the backgrounded information. In the qualification condition, the target sentence was inconsistent with the backgrounded information but was consistent with a briefly stated qualification of that information. The qualification version of each of O'Brien et al.'s original items was modified to ensure that participants would perceive the "updated" information as consistent with the protagonist's behavior depicted in the target sentence. This modification was pre-tested in a separate study prior to running the present experiment. As in O'Brien et al., the target sentence was identical across all three versions. Each participant saw six passages in the three experimental conditions. Passages were counterbalanced across condition such that, across all participants, each passage appeared in each condition an approximately equal number of times. Six additional items were filler passages, and basically were similar in form to the items in the consistent condition. These items were randomly interspersed among the experimental passages but were not analyzed.

Each passage was preceded by a "Ready" signal. Participants read each passage at their own pace by pressing a line advance key. Each line of the passage appeared in the center of the computer screen, and was replaced by the next line when the key was pressed. The time between key presses was stored as reading time. Immediately following the target sentence, a warning signal (XXX) appeared in the center of the screen for 500 milliseconds and was followed by a word representing the inconsistent information. Participants pronounced the target word as quickly as possible, into a microphone. This utterance activated a voice-key and resulted in the target word disappearing from the screen. Immediately following, the time interval spanning the onset of the target and the participant's first utterance was briefly displayed on the screen, and was stored as naming time. A second warning signal (???) then appeared for 500 milliseconds and was followed by a comprehension question. Participants responded by pressing a key marked YES or NO. An error message appeared if the question was answered incorrectly. An instruction to read more carefully appeared on the screen after each incorrect response following the fifth error.

The primary data were the target sentence reading times and the target word naming times in the three experimental conditions. The data from nine participants were omitted from the analyses. These participants were omitted either because they answered four or more comprehension questions (16.7%) incorrectly or because of problems prematurely tripping the voice key during the experimental session. The analyses, therefore, were based on data from 101 participants. The analysis of the reading times showed, as predicted, that participants read the target sentences more slowly in the inconsistent condition (m = 2487 msec) than in the consistent condition (m = 2221 msec) (p < .001). Also, similar to the findings of O'Brien et al. (1998), reading times in the qualification condition (m = 2282 msec) were slower than those in the consistent (albeit, not significantly so) and faster than those in the inconsistent condition (p < .001). These data are consistent with the notion that, in the qualification condition, participants consulted both the backgrounded, inconsistent information as well as the "updated" qualification information at the point they encountered text material that is relevant to previously read information.

The naming time data fell into line with the automatic retrieval hypothesis. As originally predicted, naming times were faster in the inconsistent condition (m = 449 msec) than in the consistent condition (m = 462 msec), p < .05. However, participants were equally as fast to name the target word in the qualification condition as in the inconsistent condition (m = 449 msec). These data indicate that readers have fast, direct access to backgrounded information of the protagonist, even if that information is inconsistent with an updated mental model of the protagonist. This is consistent with the notion that readers access backgrounded information via a passive, fast-acting resonance process (Myers & O'Brien, 1998).

Most writing research shows that writing is a very complex activity, mainly because writers have to manage simultaneously or successively a great number of processes. They have to select and to organize information retrieved from long term memory (planning process), to choose correct syntactic frames and appropriate lexical items (translating process), to reread and to revise the already produced text (revising process) and, finally, to program and execute the necessary motor movements for the realisation of the written text (graphic execution). The study of these different processes has led to the elaboration of models, as Hayes and Flower's (1980) well known model, or more recently, Hayes' (1996) and Kellogg's (1996) models. The research proposed here works from the perspective of these recent models. The objective was to analyze how novice writers (3rd and 5^th graders), French and Dutch, managed in real time the processes of producing a written text, and more particularly the recovery of pieces of information, and their syntactic and lexical organization in a text. The results of two experiments have therefore been compared: The first one carried out in French by Chanquoy, Foulin, and Fayol (1990), and the second led with Dutch children by Schilperoord, Chanquoy, and van den Bergh (1997).

These two experiments used, in writing, the method elaborated by Holmes (1984, 1988) for speaking. This method consisted in asking participants to produce an ending to a given text respecting a certain number of constraints. By following the main principles of this method, 3rd and 5^th graders were asked to produce text endings following three constraints: The predictability (predictable vs. unpredictable text content), the type of text (descriptive vs. narrative) and the syntactic structure (one sentence with two [or three in Dutch] clauses vs. two [or three] sentences). Thus, these two experiments aimed to study the impact of these different constraints on interclause (or inter-sentence) and pre-writing pause duration, and on writing speed. Each child, French or Dutch, wrote 8 endings, which had to respect the three constraints described above (predictability, type of text, and syntactic complexity). During all the time of their writing, children were videorecorded (the sheet of paper, the hand and the pencil) with a camera allowing very precise chronometric measures. At the end of the experiment, three measures were collected: The duration of pre-writing pause (time between the end of the instructions and the beginning of the graphic activity), the duration of interclause (or inter-sentence) pauses, that corresponded to the time between the end of the writing of the last word of a clause (or a sentence) and the beginning of the writing of the first word of the following clause (or sentence), and the writing speed (measured in number of characters written per second).

The results of the French study showed that syntactic, semantic and textual variables played an important role in the temporal management of writing, both at the level of pause duration (i.e., pre-writing and interclausal pauses) and writing speed. Thus, predictable endings led to shorter pre-writing and inter-clause pause durations than unpredictable endings. Narrative endings had shorter pauses than descriptive endings (this result was however only significant for 5th graders). Pauses were shorter and writing speeds were more rapid for endings composed of one sentence (with 2 clauses) than for those of two sentences. Finally, concerning the grade, 5^th graders made shorter pauses and wrote more rapidly than 3rd graders.

The results of the Dutch study showed that Dutch children did not manage their writing, according to the imposed constraints, in the same way as French graders. More precisely, both Dutch graders made significantly shorter pauses when they had to write unpredictable endings than when these endings were predictable; This was significant both for pre-writing and interclause pause durations. However, 3rd and 5th graders wrote unpredictable endings significantly slowly than predictable endings. For both grades, pre-writing pause durations did not significantly vary for endings built with one or three sentences. Conversely, interclausal pause durations were longer for endings with three sentences. Whatever the grade, pre-writing pauses and interclause pauses were longer when the endings were narrative. Finally, 3rd graders made shorter pauses than 5th graders. The comparison of the two experiments showed that:

Concerning the grade, while French 3rd graders made longer interclausal pauses than 5th graders, the opposite result was observed for Dutch participants: 5th graders' pre-writing and interclausal pause durations were longer than 3rd graders' durations. For both populations, the writing speed was slower for 3rd graders than for 5th graders. With respect to the constraint of predictability, and as already mentioned, in French, unexpected endings led to pre-writing and interclausal pause durations longer than predictable endings, while the opposite result was found in Dutch. Moreover, while this variable had no effect on the writing speed in French, the writing speed was more rapid for predictable endings in Dutch. For the syntactic constraint, the results were close for pre-writing and interclausal pause durations in French and in Dutch. However, while the writing speed did not significantly vary in Dutch, it was more rapid for endings built with one sentence in French, compared to the two sentence endings (but only for 3rd graders). Concerning the type of text, no effect of the writing speed was noticed both for French and Dutch children. However, while pre-writing pause durations did not significantly differ in French endings, for the Dutch children these pauses were longer in narrative than in descriptive endings. In addition, an opposition between the two populations was observed for interclausal pause durations: these durations were longer for French descriptive endings (in 5th graders) and shorter for Dutch descriptive endings, compared to narrative endings. Thus, while strong semantic constraints led to an increase of pause durations and to a decrease of writing speed in French children's texts, this phenomenon was not observed for the Dutch children. These latter results were very complex to analyse and therefore we investigated them further, notably by analysing verbal protocols collected only from Dutch children. For example, the finer analysis of Dutch children's ending showed that these children used, so as to palliate difficulties of the task, strategies that had not been observed in French participants. Concerning the predictability constraint, the element of surprise in unpredictable endings was simply brought about by a negation or by an inversion of one or several sentences provided in the beginnings. This phenomenon could then explain why Dutch children wrote unpredictable endings more rapidly.

Investigate listening comprehension is a way to specify the nature of the deficits in reading comprehension. Two hypotheses can be considered: The linguistic hypothesis views poor readers as suffering problems in language processing whereas the cognitive hypothesis involves more basic deficits (attention, working memory capacity, processing speed...). Previous investigations have shown that poor readers are deficient in processing cohesion devices such as pronouns in written texts. Their difficulties are stronger in situations involving a high cognitive load (Ehrlich & Rémond, 1997). However, the specificity of these difficulties to the written modality and the factors responsible for them remain an open question. The purpose of the two experiments reported in this communication was to investigate the on-line processing of cohesion in a listening situation, in two groups of children contrasted according to their reading comprehension ability. The experiments studied how children integrate different kinds of information for linking sentences by means of a cross-modal naming task used by Marslen-Wilson, Tyler and Koster (1993) in adults. Subjects heard short texts followed by a sentence fragment which had an anaphor in subject position followed by a verb phrase. At the end of the fragment, a visual probe was presented which subjects had to name as quickly as possible. The visual probe was appropriate or inappropriate. The appropriateness effect on the naming latencies of the visual probe reveals the processing of the anaphor.

Experiment 1. On-line processing of two types of subjects anaphors: Repeated noun phrase and pronoun

Subjects. 106 7-year-old children (first grade) were submitted to a reading comprehension test and to a listening comprehension test (French adaptation of Neale, 1988). From the distribution of reading comprehension scores, two groups were formed: Good readers (N = 31) and poor readers (N = 35).

Material and Procedure. 16 short texts were constructed containing a context sentence and a critical sentence. Children heard the text. The last word of the critical sentence was presented visually. Children had to name this visual probe. In addition, 16 filler texts were presented with a different structure. Example of an experimental text: “After a long trip around the world, Elodie had diner with Sebastien in a restaurant (Context sentence). She glady spoke with (Critical sentence) / (Visual probe).

In all texts, the two protagonists (A and B, here Elodie and Sebastien) in the context sentence differed in gender, the verb of the critical sentence was neutral with respect the protagonists. The visual probe was always an object pronoun (lui ou elle). In the critical sentence, the variables were the following: (1) Type of subject anaphor: Repeated noun phrase (Elodie) or pronoun (she); (2) Subject anaphor: Protagonist A or protagonist B; (3) Visual probe: The probe A (him) is appropriate with the protagonist A in subject position (Elodie), so the probe A is inappropriate with the protagonist B as subject. The probe B (she) is appropriate with the protagonist B in subject position (Sebastien). So the poste B is inappropriate with the protagonist A as subject. The difference between the naming latencies between probe A and probe B for each subject anaphor (appropriateness effect) reveals the processing of this subject anaphor.

Results. If poor readers suffer a specific deficit in anaphor processing (linguistic hypothesis), differences in the appropriateness effect between the two groups are expected. Naming latencies were significantly longer for poor readers. They also were significantly longer for repeated noun phrases than for pronouns. The reading comprehension ability did not interact significantly with any other factors. However planned comparisons showed similarities and differences in good and poor readers. For repeated noun phrases, good and poor readers showed similarly an appropriateness effect only when they refered to the protagonist A. For pronouns, good readers showed an appropriateness effect when pronoun refered to protagonist A, whereas poor readers tended to show this effect (non significantly) when pronoun refered to protagonist B. In conclusion, poor readers seems to have specific difficulties in processing pronouns, but results are not conclusive.

Experiment 2. On-line integration of lexical properties of the pronoun and semantic verb bias

Subjects. Two groups of good readers (N = 29) and poor readers (N = 33) were formed from a pool of 105 8-year-old children (second grade).

Material and Procedure. The cross-modal naming task was used as in Experiment 1. The experimental texts had the same structure as in Experiment 1, but the subject anaphor was always a pronoun. A new variable was manipulated : Three sets of texts varied as a function of the semantic feature of the verb of the critical sentence. Set 1: The verb was neutral as in Experiment 1. Set 2: The verb biased towards the protagonist A. Given that the pronoun refered to protagonist A or protagonist B, the lexical cue on the pronoun and the verb converged towards probe A in half texts. In other texts, there was a conflict between pronoun and verb bias. Set 3: The verb biased towards the protagonist B. In a symetric way to set 2, pronoun and verb bias converged towards probe B or were in conflict.

Results. Naming latencies were similar for good and poor readers. A significant interaction was observed between reading comprehension ability and the factors underlying the appropriateness effect. In sentences with neutral verb, good readers were able to process the lexical cue of the pronoun as adults in Marslen-Wilson and al's study. Poor readers processed the lexical cue when pronouns refers to protagonist B, but not when it refers to protagonist A. In sentences with verb bias, the appropriateness effect was different for good and poor readers. When the lexical cue of the pronoun and verb bias were convergent, the appropriateness effect was smaller in poor readers than in good readers. When the two factors were in conflict, the weight of the lexical cue of the pronoun was dominant relatively of verb bias, especially in good readers.

To summarize, results supported the linguistic hypothesis : 1- Poor readers have difficulties to process the lexical cue of the pronoun in some conditions. 2- Poor readers do not integrate efficiently lexical cue of the pronoun and semantic inference from the verb.

Difficulties observed in the processing of cohesion in written texts are not specific to this modality. Poor readers are also deficient in understanding spoken utterances and results support the linguistic hypothesis.

Ehrlich, M.-F. & Rémond, M. (1997). Skilled and less skilled comprehenders: French children's processing of anaphoric devices in written texts. British Journal of Developmental Psychology, 15, 291-309.

Marslen-Wilson, W.D., Tyler, L.K. & Koster, C. (1993). Integrative Processes in Utterance Resolution. Journal of Memory and Language, 32, 647-666.

Two categories of studies are examined: Some mention intermediate representations between the instructions and the schema, others are more centered on the activity of the participant rather than on the formal properties of the problems.

Models which include a situation model: The authors (Reusser, 1989; Nathan, Kintsch & Young, 1992; Stern & Lehrndorfer, 1992; Staub & Reusser, 1995) bring into play what they call the situation model, a more qualitative and less mathematical representation of the text of the problem, in order to construct more efficient representations for solving it. The models constructed by these authors show that for them, taking into account the quantitative elements of the instructions (numerical values and quantitative relations) is not the only thing brought into play in order to solve the problem. They also attribute importance to “ the intermediate representations between the body of the text and the mathematisation of the problem ” (Kintsch, 1998, p. 347); the activation of the schema is not called into question, but it passes through different paths (Kintsch, 1998).

Contextualist or procedural models: In the second category are to be found models known as contextualist (Bastien, 1997), constructed in a school context (Greer, 1997; Reusser and Stebler, 1997), and which for the purposes of this study will be called procedural since they are centered more on the activity of the participant than on the formal properties of the problem. In these models reading activates more or less efficient chain of rules of actions: For example, Bouchafa (1985) suggested a model for the processing of problems composed of three successive modules: a) extraction of the numerical data in the order in which they appear in the text; b) partition of this set of numbers on the basis of various criteria like “ belong to the same set ”, or “ same size ”...; c) selection of arithmetical operators to place between the numbers on the basis of “ semantic triggers ” or the activation of elementary schemas. We accept Bouchafa’s model with the b) partition-criterion: the criterion of “ linear partition ”; groupings of data are made by order of appearance, the first with the second, and the combination obtained with the third etc. This criterion is natural, insofar as reading is a linear activity with privileged places for the integration of information (Passerault & Chesnet, 1991). In the case of instruction problems, the information is also treated as the reading progresses, including numerical information (Fayol, Abdi & Gombert, 1987; Reusser & Stebler, 1997).

The objective of the current study was to determine whether the processing brought into play during the reading-solving of problems by 11 years old children belongs to the first category of models or to the second. With this in mind, the pupils thus had to solve problems of the distributive type; for example: For the school party, the school cook puts on the 13 tables [a basket containing] 6 bananas and 8 apples on each. What is the total number of items of fruit that she will use? They could be solved by two different strategies: Development (13X6 + 13X8) or factorization 13 (8 + 6). The formulation of instructions varied between two crossed variables: presence/absence of a structuring term such as “ basket ”, and the order of numerical data. The dependent variable was the choice of the solution strategy. The solution models which give importance to the situation model (first category) enable the effect of the presence of a structuring term, and the absence of order effect; whereas the linear procedural model predicts an influence of item order and no effect of the structuring object.

Participants. 182 pupils from 10 different grade 5 classes in 7 schools from the Poitiers region in France. Average age: 11, s.d. = 0.55.

Material. Three distributive problems, each presented in four versions, themselves the result of crossing two variables: Presence/absence of a structuring element (S+/S-) and the placing of the scalar (O2). In the example above, the structuring element is between square brackets and the scalar (13) is at the beginning of the word problem.

Procedure. The distributive problems were combined with four additional distracting problems, and the experiment was carried out in two sessions, with an interval of about 10 days; each pupil solved three distributive problems under the same experimental conditions, one during the first session and two during the second. The order of the problems was counterbalanced.

Performances. 74% to 81% of the problems were solved correctly (92% to 93% if calculation errors were not taken into account).

Strategies. Each factorisation (correct factorisations, e.g. (6+8) x 13; or incomplete factorisations: factorisations initiated (6+8), but not completed or erroneously completed) was coded 1 and the rest zero and then a total was made across all the problems; an indicator of factorisation was thus obtained, varying from 0 to 3 for each participant.

Factorisation was stronger when there was a structuring element in the text (M = 2.13) than when there was not (M = 1.66), (F[1, 174] = 4.31; p < 0.039), but the place of the scalar had no effect (m=1.70 and 2.08; F[1, 174] = 3.01; p > 0.08), and there was no interaction (F[1, 174] <1). According to our hypotheses, these results are compatible with a processing model which takes into account the situation model rather than with the linear procedural model. Despite the absence of interaction, the superiority of the factorisations in the presence of a structuring element disappeared when the scalar was placed at the end of the problem (F[1, 174] <1). However, when there was no structuring element, the effect of the placing of the scalar approached significance (F[1, 174] = 3.66; p < 0.057).

According to our hypotheses, the results of the canonical analysis, with as the only significant effect the presence of a structuring element, support the idea of a processing model which brings in the situation model. However, the effect of the structuring element disappears when the scalar is placed at the end; for example when the data order was 6 bananas, 8 apples, and 13 tables, the grouping ((6,8), 13) was observed, which can be explained either by the structuring element when it is present, or by linear processing. On the contrary, when the scalar was at the beginning (13 tables, 6 bananas, 8 apples), the grouping (13, (6, 8)) observed predominately in S+ can only be explained by the existence of a situation model which integrates the presence of the structuring element; otherwise, the linear processing would give groupings of the type ((13, 6)...), and this is what was observed in S-. It therefore seems that processing which uses situation models dominates, but when it cannot operate, due to the presence of the structuring term in the instructions, the linear procedural model takes over the role.

The experiment showed the effect of a term which structured the elements of the text on the choice of the solving strategy; this term only came into play by means of the transitory situation model constructed during reading. This result corresponds with Reusser’s (1989) Situation Problem Solver (SPS) and, generally, the models which integrate more qualitative representations than the strict formal schema. However, in the absence of a structuring term, the representations are less organised, and it seems that linear processing takes over.

Metacognitive knowledge in reading may be defined as the knowledge a person possesses about a) the structure and function of written discourse (Gombert, 1990), and b) three broad categories of reading skills: Evaluation, planning, regulation (Paris & Jacobs, 1984). Evaluation refers to the assessment of the task difficulties and one's cognitive abilities. Planning involves the selection of specific strategies to reach specific goals. Regulation concerns monitoring and remediating difficulties during the course of reading. Several authors have pointed out the role of metacognitive knowledge in text comprehension (Garner, 1987; Brown, Armbruster & Baker, 1986; Hacker, 1997). For example, Garner (1987) suggested that in order to read and study expository texts effectively, students must possess both conceptual knowledge (about the domain or situation dealt with in the text), text-processing strategies (e.g., ongoing summarization and backtracking) and "schema for exposition". Garner defined the latter as a person's knowledge about textual and metatextual signals of the importance or salience of information in a particular genre or type of text. Brown, Armbruster and Baker (1986) reported some differences in metacognition between readers at different ages or reading levels. And Hacker's results showed that metacognitive deficiencies could persist in college students. A unique French study (Ehrlich, Kurtz-Costes, Rémond & Loridant, 1995) showed that the level of metacognitive knowledge could explain a specific proportion of variance in comprehension (multiple regression analysis), even with decoding and motivation variables entered in the regression. However the relationship of metacognitive knowledge and the efficiency of comprehension processes is still unclear. Firstly, studies have focused on children, so that we know very little about knowledge in adults. Second, few studies have attempted to relate qualitatively subjects' comprehension level and their actual knowledge about comprehension: For example, which category(ies) of knowledge are the poor comprehender missing?

The purpose of the present studies was (a) to design a tool in order to assess explicit metacognitive knowledge in both young adults and children and (b) to analyze the development of metaknowledge in children and qualitative individual differences between good and less good comprehenders and (c) to describe knowledge about reading in University students.

A questionnaire (MK23) was designed to collect qualitative data about children and adults' metacognitive knowledge. Four dimensions of metaknowledge were considered:

1) Evaluation, e.g.: "When you're reading, what is difficult for you?"; "Do you always understand what you read?" (5 items);

2) Functions of text features and documents, e.g.: "What is a table of content, what is it for?"; "Do you think that initial sentences in a paragraph or a text are special? Do they have a special importance?" (6 items);

3) Planning, e.g.: "What is the best way to remember what you have read?"; "If you have only 3 minutes to read a 2-page text about which you will have to answer questions later, how do you proceed?" (7 items);

4) Regulation, e.g.: "What do you do when you read a sentence that you don't understand?"; "Do you happen to look back and read again some sentences in a text? If so, why?" (5 items).

Questions aimed at assessing knowledge about these four dimensions were constructed from scratch or based on previous works (Ehrlich et al., 1995; Myers & Paris, 1978).

84 children from grades 3 and 5 (9 and 11 year old) and 66 3rd year psychology students at the University of Poitiers, responded to the MK23 questionaire. Children were interviewed individually and the interviewer transcribed their answers on a scoring sheet. Adults wrote down their answers. Furthermore, children read a 424-word written text for comprehension and answered 24 literal and inferential questions, which provided a comprehension score. Each child was asked to provide an estimate of his/her comprehension score.

Data analysis consisted in scoring all the protocols and identified between 3 and 7 answer categories per item; then the categories were grouped into "simple" answers, "elaborate" answers, and "none/other" answers. The main results are summarized below.

A minority of children demonstrated some explicit knowledge about text features and comprehension activities. Generally, the children didn't provide an accurate evaluation of their comprehension level; they explained their reading difficulties by referring to elementary processes, such as word identification, and not to deeper processes, such as knowledge base or global coherence; they could define text features (paragraphs) and documents (table of contents) but not their function in comprehension; they didn't know many strategies to plan or regulate their activity and they didn't report strategy variation as a function of reading goals. Simple and elaborate answers moderately increased between the ages of 9 and 11 (a significant difference was found only on the "text features" scale). Nevertheless good and poor comprehenders differed in the quality and amount of knowledge, particularly on "planning" and "regulation" subscales.

As expected, elaborate answers increased in adults; in particular, evaluation knowledge seemed to develop during secondary and higher education, as evidenced in University students' rather high scores to these items. In contrast, less than half the undergraduate students provided elaborate answers to questions about planning and regulation activities: Most of them attributed their failure to local features or elementary processes, and overestimated their comprehension level. At last, many adults didn't think about the most useful strategies for solving difficulties.

In sum, children's knowledge about reading appeared to be related to their performance in text comprehension. In adults, large individual differences were observed too, which could be related to their academic success or failure. While a great deal of attention has been given to the development of metacognition in children, very little is known on the relationship between meta-knowledge and cognitive performance on comprehension tasks in adults. Yet this issue is becoming critical as many higher education institutions promote learning situations which rely on students' ability to interact with complex information systems. Further studies will assess the possible relationship between students' metaknowledge and their academic success.

Brown, A.L., Armbruster, B.B. & Baker, L. (1986). The role of metacognition in reading and studying. In J. Orasanu (Ed.) Reading Comprehension: From Research to Practice (pp. 49-75). Hillsdale, N.J: Erlbaum.

Ehrlich, M.F., Kurtz-Costes, B., Rémond, M., & Loridant, C. (1995). Les différences individuelles dans la compréhension de l'écrit: Facteurs cognitivo-linguistiques et motivationnels. Cahiers d'acquisition et de la pathologie du langage, 13(1), 37-58.

Garner, R. (1987). Strategies for reading and studying expository text. Educational Psychologist, 22, 299-312.

Gombert, J.E. (1990). Le développement métalinguistique. Paris: Presses Universitaires de France.

Hacker, D.J. (1997). Comprehension monitoring of written discourse across early-to-middle adolescence. Reading and Writing: An Interdisciplinary Journal, 9, 207-204.

Myers, M., & Paris, S.G. (1978). Children's metacognitive knowledge about reading. Journal of Educational Psychology, 70(5), 680-690.

Paris, S.G. & Jacobs, J.E. (1984). The benefits of informed instruction for children's reading awareness and comprehension skills. Child Development, 55, 2083-2093.

Several studies have been conducted to test for both assembled and adressed phonologic mediation in reading comprehension (Coltheart, Curtis, Atkins & Haller, 1993). It is clear now that adult skilled reading of frequent words implies two kinds of processes. The adressing refers to a recognition of extracting some visual properties of the stimulus, that will later be used to search the mental lexicon. The assembling refers to the analysis of words' components. The developmental question is how this balance between these two types of processing is achieved. Two alternatives have been proposed. Either the beginning readers develop a sequential reading (letter-by-letter or grapheme-by-grapheme strategy) and later they use bigger units or the beginning readers start reading by focusing on global characteristics of words and later they would focus on detailed information. These two alternatives coincide with the distinction between developmental models. Thus, the stage developmental models (Frith, 1986; Harris & Coltheart, 1986; Morton, 1989) claims that children in the phonological stage first use small units like letters and only later employ larger unit such as morpheme or analogies based on word rimes. The second alternative proposes exactly the reverse. Goswami and Bryant (1990) proposed that the use of analogies based on word rimes appears first and phonological procedure based on grapheme-phoneme correspondence only later. Thus, studies in English language suggest that onset-rime awareness might be connected to children making analogies between the spelling patterns representing those linguistic units in words (Goswami, 1994; Goswami & Mead, 1992; Muter, Snowling & Taylor, 1994; Peterson & Haines, 1992). The first kind of models considers that reading by analogy is characteristic of the last stage of reading acquisition. The second type of models suppose that children start the reading process through an analogical process. More precisely, Goswami and Bryant (1990) think that the children's capacity of analyzing the speech stream into syllables and more specifically into onset and rimes would allow them to discover the existence of analogous and, therefore, to expand their reading skill to new words. Thus, the early stage of reading acquisition distinguish between a small unit approach and a large unit approach. They differ mainly in theirs claims about the size of the linguistic units which are important in the initial stage of reading acquisition. The purpose of this paper is to address the predictions of the " large unit theory " in French. This is approached by using an identification lexical task with pre-literate children and a decision task in the second year of learning.

44 pre-literate children of kindergarden (mean age 5;7) were tested to recognize words in a forced-choice task. Two groups (good phonological awareness and poor phonological awareness) were proposed to recognize target words (CVC) with a clue word (analogy condition) or not. Target words and clue words share some orthographic units, either rime (lac-sac), either body (bus-but) either onset+coda (mur-mer). It is shown that only good phonological awareness group significantly recognize in the analogy condition target words which share the rime with clue words, as in Goswami's studies.

In this lexical-decision task children indicated whether presented letter strings constituted words or not. In this experiment we manipulated type of misspellings (control and analogous pseudo-words), base-word frequency and the reading level. The main question was whether second graders show evidence of analogy. This was investigated by means of a lexical-decision task in which subjects have to decide whether or not individually presented letter strings constitute words. This experiment compared the children's speed and accuracy in lexical decision task of bisyllabic nonsense words. The children were from second grade. A two-groups reading level design was used in this study (Khomsi, 1997). A frequency table (Brulex) was consulted for the selection of twelve bisyllabic words (6 high-and 6 low frequency). The twelve base words were changed into 24 analogous pseudo-words. The stimuli consisted of lists of bisyllabic nonsense words that either shared orthographic units with real words in beginning (garale/garage is a analogous pseudo-word on the beginning APWb) or shared orthographic unit with real words at the end, that is analogous pseudo-word on the end (marage/garage is APWe) and control pseudo-words (pudole). Finally, 36 words were used to serve as fillers. Proportions of correctly identified errors- that is, correct non-word decisions- were calculated for two kinds of analogous pseudo-words and control pseudo-words derived from frequent and infrequent base words.

A 2 (reading level: Good versus poor readers) by 3 (type of misspelling: APWb, APWe, control) and by 2 (base-word frequency: Frequent versus unfrequent) ANOVA on correctly identified errors was performed on subject and item means. Subject's mean decision latencies of correct no-responses to analogous pseudo-words and correct no-responses to control misspellings were also computed. The results of the lexical decision task show that the analogous pseudo-word APWb (garale) were processed faster than the analogous pseudo-words who have kept the final orthographic unit (marage). No frequency effect was found. In good readers, positive evidence was obtained for analogical process because fewer analogous pseudo-words APWb were correctly classified as non-words than control pseudo-words and analogous pseudo-words APWe.

Many aspects of the results of these experiments are apparently consistent with the theoretical viewpoint set out by Goswami (1986, 1993). The results show that the reading by large-unit analogies determines the initial stage of acquisition but a late influence occurs with second graders. Therefore, we assume that analogical procedures are used for both pre-literate children and second graders. In other words, the ability to draw analogies might be thought of as another type of self-teaching mechanism for the beginning readers. In this perspective, the analogical processes correspond to the use, by learner, of his or her own basis of knowledge (phonological and visuo-orthographic) in order to increase his or her lexical knowledge. It is thus concluded that analogical processes may hightlight some developmental stages and processing strategies in the reading acquisition.

Funding for this research was made available from the Spencer Foundation (Chicago, Illinois, USA), "Developing literacy in different contexts and different languages", Ruth Berman, Principal Investigator.

Studies of the acquisition of monologues have concentrated essentially on narrative discourse. Few have attempted comparative analyses between different genres. In this paper we will study the acquisition of discourse in two textual types (narrative and expositive), in two modes of production (spoken/written) and two different orders of production (spoken-written/written-spoken) in French children and adults. Our research compares the different linguistic forms used to structure the discourse as a function of these variables, paying particular attention to genre.

The general aim of this study is to access the impact of literacy on the child’s developing syntactic competence. Blanche-Benveniste uses the term "first grammar " to refer to the grammar used before schooling and "second grammar " to refer to the grammar encountered in school. Romaine (1984) underscores, the impact of learning to read and write "to successful participation and progress in school" (Romaine, 1984, p. 196). Whereas all speakers acquire the first grammar, some of them will have more difficulties to control the second. " Although all children learn to speak and hear without formal instruction, this is not true for reading and writing (Romaine, 1984, p. 207). For Berman (1996), literacy refers to the capacity " to use decontextualised written texts as a source for deriving information and as a means for conveying information to others (…) Cognitively, literacy involves the ability to use writing not merely for communication, but as a epistemic tool, as a means for analyzing and hence transforming one’s own knowledge-base ".

In this present communication we concentrate on synctatic packaging in both narrative and expositive texts. Speakers/writers must organize their text at both the level of coherence or meaning, as well as at the level of cohesion, or linguistic structure. Our study examines clause combining from both a syntactic and a semantic perspective. We use the notion of t-units as defined by Biber (1998), in accordance with Hunt (1965), as an independent clause with all its dependent clauses. The level of t-units provides e measure the syntactic complexity. From a developmental perspective we examine both the size of t-units and their internal structure across spoken and written texts and narrative and expository texts. In addition we examine the semantic relations between the clauses, which make up the t-unit.

The work presented is part of larger international project undertaken in collaboration with Israel, Netherlands, Spain, Sweden and United-States. Like the other countries, the French team has included four age groups: three for children and one for adults. The age of the child groups are 9/10-, 12/13- and 15/16-years-old. The mothers of all the children have a baccalauréat. The adults have a level of education of at least 5 years of study after the baccalauréat. Each age group consists of forty subjects. Subjects produce one narrative and one expository text, each text in a spoken and a written mode. Thus there are four texts for each subject, the data collection is organized in two separated parts: session one and session two. Each subject takes part in the two sessions. In addition, the order variable is controlled resulting in four orders, shown in table 1.

Hypotheses: (H1) The number of clauses per t-units increases with age; (H2) The inventory of clause type changes as a function of age, text type and modality of production; (H3) The variable text genre will prove to play a stronger role than modality of production.

The results confirms the first two hypotheses: The number of clauses per t-unit increases with age. The inventory of subordination types increases with age. Text genre is as important if not more important than modality as a variable. However, in terms of semantic relationships established between successive clauses, the developmental progress is much more difficult to interpret.

There is now substantial evidence from studies of reading acquisition to suggest that phonological skills are related to the development of young children's reading ability in alaphabetic orthographies (Bradley & Bryant, 1985; Stuart & Coltheart, 1988; Wagner, 1988; Wagner, Torgeson & Rashotte, 1994). In French, the syllable is a phonologically and orthographically well-defined unit. It proved of particular importance in speech perception of words (Melhler, Dommergues, Frauenfelder, & Segui, 1981) and in speech production (Ferrand, Segui, & Grainger, 1996). So the role of syllables as phonological units in hearing and in deaf children has still not raised a lot of research in French. Children with severe/profound prelingually deafness encounter considerable difficulty in achieving fluent reading skills. Generally, it has been hypothesised that the relatively poor reading skills of mostly deaf individuals results from deficiencies in phonological processing. Indeed, the lack of accurate and specified phonological representation results from the lack of adequately developed spoken language. For example, Conrad (1972) found that hearing subjects confused words which sounded similar, whereas deaf subjects confused similar spelled words. More recently, Merrils, Underwood and Wood (1994) suggested that deaf readers relied more on visual word features than hearing readers did. In the same way, Waters and Doehring (1990), Harris and Beech (1997) did not provided evidence for deaf subjects sensitivity to the regularity of pseudohomophones. Moreover, Campbell and Wright (1990) used a paired-association task to show that unlike hearing children, deaf children did not use picture-name rhyme spontaneously as a cue to recall pairs. In contrast to these results, some authors found that deaf children are able to develop and use a phonological code in various tasks. For example, deaf subjects were able to read pseudo-words just like hearing subjects (Leybaert & Alegria, 1993). Evidence showed that pseudo-words with a complexe phonology were more difficult to read than those with a simple phonology. Deaf subjects were also more accurate in reading regular than irregular words. Transler, Leybaert and Gombert have shown that deaf children like hearing children use syllables as phonological units in print processing situations. Children were asked to copy written words and pseudo-words and measured the number of glances at the item, copying duration and locus of the first segmentation. Like hearing children, deaf children used syllables as copyong units. They only differed in a condition in wich orthographic and phonological criteria were differentiated. We also collected data providing evidence that profoundly deaf children (8;6 years) like hearing (see Colé, Magnan, & Grainger, in press; Magnan & Colé, 1999) were sensitive to the syllable compatibility between target and words, but only in case of short and familiar words (Paire-Ficout & Magnan, accepted).

The aim of this present study is to compare deaf children who have been exposed to Cued Speech (CS) and deaf children who have not been exposed to CS. In CS, speech/reading is combined with manual cues which disambiguate it. The central hypothesis is that young deaf children with greater exposure of Cued Speech would be more sensitive to syllabic congruency. In order to test this hypothesis, we used a visual version of the monitoring technique employed in studies of speech perception (Mehler et al., 1981). Subjects were asked to detect a letter string target presented visually at the initial position of words. The target structure was either a consonant-vowel (CV) or consonant-vowel- consonant (CVC) structure and either did or did not correspond to the structure of the first syllabic part of the target bearing words. Thus, for example, children were asked to detect a letter string BA (CV target) in BALLON (CV word) or in BALCON (CVC word) and the letter string BAL (CVC target) in BALLON and BALCON. This task can be performed phonologically or visually and if the beginning readers process written language on visual/orthographic grounds only, the detection times should only be function of the size of the target sequence so that detection times should be faster for CV targets (2 letters long) than for CVC targets (3 letters long) whatever the syllabic congruency is. In contrast, if the task is performed phonologically, the detection times should be function of the syllabic congruency between the structure of the target sequence and the one of the first part of the test word. As an example, detection times should be faster when detecting SO in SOLEIL (sun) than SOL in SOLEIL and detecting SOL in SOLDAT (soldier) should be faster than detecting SO in SOLDAT. In the latter case, an interaction between the size of the target sequence and the syllabic structure of the test word is expected. The first data seem indicate that Cued Speech allows deaf children to be more sensitive to the syllable congruency and to detect the congruency more rapidly.

Numerous findings show that phonological processing skills are fundamental in the acquisition of literacy (Goswami & Bryant, 1990; Liberman & Shankweiler, 1985). According to Wagner and Torgesen (1987) phonological processing is made of three sets of distinctive component skills: Phonological awareness, phonetic recoding in working memory and phonological recoding in lexical acces. Phonological recoding involves the extraction of lexical characteristics and pronunciation codes from written information. Thus, in this phase, the beginning reader uses the relations between graphic units and phonological units in reading. Most current theoretical models of reading acquisition suggests that reading acquisition rapidly allows syllable-sized units to be accessed from print (Sprenger-Charolles & Siegel, 1997), and that this type of coding continues to influence how adult readers process low-frequency words (Colé, Magnan & Grainger, 2000; Ferrand, Segui & Grainger, 1996). The syllabic effects obtained with good readers after the first year of schooling (Colé & Magnan, 1997; Magnan & Colé, 1999) may be specific to French and more precisely to the kind of orthographic system. French is a language where the syllable is likely to be a critical unit in the processing of speech. We hypothesize that French children at the start of reading acquisition use their syllabically structured spoken-words system. French, although not perfectly regular in the transcription from phonemes to graphemes is more regular than the other way around. Thus, it is a much more transparent language than English and has a much more predictable relation between graphemes and phonemes. The main purpose of this work is to show that this syllable-based procedure is influenced both by the phonological and orthographic characteristics of French. More precisely, during reading acquisition this syllabic process might be affected when the orthographic transcription of syllables is difficult i.e contains a consonant cluster in the onset (CCV). So we used words with first syllables that possess a consonant cluster in the onset (CCV) or a rime with a coda (CVC). In French, performance on closed syllables with a coda (CVC) can be expected to be better than performance on open syllables (CCV). More precisely, inversions of letter sequences or omissions of one letter, often pointed by teachers, (“cravate” (tie) is read “ carvate ” or “ cavate ”) are considered like visual errors. We think that these sequential errors constitute a partial use of grapheme-phoneme correspondences (Bastien-Toniazzo, Magnan, Bouchafa, 1996, 1998; Sprenger-Charolles & Siegel, 1996). The basic structure of the syllable is a simple onset and a rime without coda (Clement & Keyser, 1983). This single consonant and vowel (CV) exists in all languages and leads to superior performance in the early acquisition of reading.

Subjects. A group of 20 children (mean age in december was 6 years 7 months) was tested in the months of December and June during the first grade which is the first year of reading instruction in France. Using a reading test (Khomsi, 1990), we distinghish poor and normal readers. The subjects were selected according to the following criteria. All were average with regard to cognitive development. All were native French speakers from the middle social classes, who were attending their grade at the regular age. They also had normal (or corrected-to-normal) vision and normal hearing. The children came from 2 classrooms from a school in suburbs of Lyon. The teachers used the same “ mixed ” method. This method combined a “ global ” approach (using key words and short texts) and an “ analytical ” approach (focusing on simple vowels and consonants in syllables and in words).

Material and Procedure. Children were asked to detect a letter string target presented visually at the initial position of word presented below. The target was either a consonant-consonant-vowel (CCV) or a consonant-vowel-consonant (CVC) structure and either did or did not correspond to the first Syllabic part of the target bearing word. For example, children had to detect CAR (CVC target) in CARNET (CVC word) or in CRAPAUD (CCV word) and CRA (CCV target) in the same words. The same word was associated with two different experimental conditions. Four experimental situations were proposed to children. The task was performed with familiar words. Familiarity was assessed by a self made test. The word length was controlled (6/7 letters). The same words were used for two sessions. The children were tested individually in a 15-20 minutes session. Sequences were presented in the center of an computer monitor as follows. A fixation point was presented during 500 milliseconds. After a 500 milliseconds blank interval, the target was presented on the screen during 500 milliseconds then the test-word appeared directly. The test word remained on the screen until the child's response. He or she had to decide as quickly and as accurately as possible whether or not the target occurred At the initial position of the test-word. The child had to press one button If the target and test word corresponded and another one if they did not. The "yes" response was given with the preferred hand. The experimental session began with 5 practice trials. The sequences order was randomized. We think that the first good readers after one year of schooling will be sensitive to the congruence of the target and the word. However, no such matching is expected with consonant clusters at the beginning of learning. Our hypothesis is that the processing varies as function of the complexity of syllables. More precisely, we think that CVC words would produce significantly fewer errors than CCV words. The consonant clusters could involve a delay in the processing. The congruence is expected for CCV words and CVC words in the session 2 only for the good readers.

The results suggest that good readers activate orthographic and phonological informations more efficiently than poor readers. Our results agree with the recent ones of Booth, Perfetti and Mac Whinney (1999), showing that becoming a skilled reader involves the development of fast and automatic orthographic and phonological processes. We think that poor readers have less efficient orthographic and phonological processes that may result from the imprecise and sparse nature of their representations (Perfetti, 1992).

Language teachers are faced with a difficult dilemma. They know that students benefit from practice, and from getting feedback on their compositions. But the more in-depth the feedback is, the more time it takes to produce the feedback. With finite amounts of grading time, language teachers are thus faced with a tradeoff: Either assign less writing or provide quicker, less-detailed feedback. This paper describes a computer tool which uses a variety of language processing technologies to address this dilemma in a novel way. The tool is called Select-a-Kibitzer (hereafter SaK). SaK is based heavily on theories of language composition that describe the process of writing as one of attempting to simultaneously solve multiple, potentially conflicting constraints (Flower, 1994). These constraints come from a wide range of sources. Some constraints depend on the student's knowledge (or lack of knowledge) of their language. Spelling, grammar, and word sense selection are difficult for many students. Other constraints are conceptual. Students have difficulty reasoning about what they know about an assigned topic and what they can relate in a composition. Other constraints are socio-linguistic in nature. There are commonly differences between the teacher's language, and the language the students naturally speak. The existence of so many different constraints makes the composition process very difficult. What makes it more difficult is that students normally do not even know explicitly what the different constraints are. Thus, they are unable to reason about how to reconcile them. SaK attempts to make the student's task easier by making the constraints explicit. It does this by using different computer agents to represent different constraints.

The knowledge which these agents use is called from a variety of sources. Lexical knowledge is extracted from the WordNet lexicographer's database (Fellbaum, 1998). This knowledge source helps determine not just the spelling and sense of a word, but also related words which are more or less specific (hyponyms and hypernyms). Knowledge about the syntactic structure of the student's text is extracted by parsing the text with a readily-available natural language parser (Allen, 1995). SaK process WordNet definitions into a form in which they can be used as an online lexicon for the Allen parser. The parser provides a parse tree for each of the student's sentences.

The sources mentioned above provide important information, but it is also information that is fairly easy for teachers to give feedback on. The most important source of information for SaK is that about the meaning of the text. Latent Semantic Analysis (LSA: Foltz, Kintsch, & Landauer, 1998; Landauer & Dumais, 1997) is a statistical technique which gives impressive results in computing the semantic similarity between texts. By themselves, semantic similarity judgments between texts can be a valuable tool to evaluate text meanings. Given a set of "target essays" for a particular assignment that cover a range of relevant topics, one can determine with LSA which of these essays the student's composition is most similar to, and thereby determine the overall topic of the essay. This technique can also be used to suggest additional material that the student might want to address, and determine if the student's composition is completely irrelevant to the assignment.

The potential uses of LSA go far simple semantic judgments, however. By comparing the LSA scores of consecutive sentences in a composition, you get a measure for how coherent the overall composition is. You also can tell where breaks between different subtopics are. LSA can also be used to provide an automatic summary of a text, by finding the breaks between subtopics, and then taking the most central sentence from each (Wiemer-Hastings & Graesser, in press).

The individual agents of SaK are implemented using the Soar problem-solving architecture (Laird & Rosenbloom, 1996), which allows us to express the agents knowledge in the form of condition-action rules. SaK places the information from each of the linguistic information sources into Soar's working memory. SaK's agent rules use this information along with contextual knowledge about the student and the particular assignment to decide what type of feedback should be given to the student about his or her composition. SaK then displays that feedback to the student. The key is that, unlike a standard computer style checker, the feedback does not come from a single, "all-knowing" evaluator, but is associated with different characters. Each different agent has an iconic picture (for example, a cowboy, dog, or William Shakespeare). Other studies show that information presented by more-or-less lifelike agents is more readily understood and remembered by students (for example, Lester et al, 1997). Each agent comments on a single aspect of the composition. Some of the comments are negative, pointing out errors or misconceptions by the student. Others are positive, praising some aspect of the student's composition. By recognizing the correspondence between the agents and the types of feedback, the student can recognize that there are different aspects of the writing process which must be taken into account. This knowledge will allow them to think explicitly about what they can do to improve their compositions.

The current status of SaK is that it is has all the pieces in place to collect the linguistic knowledge, to fire the rules which invoke different feedback items, and to present them graphically to the student. What SaK currently lacks is the principle pedagogical knowledge concerning what types of feedback are most appropriate at what stages of the composition process. The authors welcome ideas concerning that aspect of the system.

Allen, J. (1995). Natural Language Understanding. Addison-Wesley, Reading, Mass.

Fellbaum, C. (1998). WordNet: An electronic lexical database. MIT Press, Cambridge, MA.

Flower, L. (1994). The construction of negotiated meaning: A social cognitive theory of writing. Southern Illinois University Press, Carbondale, IL.

Foltz, P. W., Kintsch, W., & Landauer, T. K. (1998). The measurement of textual coherence with Latent Semantic Analysis.

		Coherent	Incoherent	Neutral
	First	3426	3500	3555
Short	Second	3043	3154	3196
Long	First	3355	3815	3548
	Second	3159	3454	3267

	Enablement	Disablement	Re-enablement
Experiment 1A	2,074 (317)	2,281 (391)	2,096 (363)
Experiment 1B	2,236 (507)	2,413 (483)	2,228 (329)

	Target-sentences
Context sentences	actual animal Guy is agile	unassociated Guy likes music	conventional Guy is funny
actual animal "Guy likes to climb"	86% (2.18 s)	5% (3.21 s)	41% (2.65 s)
unassociated "Guy goes to concerts"	28% (4.28 s)	35% (3.53 s)	41% (3.36 s)
conventional "Guy likes to joke"	36% (3.04 s)	6% (2.46 s)	91% (2.10 s)

Experimental Conditions	Recall of clause 1	Recall of clause 2
"And condition"	3.16	4.75
"Therefore condition"	2.5	3.98
"Because Condition"	0.88	3.21

Span	Explicit	Implicit	Alternative	Control
Experiment 1 (Lexical Decision Task)
High	716 (40.5)	811 (55.3)	871 (57.3)	922 (59.2)
Low	613 (40.5)	716 (55.3)	710 (57.3)	713 (59.2)

Experiment 2 (Picture Naming Task)
High	728 (30.9)	894 (37.6)	991 (39.8)	1009(41.2)
Low	673 (30.9)	753 (37.6)	834 (39.8)	881 (41.2)

Tenth Annual Meeting of the Society for Text and Discourse

Lyon, France
July 19 – 21, 2000

	Neutral	Control	Simple	Associative	Causal
L.D.	688	677	675	645	654
Judgment	14	26	76	75	84

	C-Simple	C-Associative	C-Causal	Simple	Associative	Causal
L.D.	658	636	627	649	624	639
Judgment	46	45	77	80	81	85

	Control	Simple	Associative	Causal	Explicit
Percentage	19	68	76	77	90
Latency	-	-	2512	2205	1945

	C-Simple	C-Associative	C-Causal	Simple	Associative	Causal
Percentage	40	36	85	89	80	90
Latency	-	-	1479	1489	1411	1337

Tenth Annual Meeting of the Society for Text and Discourse

Lyon, FranceJuly 19 – 21, 2000

Lyon, France
July 19 – 21, 2000