Proceedings of the Annual Meeting of the Cognitive Science Society

The constructivist paradigm, which views learning as an active process in which students participate by engaging in activities that facilitate the construction of internal representations, is fast gaining currency as an innovative and effective way to overhaul classroom instruction. One way to promote constructivist learning is to firmly embed the acquisition of knowledge in realistic and stimulating contexts that challenge students to explore a variety of issues and employ a variety of their skills. This symposium will focus on four approaches (anchored instruction, casebased reasoning, goal-based scenarios, and problembased learning) designed to contextualize learning in this fashion. They do so in different, yet related ways. Examples of these approaches, ways in which they are similar and different, and important concerns regarding their effects on student learning and performance will be some of the issues that panelists will address. The panelists in this symposium are: Ray Bareiss, Institute for the Learning Sciences, Northwestern University; Cindy Hmelo, Janet Kolodner, and Hari Narayanan, EduTech Institute, Georgia Institute of Technology; Vimla Patel, Center for Cognitive Science, McGill University; and Susan Williams. School of Education and Social Policy, Northwestern University.

There have been a number of production system models which have recently made substantial advances in modeling higher-level cognition. These type of model offers only comprehensive approaches to the modeling of higher level cognition. This symposium will involve presentations by four exemplars of this approach to cognitive modeling (ACT, CAPS, EPIC, and SOAR). The presentations will try to illustrate the range of applications to which such models are appropriate, what the similarities and differences are among the various architectures, and what some of the interesting research questions are within each architecture.

Attempting to derive models of semantic memory using psychometric techniques has a long history in cognitive psychology dating back at least to Osgood (1957). Many others have used multidimensional scaling on human judgements of similarity (e.g., Shepard, 1962, 1974; Rips, Shoben, & Smith, 1973; Schvaneveldt, 1990). Recently, a small group of investigators have been using large corpora, 1 million to 500 million words, to develop cognitively plausible high-dimensional semantic models without the need for human judgements on stimuli. These models have become increasingly better at explaining a wide range of cognitive pheno

s symposium will address the new data and challenges presented by recent advances in brain imaging methodology. It will include tutorials on the available techniques: Positron Emission Topography (PET), Functional Magnetic Resonance Imaging (fMRJ) and combined fMRI & high density Evoked Response Potentials (ERP). The impact of these techniques on the physiological interpretation of human working memor y will be illustrated.

e present a new distributed connectionist model of the perception of spoken words. The model employs an internal representation of speech that combines lexical information with abstract phonological information. W e show how a single distributed representation of this type can form the basis for the perception of words and nonwords alike. The model is tested against lexical and phonetic decision data from Marslen-Wilson and Warren (1994). These experiments examined the integration of cues to place of articulation diuing lexical access and showed a pattern of results which proved difficult to accommodate in previous models. The use of a single, late, phonological representation allows this pattern of results to be simulated and has the potential to incorporate many other properties of the human system.

is commonly assumed that as a spoken linguistic message unfolds over time, it is initially processed by modules that are encapsulated from information provided by other perceptual and cognitive systems. W e were able to observe the effects of relevant visual context on the rapid mental processes that accompany spoken language comprehension by recording eye movements using a head-mounted eye-tracking system while subjects followed instructions to manipulate real objects. Under conditions that approximate an ordinary language environment, incorporating goal-directed action, the visual context influenced spoken word recognition and mediated syntactic processing, even during the earliest moments of language processing.

Non-cognate French-English homographs, i.e., identical lexical items with distinct meanings in French and in English, such as pain, four, main, etc., are used to study the organization of bilingual memory. Bilingual lexical access is initially shown to be compatible with a parallel search through independent lexicons, where the search speed through each lexicon depends on the level of activation of the associated language. Particular attention is paid to reaction times to "unbalanced" homographs, i.e., homographs with a high frequency in one language and a low frequency in the other. It is claimed that the independent dual-lexicon model is functionally equivalent to an activation-based competitive-access model that can be used to account for priming data that the dual-lexicon model has difficulty handling.

A distributed attractor network is trained on an abstract version of the task of deriving the meanings of written words. When processing a word, the network starts from the final activity pattern of the previous word. Two words are semantically related if they overlap in their semantic features, whereas they are associatively related if one word follows the other frequently during training. After training, the network exhibits two empirical effects that have posed problems for distributed network theories: much stronger associative priming than semantic priming, and significant associative priming across an intervenmg unrelated item. It also reproduces the empirical findings of greater priming for low-frequency targets, degraded targets, and high-dominance category exemplars.

We present an experiment where subject's subitizing performance for linear dot arrays was analyzed using Differential Time Accuracy Functions. This technique uses accuracy and reaction time data to decompose overall response latency into stimulus-limited and post-stimulus processing. Our results show that subitizing is a phenomenon produced by the effects of increased numerosity on stimulus?limited processes alone. They also suggest that the familiar guessing strategy for the largest arrays in reaction time measures of subitizing results from a reduction in post?stimulus processing. Subjects appear to extract the perceptual characteristics of all arrays but presumably fail for the largest and therefore default to guessing. Existing theories of subitizing are evaluated in light of these results.

When a wheel rolls along a flat surface, a point on its perimeter traces a cycloid trajectory. However, subjects perceive the point's path not as the cycloid, but as the curtate cycloid, containing loops where the point contacts the surface. This is the curtate cycloid illusion. I hypothesize that the illusion occurs because the cognitive system does not have sufficient activation, or capacity, to both maintain an updated representation of the wheel's translation and compute its instant centers, the point about which the wheel is rotating at a given instant. This hypothesis is supported by showing that illusion susceptibility is decreased when the competing instant center demand is reduced, either by giving subjects practice at instant center computation (Experiment 1) or by eliminating the contour containing the instant centers subjects are most likely to compute (Experiment 2). Experiment 3 demonstrates that heightened instant center demands have less effect on illusion susceptibility when they are confined to irrelevant portions of the wheel's contour. A general form of the capacity account may explain illusions in the perception of many kinematic systems and point the way toward theoretical unity in the study of the perception of motion and events.

develop a model of scan path generation based on the output of low level filters. The highest variance of Gabor jet filters computed over orientations are used as the object of attention. These points are held in a feature map which is inhibited as attention points are visited, creating a new attention point elsewhere. Scan paths generated this way can be used for recognition purposes where "single-shot" methods, such as PCA, would fail because the image is not registered.

The ACT-R production-system theory (Anderson, 1993) has been extended to include a theory of visual attention and pattern recognition. Production rules can direct attention to primitive visual features in the visual array. When attention is focused on a region, features in that region can be synthesized into declarative chunks. Assuming a time lo switch attention of about 200 msec, this model proves capable of simulating the results from a number of the basic studies of visual attention. W e have extended this model to complex problem-solving like equation solving where we have shown that an important component of learning is acquiring more efficient strategies for scanning the problem.

Both human learners and Case-Based Reasoning systems have applied metacognitive strategies such as self-questioning to improve the learning process. Whereas case-based reasoning systems do not allocate attention to reasoning strategies in order to facilitate strategy selection, previous work on attention in human thinking has focused on the selection of domain objects. We describe a computational model of metacognitive attention which integrates metacognitve approaches in case based reasoning with the concept of attention which is applied to the reasoning process itself. An example of our implementation, lULJAN, will illustrate the process of allocating metacognitive attention

Case-based comparative evaluation appears to be an important strat?egy for addressing problems in weak analytic domains, such as the law and practical ethics. Comparisons to paradigm, hypothetical, or past cases may help a reasoner make decisions about a current dilemma. W e are investigating the uses of comparative evaluation in practical ethical reasoning, and whether recent philosophical models of casuistic rea?somng in medical ethics may contribute to developing models of com?parative evaluation. A good comparative reasoner, we believe, should be able to integrate abstract knowledge of reasons and principles into its analysis and still take a problem's context and details adequately into account. TRUTH-TELLER is a program we have developed that compares pairs of cases presentmg ethical dilemmas about whether to tell the truth by marshaling relevant similarities and differences in a context sensitive manner. The program has a variety of methods for reasoning about reasons. These include classifying reasons as prin?cipled or altruistic, comparing the strengths of reasons, and qualifying reasons by participants' roles and the criticality of consequences. W e describe a knowledge representation and comparative evaluation pro?cess for this domain. In an evaluation of the program, five professional ethicists scored the program's output for randomly-selected pairs of cases. The work contributes to context sensitive similarity assessment and to models of argumentation in weak analytic domains.

n essential component of opportunistic behavior is oppor?tunity recognition, the recognition of those conditions that facilitate the pursuit of some suspended goal. Opportunity recognition is a special case of situation assessment, the pro?cess of sizing up a novel situation. The ability to recognize opportunities for reinstating suspended problem contexts (one way in which goals manifest themselves in design) is crucial to creative design. In order to deal with real world oppor?tunity recognition, we attribute limited inferential power to relevant suspended goals. W e propose that goals suspended in the working memory monitor the internal (hidden) represen?tations of the currently recognized objects. A suspended goal is satisfied when the current internal representation and a sus?pended goal "match". W e propose a computational model for working memory and we compare it with other relevant theo?ries of opportunistic planning. This working memory model is implemented as part of our IMPROVISER system.

mputer models of case-based reasoning (CBR) generally guide case adaptation using a fixed set of adaptation rules. A difficult practical problem is how to identify the knowledge required to guide adaptation for particular tasks. Likewise, an open issue for CB R as a cognitive model is how case adaptation knowledge is learned. W e describe a new approach to acquiring case adaptation knowledge. In this approach, adaptation problems are initially solved by reasoning from scratch, using abstract rules about structural transformations and general memory search heuristics. Traces of the processing used for successftil rule-based adaptation are stored as cases to enable future adaptation to be done by case-based reasoning. Whe n similar adaptation problems are encountered in the future, these adaptation cases provide task- and domain-specific guidance for the case adaptation process. We present the tenets of the approach concerning the relationship between memory search and case adaptation, the memory search process, and the storage and reuse of cases representing adaptation episodes. These points are discussed in the context of ongoing research on DIAL, a computer model that learns case adaptation knowledge for case-based disaster response planning

ime-accuracy functions are obtained by measuring the accuracy of a subject's responses at various levels of stimulus presentation time. Unlike reaction time (RT) measurements, which convey information about the entire set of processes taking place between the onset of the stimulus and the production of a response, time-accuracy functions (TAFs) focus on a subset of those processes, namely stimulus-limited processes. Stimulus-limited processes are responsible for the extraction of the perceptual information that is necessary for the elaboration of a response. Post-stimulus processes take care of selecting and executing the response based on the information extracted by stimulus-limited processes. This paper presents a method of analysis that allows us to (a) extract estimates of the duration and variance of stimulus?limited processes from individual TAFs and (b) combine these estimates with RT data in order to induce the duration and variance of post-stimulus processes. The method is illustrated with data from a subitizing (speeded enumeration) task.

A certain indecisiveness and lack of common purpose seems to be a feature of cognitive science at the moment. W e are in this paper that it can be explained in part by cognitive science's lack of success so far in connecting its scientific, computational image (better, images) of cognition to what we experience of people in ordinary life: in society, law, literature, etc. Following Sellars (1963), we call these two ways of representing cognizers the scientific image and the manifest image. The scientific image sees persons, and also artificial cognitive systems, as vast assem?blages of postulated units of some kind. In the manifest image by contrast, persons are seen as unified centre of representation, deliberation and action, able to reach focused, unified decisions and take focused, unified actions. Since the manifest image is the murkier of the two, more of the paper is devoted to it than to the scientific image. The manifest image is richer and more diverse than might at first be thought.

The field of cognitive science is inherently multi?disciplinary. However, it is unclear to what extent truly interdisciplinary work occurs in cognitive science. That is, is cognitive science merely a collection of researchers from different disciplines working separately on commo n problems? Data gathered from a recent cognitive science conference are presented. Interestingly, a significant proportion of interdisciplinary collaborations were found. Analyses were also conducted on the impact of same vs. different backgrounds on the structure of collaborations, and It was found that interdisciplinary collaborations involved more equally distributed contributions among the authors than did intradisciplinary collaborations.

An extension of Klahr and Dunbar's (1988) Dual space model of scientific discovery is presented. W e propose that, in addition to search in an experiment space and a hypothesis space, scientific discovery involves search in two additional spaces; the space of data representations and the space of experimental paradigms. That is, discoveries often involve developing new terms and adding new features to descriptions of the data, and the also often involve developing new kinds of experimental procedures. The 4-space model was motivated by the analysis of human performance in a discovery microworld. A brief description of the data is presented. In addition to the general 4-space framework, a description of the component processes involved in each of the four search spaces is also presented

The semantics of spatial relations have been intensively studied in linguistics, psychology, and cognitive neuroscience. Angle, distance, and shape are widely considered to be the key factors when establishing spatial relations. In this work an empirical study shows that previous theories overemphasize variation and we clarify the interdependencies between angle, distance, and shape with respect to the acceptability of projective relations. It turned out that the angular deviation plays the key role for relations of this class. The degree of deviation was dependent upon the extension of the reference object perpendicular to the canonical direction of the relation. There was no major effect due to the distance. However, distance interacted with the angular deviation if the located object was very close to the reference object. The experimental results can now be used as a theoretical framework for validating existing computational models of projective relations for their cognitive plausibility

This paper outlines a general framework for giving the meanings of temporal prepositions (and some related adverbials) in English. The framework takes the form of a method for translating English sentences involving these adverbials into an expressively limited temporal logic, whose operators permit only restricted quantification over subintervals of a given interval. We illustrate our approach with reference to the temporal adverbials one Monday, every Monday, on Monday, in July, for five minutes and in two hours. W e pay special attention to sentences containing multiple temporal adverbials. In particular, we show how some of our Unguistic intuitions concerning the acceptability of combinations of temporal adverbials prevent us from entertaining sentences that are either logical falsdioods or logically equivalent to simpler sentences.

Analogies between material object exchange and communication abound in figures of speech, e.g. "exchange of ideas" But, as transfer of information entails no loss of it to the donor, the obvious analogy fails. To explicate, I consider first a formal, minimal naive theory OT M of object location/possession and transfer. Failure of the obvious analogy translates as absence of any intuitable model of communication related to OT M by an isomorphism which maps people ("possessors") to people, and objects ("possessions") to ideas ("propositions", "infons"). Isomorphisms to a counterintuitive model MC M of communication and belief are, however, exhibited which map objects to people ("believers") and persons to ideas. Under the interpretation appropriate to MCM , the schemata of crucial postulates of OT M instantiate to epistemic instances of the Laws of Contradiction and Excluded Middle. MC M features complementary ideas which, as it were, appropriate or lose adherents. Empirical instantiations of this apparently counterintuitive theory are shown to occur in the lexicologies and ideologies of possession by ideas (and, perhaps, by their yet more anthropomorphic spirit avatars) and in the grammar of expressions for a change of mind. Thematic role structure, relations to "middle' constructions and, briefly, use in verbal action are discussed. I conclude that the mental leap reflected in the linguistic data warrants use of moderately formal tools to investigate open class lexica of natural languages for underiying theories.

It is generally accepted that in the semantic interpretation of compound nominals there is a set of possible relationships that could apply between the nominal constituents. This, however, has not been reflected adequately in the literaUire, which favours very deterministic processing or analyses performed on a pragmatic level. This study extends the existing set of relationships described by Levi (1978), postulating a set of rules to predict a subset of these relationships for a particular compound using a unification-based formalism with typed featurestructures. The system shows that by operating on a purely semantic level a small set of valid predicates for the meaning of the whole compound can be obtained.

Recent Positron Emission Tomography (PET) and other studies have produced detailed information about the areas of the brain involved in word association tasks, their functional roles in learning word associations, and the changes in activity in these areas during learning. We present a dynamic neurond model that replicates observed human cognitive behavior in learning word as?sociations while satisfying salient neuroanatomical and neuropsychological constraints. The model captures the observed dynamics of cortico-thalamo-basal ganglionic loops.

In drawing a figure, hemispatial neglect patients typically produce an adequate representation of parts on the right of the figure while omitting significant features on the left. This contralateral neglect is influenced by multiple spatial reference frames and by the hierarchical structure of the object(s) in the figure. The current work presents a computational characterization of the interaction among these influences to account for the way in which neglect manifests in drawing. Neglect is simulated by a "lesion" (monotonic drop-off from right to left) that can affect performance in both object-centered and viewer-centered reference frames. The joint effects of neglect in both these frames provide a coherent account of the drawing performance of a patient, JM, and may be extended to account for the copying performance of other patients across a range of objects and scenes.

fficient categorizations of complex stimuli require effective encodings of their distinctive properties. In the object recognition literature, scene categorization is often pictured as the ultimate result of a progressive reconstruction of the input scene from precise local measurements such as boundary edges. However, even complex recognition tasks do not systematically require a complete reconstruction of the input from detailed measurements. It is well established that perception filters the input at multiple spatial scales, each of which could serve as a basis of stimulus encoding. Whe n categorization operates in a space defined with multiple scales, the requirement of finding diagnostic information could change the scale of stimulus encoding. In Schyns and Oliva (1994), we showed that very fast categorizations encoded coarse information before fine information. This paper investigates the influence of categorization on stimulus encodings at different spatial scales. The first experiment tested whether the expectation of finding diagnostic information at a particular scale influenced the selection of this scale for preferred encoding of the input. The second experiment investigated whether the multiple scales of a scene were processed independently, or whether they cooperated (perceptually or categorically) in the recognition of the scene. Results suggest that even though scale perception is mandatory, the scale of stimulus encoding is flexibly adjusted to categorization requirements.

There are striking parallels between ecological psychology and new trends in robotics and computer vision, particu?larly regarding how agents interact with the environment. W e present some ideas from ecological psychology, in?cluding control laws using optical flow, affordances and action modes, and describe our implementation of these concepts in a small mobile robot which can avoid obstacles and play tag solely using optical flow. This work ties in with those of others arguing for a methodological approach in robotics which foregoes a cen?tral model/planner. Ecological psychology may not only contribute to robotics, but robotic implementations in turn provide a test bed for ecological principles and sources of ideas which could be tested in animals and humans.

When tested on siuprise or preferential looking tasks, young infants show an understanding that objects continue to exist even though they are no longer directly perceivable. Only later do infants show a similar level of competence when tested on retrieval tasks. Hence, a developmental lag is apparent between infants' knowledge as measured by passive response tasks, and their ability to demonstrate that knowledge in an active retrieval task. W e jwesent a connectionist model which learns to track and initiate a motor response towards objects. The model exhibits a capacity to maintain a representation of the object even when it is no longer directly perceptible, and acquires implicit tracking competence before the ability to initiate a manual response to a hidden object. A study with infants confirms the model's prediction concerning improved tracking performance at higher object velocities. It is suggested that the developmental lag is a direct consequence of the need to co-ordinate representations which themselves emerge through learning.

In an experiment, subjects learned about new categories for wbich tbey had prior beliefs, and made probability judgments at various points during the course of learning. The responses were analyzed in terms of bias due to prior beliefs and in terms of sensitivity to the content of the new categories. These results were compared to the predictions of four models of belief revision or categorization: (1) a Bayesian estimation procedure (Raiffa & Schlaifer, 1961); (2) the integration model (Heit, 1993, 1994), a categorization model that is a generalization of the Bayesian model; (3) a linear operator model that performs serial averaging (Bush & Mosteller, 1955); and (4) a simple adaptive network model of categorization (Gluck & Bower, 1988) that is a generalization of the hnear operator model. Subjects were conservative in terms of sensitivity to new information, compared to the predictions of the Bayesian model and the linear operator model. The network model was able to account for this conservatism, however this model predicted an extreme degree of forgetting of prior beliefs compared to that shown by human subjects. Of the four models, the integration model provided the closest account of bias due to prior beliefs and sensitivity to new information over the course of category learning.

Category formation is constrained by three factors: the perceptual structure of the domain being categorized, the limitations and biases of the learner, and the goals that trigger the learning process in the first place. Many studies of categorization have paid attention to the effects of the structure of the world and some to the biases due to the learner's prior knowledge. This paper explores the third factor: how the goals of the agent at the time of the learning episode affect what categories are formed. In particular it presents an information theoretical account that views categories as a means to increase the agent's chances of achieving its goals. One of the predictions of the theory is that information gain, the average reduction of uncertainty induced by a category, is maximized when the domain is partitioned into about 3 categories, the closest integer to the irrational number e. This prediction is confirmed by evidence derived from anthropological studies of folk classifications of animal and plants by different societies from around the world, and also by an informal observation of the behavior of cognitive scientists. Interestingly, e also emerges from optimization analyses of memory search as well as from experimental work on memory retrieval.

ople may be biased to leam categories which not only capture structure in the environment but organize this knowledge in a manner easy to use in reasoning. Concepts organized to contrast consistently on the same attributes as sister categories within a hierarchy may be particularly useful in guiding induction. W e assess whether systems of novel categories organized in this maimer were also easier to leam. Supervised concept learning was dramatically easier in the consistent over inconsistent contrast condition. W e tested whether several models of concept learning would show sensitivity to consistent contrast, as people did, including assessment of a model designed to use information about consistent contrast, TWILIX. None of the models tested (ALCOVE, rational analysis, and TWILIX ) showed muc h sensitivity to the Consistent/Inconsistent contrast. People may flexibly adjust their learning strategy to capitalize on simple regularities when available, in a manner not incorporated in these concept learning models.

In two experiments, children aged four years and adults were presented with unfamiliar stimuli. They had to segment them into relevant parts. Stimuli presented in a category shared global shape and features, but each occurrence of a potential feature was different. Results of the first experiment show that adults and children found the relevant features despite the differences between occurrences of potential features. Children's selections differed from adults' selections in terms of coherence of the segmentations. In the second experiment, the hypothesis that children used the global shape of the stimuli to find the relevant features was tested. The global shape of stimuli was manipulated in order to assess its role on feature selection. Results demonstrated that the number of incoherences produced by children increased when they could not rely on a global shape for their segmentation. The results are discussed in terms of the relative influence of configural and featural aspects of the stimuli. It is argued that adults rely more on feature identity than children when they segment stimuli into relevant features.

An experiment based on Allen's calculus and its transfer to qualitative spatial reasoning, was conducted. Subjects had to find a conclusion X rj Z that was consistent with the given premises X rj Y and Yr2 Z. Implications of the obtained results are discussed with respect to the mental model theory of spa?tial inference. The results support the assumption that there are preferred models when people solve spatial three-term series problems. Although the subjects performed the task surpris?ingly well overall, there were significant differences in error rates between some of the tasks. They are discussed with respect to the subprocesses of model construction, model inspection, validation of the answer, and the interaction of these subprocesses.

Diagram-based problem solving is an activity in which subjects solve problems that are specified in the form of diagrams. Since the diagram contains critical information necessary for problem solving, this is an activity that clearly requires reasoning with the diagram. Recent research on diagrammatic reasoning has uncovered many interesting aspects of this process. One such aspect that the authors have been exploring, by means of a set of verbal and gestural protocol analysis experiments, is the role of the diagram in guiding the reasoning process. The trajectory of reasoning is revealed both by the intermediate hypotheses gen?erated, and by the shifts of focus induced from problem solving protocols. In this paper w e focus on the protocols collected for a particularly interesting problem, one whose solution is ar?rived at through a pair of contradictory inferences. W e derived the reasoning trajectories of subjects by extracting the temporal order and spatial distribution of their intermediate hypotheses leading toward the final solution. These trajectories indicate that the spatio-temporal order of hypotheses depend on more than the device structure depicted in the diagram and inferred causation of events from the diagram. W e propose that subjects employ impUcit search strategies which together with their in?ternal goals to verify hypotheses and the need to replenish short term memory influence their reasoning trajectories

A complementary strategy can be defined as any organizing activity which recruits external elements to reduce cognitive loads. Typical organizing activities include pointing, arranging the position and orientation of nearby objects, writing things down, manipulating counters, rulers or other artifacts that can encode the state of a process or simplify perception. To illustrate the idea of a complementary strategy, a simple experiment was performed in which subjects were asked to determine the dollar value of collections of coins. In the no-bands condition, subjects were not allowed to touch the coin images or to move their hands in any way. In the hands condition, they were allowed to use their hands and fingers however they liked. Significant improvements in time and number of errors were observed when S's used their hands over when they did not. To explain these facts, a brief account of some commonly observed complementary strategies is presented, and an account of their potential benefits to perception, memory and attention.

We describe an extension of CaMeRa, a Computational model of M.ultiple Kepresentations in problem solving (Tabachneck, Leonardo. & Simon, 1994, 1995). CaMeRa provides a genera] architecture for LTM , ST M and their interactions, and illustrates how experts integrate pictorial and verbcd reasoning processes while solving problems. A linked production system and parallel network are used to further resolve the communication between pictorial and verbal knowledge by simulating how a diagram is understood by an expert. Low-level scanning processes and an attention window, based on both psychological and biological evidence, are incorporated into CaMeRa, and productions are developed that allow these processes to interface with the high-level visual rules and representations already in the model. These processes can explain interruptibility during problem solving, and show how understanding is reached when reading a novel diagram.

We offer evidence that the productive use of English regular past tense morphology (e.g., drived) results from competitions among lexical-level features within a single mechanism associative system. W e present error data from: (1) on-line elicited productions by adult native speakers (N = 51), and (2) cormectionist back-propagation networks trained to map stems and past tenses of 552 English verbs. The frequency of regularizations is analyzed in terms of item frequency, stem final alveolar consonant, and similarity in past tense mapping across "friends" and "enemies" in phonologically defined neighborhoods. All items were compiled from a lexicon of 1,191 verbs which represents a near-exhaustive listing of monosyllabic stem-past tense pairs in current American English. Results revealed striking similarities between the hiunan and simulation data. Regularizations were significantly correlated with item frequency, as well as phonological attributes of the stem. Crucially, regularization was a function of phonological similarity to frequent suffixed items, especially for irregulars that normally undergo a vowel-change. These results are incompatible with the view that regularization applies by default, independently of inter-item similarities which support the acquisition and processing of lexical items in associative systems.

A primary problem for a child learning her first language is that her ungrammatical utterances are rarely explicitly corrected. It has been argued that this dearth of negative evidence regarding the child's grammatical hypotheses makes it impossible for the child to induce the grammar of the language without substantial innate knowledge of some universal principles common to all natural grammars. However, recent connectionist models of language acquisition have employed a learning technique that circumvents the negative evidence problem. Moreover, this learning strategy is not limited to strictly connectionist architectures. What we call Incremental Distributed Prediction Feedback refers to when the learner simply listens to utterances in its environment and makes internal predictions on-line as to what elements of the grammar are more or less likely to immediately follow the current input. Once that subsequent input is received, those prediction contingencies (essentially, transitional probabilities) are slightly adjusted accordingly. Simulations with artificial grammars demonstrate that this learning strategy is faster and more realistic than depending on infrequent negative feedback to ungrammatical output Incremental Distributed Prediction Feedback allows the learner to produce its own negative evidence from positive examples of the language by comparing incrementally predicted input with actual input.

modular connectionist network is described that learns the German verb paradigm. The architecture of the network is in accordance with the rule-associative memory hypothesis proposed by Pinker (1991): it is composed of a connectionist short-term memory enabling it to process symbolic rules and an associative memory acting as a lexicon. The network successfully learns the German verb paradigm and generalizes to novel verbs in ways that correspond to empirical data. Lesioning the model gives further evidence for the rule-associative memory hypothesis: Whe n the lexicon is cut off, the network strongly overgeneralizes the regular participle, indicating that regular forms are produced with the shortterm memory but irregular forms rely on the lexicon. However, in contrast to the rule-association theory, the two paths are not strongly dissociated, but both the short-term memory and the lexicon work together in producing many participles. The success of the network model is seen as evidence that emergent linguistic rules need not be implemented as rules in the brain.

Current, rule-based theories of grammar do not provide much insight into how languages can develop new behaviors over time. Yet, textual data indicate that languages usually evolve new grammatical patterns by gradually extending existing ones. I show how a grammar model that is sensitive to prototype structure can model innovation as a process of extrapolation along salient dimensions of the category clusters. A Connectionist network provides a usefully interpretable implementation. Confirming evidence comes from a study of the development of English be going to as a marker of future tense.

One important finding in physics education is that very often students enter physics courses with misconceptions about the domain. A n often raised, but hardly ever thoroughly investigated question is whether and how students' misconceptions in physics come into play in solving formal textbook problems which ask for a precise quantitative solution. W e developed a cognitive computer model of the role qualitative physics knowledge plays in formal physics problem solving. O n the basis of the model it cannot only be hypothesized where misconceptions might come into play during formal physics problem solving, but also which correct qualitative physics knowledge should be applied instead in order to guide the use of quantitative physics knowledge efficiently and successfully. In particular, the model predicts that the application of misconceptions prevents the results of qualitative problem analyses from being exploited to construct additionally required formal, quantitative physics knowledge. A n empirical investigation confirmed that misconceptions frequently affect formal physics problem solving in the way predicted by the model. Commonly, subjects who applied misconceptions during problem solving reached an impasse when they tried to express the results of their qualitative problem analyses in quantitative terms. Most of the subjects were not able to resolve such an impasse successfully.

Examples play a critical role in guiding the acquisition of cognitive skills. W e have argued that students need to apply the knowledge gathered from studying examples to solve analogous problems for that knowledge to be effective. There is a tradeoff between the active nature of constructing solutions and the facilitating effect of guiding problem solving with a worked example. The present study examined the impact of self-explanations on the effectiveness of examples in guiding later problem solving. W e found that within a learning environment which provided direct support for the self-explanation of worked examples, such study could be as effective as direct problem solving practice.

It is hypothesized that when a set of steps in an example solution are labeled, the label can serve as a cue to the learner to group those steps and to attempt to determine their purpose. The resulting subgoal that represents the steps' purpose can aid transfer to novel problems that involve the same subgoal but require new or modified steps to achieve it. The present experiment tested the label-asgrouping-cue hypothesis by examining transfer performance by learners with different math backgrounds who studied examples that used either no labels or labels that varied in meaningfulness. Learners with a stronger math background transferred equally well regardless of the meaningfulness of the label, and better than learners not receiving labels in their examples, while learners with weaker math backgrounds transferred successfully only when they studied examples using meaningful labels. This result is consistent with the claim that the presence of a label, rather that only its semantic content, can be sufficient to induce subgoal learning if the learner has sufficient background knowledge.

When solvers have more than one strategy available for a given problem, they must make a selection. As they select and use different strategies, solvers can learn the strengths and weaknesses of each. W e study how solvers learn about the relative success rates of two strategies in the Building Sticks Task and what influence this learning has on later strategy selections. A theory of how people learn from and make such selections in an adaptive way is part of the ACT-R architecture (Anderson, 1993). W e develop a computational model within ACT- R that predicts individual subjects' selections based on their histories of success and failure. The model fits the selection behavior of two subgroups of subjects: those who select each strategy according to its probability of success and those who select the more successful strategy exclusively. W e relate these results to probability matching, a robust finding in the probability?learning literature that occurs when people select a response (e.g., guess heads vs. tails) a proportion of the time equal to the probability that the corresponding event occurs (e.g., the coin comes up heads vs. tails).

planation and evidence play important and non?interchangeable roles in argument. However, previous research has shown that subjects often confuse explanation and evidence (Kuhn, 1991). This study investigates the circumstances under which this confusion occurs. In Experiment 1, subjects generated arguments about issues of popular interest such as problems in schools and drug abuse. In Experiments 2 and 3, subjects rated the strength of evidence presented to them. The results of the protocol analyses and ratings tasks suggest that subjects tend to overestimate the strength of explanations when they lack sufficient knowledge of the domain or when they are unable to generate alternatives to the hypotheses presented to them. W e consider reasons why relying on explanations in these circumstances might be a valuable heuristic

This paper describes and evaluates a computational model of anomalous data integration. This model makes use of three factors: entrenchment of the current theory (the amount of data explained), the relative probability of the contradictory explanations (based on conditional probabilities as part of the domain-knowledge), and the availability of alternative explanations based on learning. In an experimental study we found that the enu-enchment of a theory and the availability and likelihood of an alternative explanation influenced solution speed and the correctness of inferred causal explanations. However, in detail, the single levels of both factors were not cleariy distinguishable and did not follow the predictions. These findings suggest that entrenchment itself is not a major factor in determining the difficulty of a task. Instead, we hypothesize that task difficulty is dominated by a person's ability to construct an alternative explanation of a given situation, a factor that is only indirectly related to entrenchment.

Objects with varying amounts of domain knowledge read texts on two controvCTsial issues: whether the U S should participate in the Persian Gulf War and w^iether abortion should be legal. Each text contained ten arguments for each side of the issue. Subjects with the most knowledge about the topics recalled rou^ly equal numbers of arguments from either side of the issue, while subjects with less knowledge recalled more arguments for the side they agreed with. The results were replicated with a third topic, the OJ Simpson case. The results of both experiments suggest that recall bias due to attitude may be eUminated by the possession of domain knowledge. ImpUcations for instructional programs using expert models are discussed.

Decisions made by physicians in intensive care medicine are often complex, requiring the consideration of information that may be incomplete, ambiguous, or even contradictory. Under conditions of complexity and uncertainty, individuals may cope by using simplifying decision strategies. The research described in this paper examines the strategies used by physicians in coping with complexity in decision making. Six residents (intermediates) and three specialists in intensive care were each presented with 12 cases of intensive care respiratory problems of varying levels of complexity. The subjects were asked to think-aloud as they worked through the problems and provided a management and treatment plan for each case. The audiotaped protocols were coded for key process variables in decision making and problem solving. Despite the incompleteness and ambiguity of the information available, the confidence of physicians in their decision making was consistently high. The strategies used by intermediates and experts in dealing with the more complex cases varied considerably. Expert physicians were found to focus on the assessment of the decision problems to a greater extent than intermediates. Implications for research in decision making and medical cognition are discussed.

The connectionist model lAK (Information evaluation using configurations) for classification learning is presented here. The model can be placed between feature based (e.g. Gluck & Bower, 1988) and exemplar based models (e.g. ALCOVE , Kruschke, 1992). Specific to this model is that during learning, sets of input features are probabilistically sampled. These sets are represented, in a hidden layer, by configuration nodes. These configuration nodes are connected to output nodes that represent category labels. A further characteristic of the lAK model is a mechanism which enhances retrieval of information. Simulations with the lAK model can explain different phenomena of classification learning which have been found in experimental studies: A Type 2 advantage without dimensional attention learning observed by Shepard et al. (1961); a generalisation of prototypes; a generalization based on similarity to learned exemplars; a differential forgetting of prototypes and exemplars; a moderate interference (fan effect) caused by stimulus similarity; and the missing of catastrophic interference even in A-B/A-Brdesigns.

Is implicit learning an independent and automatic process? In this paper, ! attempt to answer this question by exploring whether implicit learning occurs even despite the availability of more reliable explicit information about the material to be learnt. I report on a series of experiments during which subjects performed a sequential choice reaction task. On each trial subjects were exposed to a stimulus and to a cue of varymg validity which, when valid, indicated where the next stimulus would appear. Subjects could therefore optimize their performance either by implicitly encoding the sequential constraints contained in the material or by explicitly relying on the information conveyed by the cue. Some theories predict that implicit learning does not rely on the same processing resources as involved in explicit learning. Such theories would thus predict that sensitivity to sequential constraints should not be aftectcd by the presence of reliable explicit information about sequence structure. Other theories, by contrast, would predict that implicit learning would not occur in such cases. The results suggest that the former theories arc correct. I also describe preliminary simulation work meant to enable the implications of these contrasting theories to be explored.

Previous research has indicated that breaking a task into subtasks can both facilitate and interfere with learning in neural networks. Although these results appear to be contradictory, they actually reflect some underlying pnnciples governing learning in neural networks. Using the cascade-correlation learning algorithm, we devised a concept learning task that would let us specify the conditions under which subtasking would facilitate or interfere with learning. The results indicated that subtasking facilitated learning when the initial subtask involved learning a function compatible with that characterizing the rest of the task, and inhibited learning when the initial subtask involved a function incompatible with the rest of the task. These results were then discussed with regard to their implications for understanding the effect of knowledge on concept learning.

Athough both prepositional and perceptual representations are viewed as central to human memory, propositional representations are typically assumed to underlie conceptual knowledge. Propositional models of concepts, such as feature lists, frames, and networks, embody this assumption. Recent theories across the cognitive sciences, however, have proposed that perceptual representations are central to conceptual processing. These perceptual representations are postulated to be schematic, dynamic, and multimodal images that have been extracted from perception and experience. In the experiment reported here, we used the property verification task to determine the extent to which people use perceptual representations during conceptual processing. A regression analysis revealed two kinds of evidence for the spontaneous use of perceptual representations: First, neutral and imagery subjects showed a similar pattern of reaction times on the task. Second, perceptual variables, such as the property size, predicted verification times.

This paper describes a framework for the representation and interpretation of indirect speech acts, relating them to the politeness phenomenon, with particular attention to the Ccise of requests. The speech acts are represented as actions of a plcm hbreiry 2ind are activated on the basis of the presence of syntactic and semeintic information in the linguistic form of the input utterance. The speech act cuicdyzer receives in input the senicintic representation of the input sentence and uses the politeness indicators to chmb up the decomposition and generalization hierarchies of acts encoded in the librciry. During this process, it eliminates the indicators and collects the negated presuppositions (represented cis effects of the indirect speech act) that characterize the politeness forms. Some cycHc paths in the hierarchy allow the system to cope with complex sentences including nested politeness indicators. In the proper places of the hierarchy the semantic representation of the input sentence is converted into a domain action in order to start-up, when needed, the domsiin-level plan recognition process.

Content planning systems generate explanations to achieve a communicative intent, often with respect to a particular audience. However, current research in con?tent planning does not take into consideration the fact that an addressee may stop paying attention to an expla?nation because of boredom or cognitive overload. In this case, the generated explanation fails to achieve the com?municative intent. In this paper, we present a computa?tional representation of boredom and cognitive overload, and cast the problem of content planning as a constraint?based optimization problem. The objective function in this problem is a probabilistic function of a user's beliefs, and the constraints Jire restrictions pl

We describe an experiment to test the view that structural focusing and thematic role focusing are distinct. Subjects were presented >Mth 2-clause sentences containing because or so The first clause introduced two individuals occupying the thematic roles Goal and Source, while the subject of the second was either a pronoun or repeated name. Results showed that reading times for the second clause were facilitated when the pronomis referred to the Goal rather than Source, particularly when the clauses were connected b> so. This facilitation occurred regardless of the surface position of the Goal and regardless of tlie type of anaphor, pronoun or repeated name. With pronouns, facilitation also occurred when the antecedent was in the first position in its clause, but onh when the antecedent was the Source. With Repeated Names, reading times were slowed when the antecedent was in the first position, regardless of its thematic role. These findings suggest that there are two foci in an utterance, one containing the first noun phrase in the utterance and the other containing the preferred thematic role. W e suggest that the focus based on initial mention corresponds to the forward lookmg center described by Grosz et al. (1963) and that the focus based on thematic roles is part of the global focus (Grosz and Sidner, 1986). W e also discuss the implications of our results for Sanford and Garrod's (1981) scenario mapping model.

We investigated how college students answer questions about the content of illustrated expository text. Subjects studied illustrated texts describing causal event chains that unfold during the operation of everyday machines. Subjects subsequently provided written answers to questions about events occurring in each machine. Four types of questions were asked: why did event X occur?. how did X occur?, what are the consequences pf X occurring?, and what if X didn't occur?. In our analysis of the answer protocols, we adopted the theoretical framework of the QUES T model of human question answering (Graesser & Franklin, 1990). The present study supported predictions generated from three components of the QUEST model: question categorization, utilization of information sources, and convergence principles. Our results also revealed two novel findings. First, subjects had a bias toward sampling information from the text more than from the picture. Second, subjects tended to sample infontiation depicted in both the text and the picture.

A better understanding of the nature of consultations between professionals engaging in the collaborative process of solving complex problems — expertise in use — offers the potential to reshape our ideas about how to design computer systems that can engage in collaborative problem solving with their human cohorts. The research reported here has sought to account for key behaviors contributing to successful consultation, as identified by a cognitive task assessment of human-human consultation discourse in the medical teaching rounds setting. W e have come to view the communication acts of the presenter/investigator as evidence of his deliberate intention to indirectly construct a particular model of the patient's case — his model — in the expert's mind, resulting in two separate but related diagnostic tasks for the expert: one at the patient level and one at the presenter/investigator level. This dual-diagnostic theory of expert understanding of the presenter/investigator's communication actions is partially implemented in the RUMINATE program. The theory provides insights into the expert's capacity to model aspects of the presenter/investigator's competence — insights that contribute to our understanding of expertise embedded in the context of collaborative problem solving discourse.

This study examined the effects of peer collaboration and investigated discourse activity employed by successful and unsuccessful learners in the domain of biological evolution. Participants included 108 students from grades 9 and 12 assigned to four conditions; individual-assimilation, peerassimilation, individual-conflict, and peer-conflict. Depending on the condition, students were asked to think aloud or discuss with their peers eight scientific statements presented in the order which either maximized or minimized conflict. Several measures of prior knowledge and posttest conceptual change measures were obtained. There were no significant peer effects on conceptual change; a number of interaction effects indicated that peer collaboration was beneficial for older students and when conflict was maximized. Indepth analyses of discourse activity were conducted for four successful and four unsuccessful leamers based on posttest gain scores. Unsuccessful leamers tended to assimilate information from their peers as if it were something ateady known. Conversely, successful leamers were engaged in problem-centred discourse moves treating new information from their peers as something problematic which requires explanation. Contrasts between groups indicated significant differences in problem-centred discourse moves.

A better understanding of the nature of consultations between professionals engaging in the collaborative process of solving complex problems — expertise in use — offers the potential to reshape our ideas about how to design computer systems that can engage in collaborative problem solving with their human cohorts. The research reported here has sought to account for key behaviors contributing to successful consultation, as identified by a cognitive task assessment of human-human consultation discourse in the medical teaching rounds setting. W e have come to view the communication acts of the presenter/investigator as evidence of his deliberate intention to indirectly construct a particular model of the patient's case — his model — in the expert's mind, resulting in two separate but related diagnostic tasks for the expert: one at the patient level and one at the presenter/investigator level. This dual-diagnostic theory of expert understanding of the presenter/investigator's communication actions is partially implemented in the RUMINATE program. The theory provides insights into the expert's capacity to model aspects of the presenter/investigator's competence — insights that contribute to our understanding of expertise embedded in the context of collaborative problem solving discourse.

A network exhibits strong semantic systematicity when, as a result of training, it can assign appropriate meaning rep?resentations to novel sentences (both simple and embedded) which contain words in syntactic positions they did not oc?cupy during training. Herein we describe a network which displays strong semcintic systematicity in response to unsu?pervised tiaimng. During tradning, two-thirds of all nouns are presented only in a single syntactic position (either as gram?matical subject or object). Yet, during testing, the network correctly interprets thousands of sentences containing those nouns in novel positions. In addition, the network generalizes to novel levek of embedding. Successful training requires a corpus of about 1000 sentences, and network training is quite rapid.

Fodor (1983) has claimed that informational encapsulation of the parser is the way the language system prevents extralinguistic factors firom slowing down first pass processing. However, in a naming task where the visual probe was an appropriate or inappropriate pronoun continuation to a gerundive phrase following passages in which discourse focus and verb semantics were co-varied (Marslen-Wilson, Tyler & Koster, 1993) we found appropriateness effects which suggest a role for on-line pragmatic inference in top down control of the parser. Fodor, Garret & Swinney (1993) maintain that, though the gerund is marked as requiring a subject NP, the inferential activity underlying referent assignment does not occur until an explicit anaphor (the pronoun target) is encountered. As modularity predicts a cost associated with contacting real world information, assignment times to gerunds should take longer than assignments based on lexical information. A speeded fragment completion task was used to counter Fodor's objection to a pronoun probe and to detect differences in the times taken to make anaphor assignments. The two studies reported here used the original Marslen-Wilson et al. (1993) materials. Conect assignments in the gerundive condition ("Rurming towards...") were cost free with the exception of the condition where the pragmatically most likely subject was not in discourse focus. Latencies to initiate a completion were otherwise similar regardless of whether the to-be-completed fragment contained a gerund or a disambiguating pronoun. Furthermore, in the absence of pragmatic constraints, assignment always favoured the highlighted entity. These results reproduce the critical data from the Marslen-Wilson et al. (1993) study which demonstrates context effects on first pass processing.

In this paper we describe a general connectionist model of "learning by being told". Unlike common network models of inductive learning which rely on the slow modification of con?nection weights, our model of instructed learning focuses on rapid changes in the activation state of a recurrent network. W e view stable distributed patterns of activation in such a network as internal representations of provided advice - representations which can modulate the behavior of other networks. W e sug?gest that the stability of these configurations of activation can arise over the course of learning an instructional language and that these stable pattems should appear as articulatedattractors in the activation space of the recurrent network. In addition to proposing this general model, we also report on the results of two computational experiments. In the first, networks are taught to respond appropriately to direct instruction concerning a simple mapping task. In the second, networks receive instruc?tions describing procedures for binary arithmetic, and they are trained to immediately implement the specified algorithms on pairs of binary numbers. While the networks in these prelim?inary experiments were not designed to embody the attractor dynamics inherent in our general model, they provide support for this approach by demonstrating the ability of recurrent back?propagation networks to learn an instructional language in the service of some task and thereafter exhibit prompt instruction following behavior.

This study evaluates the effect of presuppositional information associated with contrastive stress on on-line language processing. An eye-tracking methodology was used, in which eye movement latencies to real objects in a visual display are taken as a measure of on-line reference resolution. Results indicate that presupposed contrast sets are being computed on-line, and can be used to speed reference resolution by narrowing the referential domain of an utterance. In addition, presupposed contrast sets appear to play a role in managing attention in the processing of a discourse.

The mental model theory postulates that reasoners build models of the situations described in premises. A conclusion is possible if it occurs in at least one model; it is probable if occurs in most models; and it is necessary if it occurs in all models. The theory also postulates that reasoners represent as much information as possible in implicit models. Experiment 1 showed that, as predicted, conclusions about possible situations tend to correspond to explicit models rather than to implicit models. Experiment 2 yielded a discovery: there are illusory inferences with conclusions that seem plausible but that are in reality gross errors. In such cases, as the model theory predicts, subjects judge as the more probable of two events one that is impossible. For example, given that only one of the following two assertions is true: There is a king or an ace in the hand, or both. There is a queen or an ace in the hand, or both. subjects judge that the ace is more likely to be in the hand than the king. In fact, it is impossible for an ace to be in the hand.

When people think about what might have been, they construct a mental representation of the actual state of affairs, and they generate an imaginary alternative by carrying out minimal mutations to it. When they think about how an undesirable outcome might have been avoided, they mutate the events leading to the outcome in regular ways, for example, they undo the more recent event in a series of independent events. W e describe a computer simulation of the cognitive processes that underlie these effects of temporality on counterfactual thinking that is based on the idea that reasoners construct contextualized models. W e report the results of two experiments that show that the temporality effect arises because the first event provides the context against which subsequent events are interpreted. The experiments show that when the contextualizing role of the first event is decoupled from its temporal order the effect is eliminated, for both bad and good outcomes. The results rule out an alternative explanation based on the idea that the more recent event is 'fresh' in mind. The context effect in temporal mutability may shed light on the remaining primary phenomena of counterfactual thinking.

In studies of analogical transfer, subjects sometimes fail to recognize that problems are structurally isomorphic because of differences in the problems' content. One potential explanation for this finding is that differences in content lead subjects to infer that the problems have different structures. This interpretation would be supported by evidence that subjects construct differing mental models for structurally isomorphic problems. In this study, we show that subjects' gestures reveal their mental models of problems that involve discrete and continuous change. Four subjects talked out loud as they solved a set of four problems that involved constant change. All subjects produced gestures as they spoke, and their gestures revealed both continuous and discrete mental models of the manner of constant change. O n problems constructed to evoke mental models of continuous change, subjects tended to produce gestures that incorporated smooth, continuous motions. O n problems constructed to evoke mental models of discrete, incremental change, subjects tended to produce gestures that incorporated repeated, sequential, discrete motions. Subjects' gestures sometimes provided more explicit cues to their mental models than did their speech. The results indicate that subjects sometimes constructed differing mental models for structurally analogous problems.

Mathematical problems presenting themselves in the workplace and in academia are often solved by informal strategies in addition to or instead of the normative formal strategies typically taught in school. By itself this observation does little to tell us whether, when and how much these techniques should be taught. To ground arguments about the appropriate role of altemative problem-solving techniques in education, we need to first understand the demands of the tasks they address. Our focus here is on algebra and pre?algebra, or, more specifically, on the set of problems that resist solution by more elementary arithmetic methods. W e present a task analysis of this set of problems that is based on the identification of mathematical and situational problem difficulty factors. These factors provide a framework for comparing the candidate representations and strategies to meet the demands of more complex problems. W e summarize the altemative techniques that have been observed in effective problem solving and discuss their relative strengths and weaknesses. The task analysis along with this comparative analysis provides a basis for hypothesizing developmental sequences and for informing instmctional design.

Problem-based learning (PBL) reflects new conceptions of learning that have grown out of theory and research in cognitive science. PBL has been used in medical schools to enhance the development of clinical reasoning skills and to promote the integration of basic biomedical sciences with clinical applications. In this study, the effect of PBL on the development of clinical reasoning strategies, use of scientific knowledge, and accuracy are examined on a causal explanation task. Students in problem-based curricula were compared with students in traditional medical curricula. The results indicate that PBL plays a role in facilitating the development of expertise. In PBL, students learn through the transfer of hypothesis-driven reasoning skills that result in more coherent explanations. The PBL students are better able to apply their science knowledge than nonPBL students, leading to greater accuracy of hypotheses.

This paper describes the implementation and quasiexperimental evaluation of a three-week instructional project designed in accordance with theories and assumptions of constructivism and socially situated cognition. Our goal was to develop students' ability to reason about real-life problems, where "good reasoning" was conceptualized in terms of a normative thinking model derived from cognitive research in decision making, probabilistic reasoning, and argumentation. In the spring of 1994, students in two middle school classrooms worked in teams that collected evidence, constructed arguments, and prepared presentations while engaged in activities that culminated in a mock legislative hearing. Through instruction and mentoring, students were encouraged to use statistics and probability as tools for reasoning. The effectiveness of the program was evaluated by comparing the written arguments of students from the two treatment classrooms with those of students from eight comparison classrooms. Students' arguments were scored in terms of how well they captured essential features of model reasoning and avoided particular thinking fallacies. That the reasoning abilities of students developed through social negotiation and shared problem solving was su

This paper examines performance on an intelligent tutoring system designed to teach students in a first-year algebra class to solve simple linear equations. W e emphasize the effects of requiring students to complete low-level arithmetic operations on higher-level strategic decisions. On aX + b = c problems, students who were required to perform arithmetic became less likely to solve such problems by first dividing by a than students who were not required to perform the arithmetic required to carry out the ojjeration. The shift away from this strategy is in keeping with the predictions of ACT-R. W e discuss these results in terms of the educational implications of providing computational tools to students learning basic mathematics.

Our goal is to develop a tutoring system, called CATO, that teaches law students skills of making arguments with cases. CATO's domain model provides a plausible account of legal arguments with cases, but is limited in that it does not repre?sent certain background knowledge. It is important, however, that students leam to apply and integrate this background knowledge when making arguments with cases. Given that modeling this background knowledge is difficult in an ill?stiuctured domain like legal reasoning, it is worth exploring how effectively one can teach with a model that represents ar?gument structure but relatively little background knowledge. The CATO instructional envirormient, comprising a case da?tabase and retrieval tools, enables students to apply the CATO model to a specific problem. In a formative evaluation study with 17 beginning law students, we compared instruction with the CATO environment, under the guidance of a human tutor, against more traditional classroom instruction not based on the CATO model. W e found that human-led instruction with CATO is as good as, but not better than, classroom instruction. How?ever, answers generated by the CATO program received higher grades than the students' answers, suggesting that the model can potentially be employed to teach even more effectively. Examples drawn fitom protocols show that students were able to use the CATO model flexibly and integrate background knowledge appropriately, at least when guided by a human tu?tor.

Many philosophers and psychologists argue that causal inferences are solely based on the observation of contingencies between potential causes and effects. By contrast, causal-model theory postulates that the interpretation of the learning input is governed by prior causal assumptions. Simpson's paradox is an example of this basic claim of causal-model theory. Identical observations may result in dramatically different causal impressions depending on the partitioning of the event space. Tw o experiments are presented that show that participants' assessment of a contingency between a potential cause and an effect is moderated by their background assumptions about the causal relevance of additional variables, and the ordering of the learning items. Despite the fact that all participants received identical learning inputs, participants' assumptions about the causal relevance of an additional grouping variable led either to the impression that the cause facilitated the effect or to an impression that it prevented the effect. Thus, the acquisition of new causal knowledge is crucially dependent on causal knowledge that is already available at the outset of the induction process.

Rescorla and Wagner's (1972) model of associative learning (RWM ) and Cheng and Novick's (1990, 1991, 1992) Probabilistic Contrast Model (PCM) represent competing approaches to modeling the covariation component of human causal induction. Given certain patterns of environmental inputs to the learner, these models sometimes make contradictory predictions about what will be learned. Some of these situations have been tested in Pavlovian conditioning experiments using animal subjects. W e interpret these results according to PCM, and find that they are consistent with the predictions of the model. The current experiment implements similar experimental designs as a causal inference task involving humans as subjects. Tw o experimental conditions were compared to examine each model's predictions regarding when the extinction of conditioned inhibition will occur. In one condition, the RW M predicts that a previously perceived inhibitory stimulus will be judged as less inhibitory, whereas the PC M predicts that subjects will not change their causal judgments; in the second condition, the two models make the reverse claims. The data provide strong evidence favoring the PCM

This study describes a psychological experiment on biases that people exhibit in refining probabilistic causal knowledge. In the experiment, the effect of background knowledge was shown by manipulating the causal structure of prior knowledge provided to the subjects. It was found that later training instances affected the refinement of the background knowledge in different ways depending on the causal model initially given to the subjects. The two biases found in the current experiment are (1) knowledge refinement was conservative in the sense that background knowledge was modified as little as possible to account for the observed data and (2) weakening of an existing causal relationship resulted in automatic strengthening of a related causal relationship.

Explanations of many cognitive processes, including probability judgment, rely on the construct of similarity. The present paper is concerned with the similarity-based explanation of reasoning in the conjunction task. Although high positive correlations have been found between similarity and probability judgments in this task, these alone cannot validate the assumption that similarity is judged by a process of feature comparison or that similarity judgment is an explanation of probability judgment. Preliminary results from a study in which we collected written justifications from subjects who made both types of judgment suggest that these assumptions are not tenable. Subjects cited considerations of causality and statistics ~ not just feature overlap -- when judging both similarity and probability, indicating that (1) feature comparison is only one way in which people judge similarity and (2) similarity judgment can involve processes usually associated with probability judgment. These findings suggest that the role of similarity in explaining other cognitive processes needs to be revised. It is proposed that the power of similarity and probability to predict one another can be exploited for the purpose of making either type of judgment.

The Incremental Analogy Machine (lAM) predicts that the order in which parts of an analogy are processed can affect the ease of analogical mapping. In this paper, the predictions of this model are tested in two experiments. Previous work has shown that such order effects can be found in attribute-mapping problems. In the first expenment, it is shown that these effects generalise to relational-mapping problems, when subjects' error performance (incorrect mappings) is considered. It is also found that relational mapping problems are significantly harder than attribute mapping problems. In the second experiment, it is shown using relational-mapping problems, that order effects can be demonstrated for doubles (two sentences about two indiviudals) in these problems. Throughout the paper it is shown that these results are best approximated by lAM's measure of the complexity of global mappings (the remaps complexity measure), and not as has been found previously, by a measure using frequency of remaps (the re-maps measure). The empirical and theoretical significance of these results are discussed.

Analogy, conceptual change and problem reformulation have been central components in the exploration of human problem solving. A Situation Theoretic approach is developed to model analogy and conceptual change. This model is then used to relate a problem's representation to the associated cognitive difficulty. In this Unified framework the cognitive difficulty of isomorphic problem situations is defined in terms of the task, objects and relations of the problem situation. These compo?nents are then decomposed based on an Ecological Information Processing user model. The decomposition turns a problem sit?uation the structure and dynamics of the problem; the rules or constraints which are applicable; and the necessary instruc?tions for user interaction. From this, the cognitive difficulty associated with a problem representation is shown to be largely determined by the "instructional" component.

e examined similarity judgments between simple Noun-Verb?Noun statements that were matched either in their verbs or nouns (separate matches) and made either analogous or non?analogous assertions (combined matches). An analysis of written justifications that accompanied subjects' similarity judgments revealed that matching verbs and matching nouns lead to two qualitatively different types of alignments. Matching verbs (e.g., "The carpenter fixed the chair" and "The plumber fixed the radio") led subjects to construct structural alignments and evaluate the quality of the resulting analogies (e.g., "Not analogous because plumbers don't fix radios as part of their job"). By contrast, and contrary to any traditional account of similarity as a process of comparison, matching nouns (e.g., "The carpenter fixed the chair" and "The carpenter sat on the chair") led subjects to construct thematic ahgnments and evaluate similarity based on the plausibility of the resulting causal or temporal scenarios (e.g., "He sat on the chair to see whether he fixed it well").

Eureka is a problem-solving system that operates through a form of analogical reasoning. The system was designed to study how relatively low-level memory, reasoning, and learning mechanisms can account for high-level learning in human problem solvers. Thus, Eureka's design has focused on issues of memory representation and retrieval of analogies, at the expense of complex problem-solving ability or sophisticated analogical elaboration techniques. Two computational systems for analogical reasoning, ARCS/ACM E and MAC/FAC, are relatively powerful and well-known in the cognitive science literature. However, they have not addressed issues of learning, and they have not been implemented in the context of a performance task that can dictate what makes an analogy "good". Thus, it appears that these different research directions have much to offer each other W e describe the Eureka system and compare its analogical retrieval mechanism with those in ARC S and MAC/FAC. W e then discuss the issues involved in incorporating ARC S and MAC/FAC into a learning problem solver such as Eureka.

Implementing random time effects in neural networks has been a challenge for neural network researchers. In this paper, we propose a neurophysiologically inspired temporal summation mechanism to reflect real-time random dynamic processing in neural networks. According to the physiology of neuronal firing, a presynaptic neuron sends out a burst of random spikes to a postsynaptic neuron. In the postsynaptic neuron, spikes arriving at different points in time are summed until the postsynaptic membrane potential exceeds a threshold, thus initiating postsynaptic firing. This temporal summation process can be used as a metric for deriving time predictions in neural networks. To demonstrate potential applications of temporal summation, we have employed a feedforward, two-layer network featuring a Hebbian learning rule to perform simulations using the semantic priming experimental paradigm. W e are able to successfully reproduce not only the basic patterns of observed response time data (e.g., positively skewed response time distributions and speed-accuracy trade-offs) but also the semantic priming effect and the time-course of priming as a function of stimulus-onset-asynchrony. These results suggest that the proposed temporal summation mechanism may be a promising candidate for incorporating real-time, random time effects into neural network modeling of human cognition.

We present modifications to Recursive Auto-Associative Memory which increase its robustness and storage capacity. This is done by introducing an extra layer to the compressor and reconstructor networks, employing integer rather than realvalued representations, pre-conditioning the weights and presetting the representations to be compatible with them, and using a quick-prop modification. Initial studies have shown this method to be reliable for data sets with up to three hundred subtrees.

The broad context of this study is the investigation of the nature of computation in recurrent networks (RNs). The cur?rent study has two parts. The first is to show that a R N can solve a problem that we take to be of interest (a counting task), and the second is to use the solution as a platform for develop?ing a more general understanding of RN s as computational mechanisms. W e begin by presenting the empirical results of training RN s on the counting task. The task {a b ) is the sim?plest possible grammar that requires a PD A or counter A R N was trained to predict the deterministic elements in sequences of the form a"b" * where n=l to 12. After training, it general?ized to n=18. Contrary to our expectations, on analyzing the hidden unit dynamics, we find no evidence of units acting like counters. Instead, we find an oscillator W e then explore the possible range of behaviors of oscillators using iterated maps and in the second part of the paper we describe the use of iter?ated maps for understanding R N mechanisms in terms of "activation landscapes". This analysis leads to used an under?standing of the behavior of network generated in the simula?tion study.

The mere exposure effect, in which subjects prefer items they have previously been exposed to over unexposed items, is explained as the effect of competitive learning in a connectionist network. This type of unsupervised learning will cause the network to respond more strongly to patterns on which it has been trained. If it is assumed that positive affect is proportional to total activation, then the mere exposure effect is a direct consequence of this process. The addition of a habituation rule, with a dishabituating recovery element, can also explain factors which reduce or enhance the effect. These include the effect of exposure count, display presentation sequence, the complexity of the patterns, the effect of a delay after presentation, and finally, the effects of varying exposure duration. In the case of this last factor, in addition to showing that very short exposure durations can enhance the effect, the model reveals why it may be possible to respond positively to a stimulus that one cannot recall perceiving.

Agents in the real world must be capable of autonomous goal creation. One effect of this ability is that the agent may gener?ate a substantial number of goals, but only a small number of these will be relevant at any one time. Therefore, there is a need for some heuristic mechanism to control an agent's reasoning attention. Such a mechanism is presented in this paper; alarms. Alarms serve to focus the attention of the agent on the most salient goals, and thereby avoid unnecessary reasoning. In this way, a resource-bounded agent can employ modem planning methods to effectiveness

We describe a GOMS model of a ship-board Radar Operator's behavior while monitoring air and sea traffic. GOM S is a technique that has been successfully used in Human-Computer Interaction to generate engineering models of human performance. Based on the GOM S model developed, we identified those portions of the task where an intelligent agent would be most able to assist operators in the performance of their duties, and the nature of the knowledge that will be required for the task. W e present the results of a simulated execution of the model in a sample scenario, which predicted the operator's responses with a high degree of accuracy.

this paper, we examine the components of dynamic skill acquisition using a data set collected by Ackerman (1988) with the Kanfer-Ackerman Air-Traffic Controller Task©. Our analysis indicates that subjects arc improving in both the strategies they use to solve the task and the speed with which they execute the task. One strategy that subjects develop reduces the number of oven actions required lo land a plane. Another strategy that subjects develop enables them to land more planes simultaneously. A satisfactory model of this task must include both an improved strategic component and an improved speed component. The ACT-R theory (Anderson, 1993) is well suited to model these components as it is able lo separately learn over trials which strategies are better and how to execute each more efficiently.

A formulation of learning in dynamic decision-making tasks is developed, building on the application of control theory to the study of human performance in dynamic decision making and a connectionist approach to motor control. The formulation is implemented as a connectionist model and compared with hu?man subjects in learning a simulated dynamic decision-making task. When the model is pretrained with the prior knowledge that subjects are hypothesized to bring to the task, the model's performance is broadly similar to thatof subjects. Furthermore, individual runs of the model show variability in learning much like individual subjects. Finally, the effects of various manip?ulations of the task representation on model performance are used to generate predictions for future empincal work. In this way, the model provides a platform for developing hypotheses on how to facilitate learning in dynamic decision-making tasks

What role does causal knowledge play in categorization? The current study tested the hypothesis that weight given to features is determined by the specific role they play within a causal structure. After learning typical symptoms of a disease, participants were asked to judge the likelihood that new patients had that disease. Half of the patients were missing one of the typical symptoms, and the other half had an extra symptom (a symptom typical of an alternative disease). For patients with a missing symptom, likelihood ratings were lower if the missing symptom was a cause of other symptoms than if it was an effect. However, for patients with an extra symptom, there was no difference between likelihood ratings when the extra symptom was a cause or an effect. These results suggest one mechanism underlying differences between experts and novices in categorization, and suggest an explanation for why different kinds of features (e. g., molecular or functional) are important for different kinds of categories (e.g., natural kinds or artifacts).

Analogical reasoning is the process of retrieving knowledge of a familiar problem (source analog) similar to the current problem (target) and transferring that knowledge to solve the problem. The power of an analogical reasoner thus comes in part from the ability to retrieve the "right" analog when a tai^et is specified. Indexing of analogs therefore is an important issue in analogical reasoning. This issue in fact has three different aspects: (i) indexing vocabulary, (ii) learning of the indices to a new analog, and (iii) use of indices for retrieving stored analogs. W e have been exploring the hypothesis that the reasoner's mental models of the analogs give rise to the answers to these issues. W e have tested this hypothesis in the context of analogical design of physical devices. In this paper, we describe how structure-behavior-function (SBF) models of devices help in addressing the indexing issues in analogical design. W e also describe how the IDEAL system implements and evaluates the model-based scheme to indexing and index learning.

The relative importance of meaiung (semantic context) and spaces between words during reading was investigated. Sub?jects read paragraphs of coherent or incoherent text aloud; some paragraphs were presented normally, others with spaces be?tween words removed. Coherent paragrafrfis were taken from a short story. Incoherent paragraj^ had the same words and punctuation as the coherent paragrajrfis, but the order of these words was randomized, resulting in text devoid of meaning normally provided by context and syntactical structure. As expected, spaced text was read faster and with fewer pronun?ciation errors than unspaced text, and coherent text was read faster and with fewer pronunciation errors than incoherent text, regardless of the presence or absence of spaces between words. Removing spaces slowed reading down less and caused fewer pronunciation errors when the text was meaningful (coherent), than when the text was meaningless (incoherent), so spaces helped more when the text was meaningless than when the text was meaningful. The fact that spaces between words were more important for reading meaningless text than for reading meaningful text suggests that semantics, rather than spaces, are the more important determinants of reading speed and errors.

This paper presents an integration of two approaches to complex decision-making from very different traditions: from the psychology of jury decision, the Story Model, and from the philosophy of science, the Theory of Explanatory Coherence and its con^utational instantiation, ECHO . The subjects in Pennington & Hastie (1993) generated causal "stories" to represent the events related to a particular trial. These stories were modeled with ECHO , and ECH O reached the same verdicts as did the human subjects. The ECH O simulations were also linked to the trial testimony, which, despite the inconsistent nature of the testimony, actually increased the coherence of stories for two jurors with very different verdicts. Implications

The purpose of this paper is to present and contrast two approaches to representing the structure of complex, dialectical arguments. Previous research has focused mainly on representing single arguments presented by a single arguer; this analysis examines the naturalistic give and take of dialectical argumentation among fourth graders. One approach to representing dialectical arguments is the argument network approach, which views the arguments as webs of interlocking premises and conclusions. The second approach is the causal network approach, which treats many of the ideas presented in the discussions as events linked in causal, narrative sequences. The two approaches capture different but complementary aspects of the structure of the arguments.

Does affective context automatically activate congruent information in memory (e.g. positive context/ positive information) and/or inhibit incongruent information (e.g. negative context/positive information)? Context was elicited by presenting either a positive, negative, or neutral facial expression briefly on a computer screen. Immediately after setting the context, subjects saw a positive, negative or neutral word and had to pronounce it as quickly as possible. The experiment was designed to eliminate subject strategies. The results indicated that subjects pronounce words faster in a congruent context, relative to a neutral baseline. There was no evidence of a slowing down in the incongruent context. These findings suggest that affect automatically activates congruent information in memory. Results are discussed in relation to previous mood findings which suggest that affective priming is not found in semantic tasks.

This paper describes work towards an object-oriented language for cognitive modeling. Existing modeling languages (such as C, LISP and Prolog) tend to be far removed from the techniques employed by psychologists in developing their theories. In addition, they encourage the confusion of implementation detail necessary for computational completeness with theoretically motivated aspects. The language described here (OOS) has been designed so as to facilitate this theory/implementation separation, while at the same time simplifying the modeling process for computationally non-sophisticated users by providing a set of classes of basic "cognitive" objects. The object classes are tailored to the implementation of functionally modular cognitive models in the box/arrow style. The language is described (in terms of its execution model and its basic classes) before a sketch is given of a simple production system which has been implemented within the language. W e conclude with a discussion of on-going work aimed at extending the coverage of the language and further simplifying the modeling process.

As they gain expertise in problem solving, people increasingly rely on patterns and spatially-oriented reasoning. This paper describes the integration of an associative visual pattern classifier and the automated acquisition of new, spatially-oriented reasoning agents that simulate such behavior. They are incorporated into a game-learning program whose architecture robustly combines agents with conflicting perspectives. When tested on three games, the visual pattern classifier leams meaningful patterns, and the pattern-based, spatially-oriented agents generalized fi-om these patterns are generally correct. The trustworthiness and relevance of these agents are confirmed with an algorithm that measures the accuracy of the contribution of each agent to the decision-making process. Much of the knowledge encapsulated by the correct new agents was previously inexpressible in the program's representation and in some cases is not readily deducible from the rules.

Reported properties of human decision-making under time pressure are used to refme a hybrid, hierarchical re;isoner. The resultant system is used to explore the relationships among re?activity, heuristic reasoning, situation-based behavior, seiirch, and learning. The program first has the opportunity to react correctly. If no ready reaction is computed, the reasoner acti?vates a set of time-limited search procedures. If any one of them succeeds, it produces a sequence of actions to be exe?cuted. If they fail to produce a response, the reasoner resorts to collaboration among a set of heuristic rationales. A time?limited maze-exploration task is posed where traditional AI techniques fail, but this hybrid reasoner succeeds. In a series of experiments, the hybrid is shown to be both effective and efficient. The data also show how correct reaction, time-lim?ited search with reactive trigger, heuristic reasoning, and learning each play an important role in problem solving. Re?activity is demonstrably enhanced by brief, situation-based, intelligent searches to generate solution fragments.

is study set out to apply a model of situated discourse to analyze tutorial dialogue in a complex well-defined domain of problem solving in engineering. One tutor met individu?ally with three students to teach them to solve shear force and bending moments problems. The participants' discourse and actions were analyzed according to a situated discourse model. Quantitative analyses of the tutorial dialogue af?firmed that the constraints on situated discourse processing identified in the model predicted the propositional content and conversational functions of utterances produced by both the tutor and the students. Qualitative analysis of the concep?tual content of the tutor's dialogue for Problem 1 revealed several distinct types of situation models that constituted the meaning associated with the situated discourse and action. The students initially (Problem 1) received this information, their participation consisting mostly of observing, and par?ticipating in low-level algebraic procedures with tutorial guidance. Students' problem solving and dialogue on subse?quent problems (2 and 3) displayed their ability to solve problems with some tutorial assistance. These results dem?onstrate that analysis of tutorial dialogue from the standpoint of cognitive models of discourse processing can provide de?tailed information about the conceptual situation models in?volved, and the cognitive processes used by the participants.

Structural models provide an important source of hypothetical knowledge in scientific discovery. Informal Qualitative Models (IQMs) are structural models which can be applied to weak theory scientific domains. Example models are presented for the domain of solution chemistry. These models can be systematically generated, but, due to the weak theory nature of the domains to which they are applied, they cannot be verified directly. Instead, the application of IQMs to a problem can be used to drive other scientific discovery processes; in particular, the discovery of numeric laws. The HUM E system is a discovery system based around the application of IQMs. HUME's discovery goal is to construct explanations for phenomena, such as the depression of the freezing point of salt solutions, using a variety of reasoning strategies. HUM E first attempts to explain such phenomena using a pre-existing theory. If this theory is not able to provide an explanation, the system uses a combination of theory construction and numeric law discovery. The application of IQMs provides hypotheses for use by the other two processes. Used in this way, IQM application can be seen to provide a degree of explanatory support for numeric laws which would otherwise be simply descriptive generalisations of data. An example of the application of HUM E to a problem in solution chemistry is presented.

his paper presents a psychological investigation of the informal reasoning of literary experts and students as they describe a fictional narrative. The literary situation is viewed as a communicative relation between readers and writers mediated by written text. This investigation used a task of text description and applied an explicit two-stage cognitive model of literary communication to analyze the readers' verbal protocols in terms of discursive patterns and reasoning strategies. Findings suggest that the model of the communicative context which literary experts construct for their reading is instrumental in their reasoning about the text. Students it seems are ambivalent about the author?text relationship

We propose three analytic representations of collaborative problem solving. Activity nests, a generalization of goal-subgoal trees, represent functional decompositions of task activity into components, using nesting to indicate operations that satisfy task functions. Semiotic networks, an extension of semantic networks, represent meanings as refers-to relations between symbolic expressions and other signifiers, and relations in situations and situation types, along with general relations between these meanings. Contribution Vagrants, an adaptation of contribution trees (Clark & Schaefer, 1989), represent how turn sequences collectively achieve task components. W e developed these representations to analyze how pairs of middle-school students constructed tables to represent quantitative properties of a simple physical device that models linear functions. Variations between activity nests of dififerent pairs support an explanation of activity in terms of attimement to constraints and to affordances and abilities, rather than following procedures. The semiotic networks support a hypothesis that task components are completed through accomplishing alignments of refers-to relations.which is a generalization of goal satisfaction. Similarities between the contribution diagrams support a general pattern that we call the turn structure of collaborative operations, in which task information is recognized and task operations are initiated, performed, and accepted. Interaction is organized into this structure in order to support mutually aligned intentions, understandings, actions, and agreements.

PDP model of human analogical reasoning is presented which is designed to incorporate psychologically realistic processing capacity limitations. Capacity is defined in terms of the complexity of relations that can be processed in parallel. Relations are represented in the model by computing the tensor product of vectors representing predicates and arguments. Relations in base and target are superimposed. Based on empirical evidence of capacity limitations, the model is limited to processing one quaternary relation in parallel (rank 5 tensor product). More complex relations are processed by conceptual chunking (receding to fewer arguments, but with loss of access to some relations) or segmentation (processing components of the structure serially). The model processes complex analogies, such as heat flow-water flow, and atom-solar system, while remaining within capacity limitations.

An important issue in recent work on lexical representation is whether inflected past tense forms are represented as single units or in morphologically decomposed form, and whether this varies according to the regulartity of the fonns involved. W e investigated this by looking at competitor effects between homophonic past tense forms (paced/paste, made/maid) where we varied the relative frequencies of the past tense form and its homophonic competitor. In particular, if regular forms are represented in morphologicaly decomposed form, as is widely argued, and irregular forms are listed as single units, this should lead to contrasting effects. To investigate this we used two tasks - writing to dictation and cross modal priming to compare frequency effects for regular and irregular forms. The results for both type of experiment were highly consistent, showing parallel effects of frequency for both regular and irregular forms. W e discuss the implications of this for claims about the lexical representation of morphologically complex forms.

Connecionist approaches to word recognition suggest thai the units of word identification are not part of a fixed architecture, but emerge through extracting co-occurrence regularities. One implication of this idea is that unit-status, and the size of units, may be a matter of degree. This paper investigates the possible unit status of common word collocations, such as adjective-noun pairs {nexi step, large pari) and verb-preposition combinations (look out, appear in). On analogy to the pseudo-words used in word superiority experiments, I contrasted letter detection in near-collocations {next stem, barge part) and random pairs (next role, power part) with performance on collocations (which had been defined as frequent combinations in a printed corpus). Although letter detection for collocations was not better than single words, detection was impaired for random pairs relative to single words and collocations. Near-collocations had a paradoxical effect that was only partially anticipated: an enhancing effect when letter targets were in the first word, and an inhibiting effect when targets were in the second word. Because reaction times were 400msec slower in the latter case, it was inferred that the nearcollocations have a time-dependent effect, one of initial activation of neighbors, followed by inhibition

We investigate learning a set of causally related concepts from examples. W e show that human subjects make fewer errors and learn more rapidly when the set of concepts is logically consistent. W e compare the results of these subjects to subjects learning equivalent concepts that share sets of relevant features, but are not logically consistent. W e present a shared-task neural network model simulation of the psychological experimentation.

Several studies have indicated that NLP could benefit from the inclusion of a treatment of punctuation. The main impediment to the construction of any such implementation is that there no theory of punctuation upon which to base it. More basically, little is currently known about just what punctuation marks exist, how much they are used, and how they interact with each other This study aims to answer these basic questions through the analysis of a very large corpus, and some suggestions are made for the formulation of a theory of punctuation.

This study contrasts the learning of two different kinds of motion. The first of these we call extrinsic motion, or the motion of one object with respect to another, reference object. The second we call intrinsic motion, or the motion of an object or its parts expressed with respect to the object itself. An experiment tests for people's abilities to associate these two types of motion with nouns and verbs. Subjects were presented with animated events on a computer screen accompanied by sentences involving nouns and verbs. In the learning phase, each noun and verb was related to both an extrinsic motion attribute and an intrinsic motion attribute. Subjects were then tested by presenting them with pairs of events varying on only one of these attributes and asking them which event better exemplified the meaning of a particular noun or verb. The results of this experiment demonstrate a bias to associate verbs with extrinsic motion and to associate nouns with intrinsic motion. These results suggest a division of labor between noun and verb meanings, with verb meanings specialized to encode relational information, while noun meanings are specialized to encode information about objects in isolation.

In this paper, we introduce a new model of human sentence processing. The psychological issues addressed include the use of lexical information, specifically subcategorization information, during the initial stage of syntactic structure assembly, the issue of linear parsing, i.e. immediate attach?ment of words to the sentence structure, within a head-driven grammar framewoik, and the resolution of attachment ambi?guities. W e will demonsu^te that the variety of psycholinguis?tic phenomena can be accounted for by the assumption of principled behavior of linguistic signs, which are implemented in an object-oriented fashion. The model provides a serial implementation of Parametrized Head Attachment.

Students' naive conceptions of natural phenomena have been analogized to scientific theories. This theory view does not account for the instability and inconsistency of students' explanations. A n alternative is the schema view according to which students construct explanations by instantiating explanatory patterns acquired in previous learning. In two previous studies eight explanatory schemas for evolutionary change were identified through content analysis of students' explanations. It was hypothesized that if the schemas have cognitive reality, data from explanation tasks ought be consistent with data from recognition tasks. In this study students were asked to sort 24 explanations that exemplified the eight different schemas in three different ways. A hierarchical cluster analysis shows that half of the students recognized five schemas and merged the remaining ones into one broader explanatory pattem, giving partial support for the schema view. ImpHcations of the schema view for the learning of scientific ideas are discussed.

This paper implements recent proposals for enhancing the learning of mathematics by developing statistics instruction and assessment for eighth grade students that c^italizes on the use of exemplars. The goal of instruction was for small groups to learn about statistics by engaging in hands-on activities as well as to ^ply their knowledge and skills by creating statistics projects tfiat involved designing, conducting, and presenting a mini-experimeni. Performance criteria which reflected the statistical concepts taught in the instruction were explained to students to ensure their understanding of the task (i.e., project). Groups were assigned to two treatments-exemplars and nonexemplars?Av'hich differed in the degree to which criteria modeled the processes of hypothesis generation, data collection, data analysis, and graphic representation. The effectiveness of elaborating on criteria through examples and text (i.e., exemplars) or just text (i.e., nonexemplars) for enhancing learning was examined. Both treatments demonstrated significant performance gains from pretest to posttest. However, students' understanding of representative sampling was significantly better as a result of receiving the exemplars treatment than the nonexemplars treatment. Making criteria more elaborate through examples of performance can thus enhance students' understanding of more abstfact statistical concepts such as sampling.

The paper introduces a general model of recovery from errors in parsing. The mechanism proposed returns selectively on the choice points, in order to identify the one that was badly resolved, and could have caused the error. Then, it selects an alternative previously discarded and finally selectively repairs the appropriate fragments between the ambiguous region and the breakdown region. Both the psycholinguistic and the computational features of the model are put in evidence.

Features differ in their mutability. For example, a robin could still be a robin even if it lacked a red breast; but it would probably not count as one if it lacked bones. One hypothesis to explain this differential transformability is that having bones is more critical to a biological theory than having a red breast is. W e reject this hypothesis in favor of a theory of mutability based solely on local dependency links and expressed in the form of an iterative equation. W e hypothesize that features are immutable to the extent other features depend on them and offer supporting data.

We present a model of semantic memory that utilizes a high-dimensional semantic space constructed from a co-occurrence matrix. This matrix was formed by analyzing a 160 million word corpus. Word vectors were then obtained by extracting rows and columns of this matrix. These vectors were subjected to multidimensional scaling. Words were found to cluster semantically, suggesting that interword distance may be interpretable as a measure of semantic similarity. In attempting to replicate with our simulation the semantic and associative priming experiment by Shelton and Martin (1992), we found that semantic similarity plays a larger role in priming than what they would suggest. Vectors were formed for three different types of related words that may more orthogonally control for association and similarity, and interpair distances were computed for both related and unrelated prime-target pairs. A priming effect was found for pairs that were only semantically related, as well as for word pairs that were both semantically and associatively related. No priming was found for word pairs which were strictly associatively related (no semantic overlap). This finding was replicated in a single-word priming experiment using a lexical decision procedure with human subjects. The lack of associative priming is discussed in relation to prior experiments that have found robust associative priming. W e conclude that our priming results are driven by semantic overlap rather than by associativity, and that prior results finding associative priming are due, at least in part, to semantic overlap within the associated word pairs.

Understanding language is a seemingly effortless task for people. Not only can they understand the meaning of sentences with great accuracy, they do so quickly: in most cases, people understand language in linear time. In constrast, understanding language is not so easy for computers. Even ignoring problems of accuracy, natural language processing systems ane much slower than people aje. All current NL P systems that fully analyze both the syntactic structure and semantic meaning of text fall short of human performance in this respect. In this paper, we present an attempt to develop a linear time algorithm for parsing natural language using unification grammcirs. While the computational complexity of the algorithm is, in the worst case, no better than that of many other algorithms, empirical testing indicates improved average-case performance. Although linear performance has not yet been achieved, we will discuss possible improvements that may result in an average-case linear time algorithm.

When subjects generate a detailed, memory-based description of complex visual stimuli such as faces, their recognition performance can be worse than nondescribing controls. This effect, termed verbal overshadowing. typically occurs when the stimulus is difficult to describe, not normally verbalized in detail, and when subjects are naive about the task demands. Verbal overshadowing has previously been shown to effect visually based memory (for faces and colors). This experiment was designed to: 1) detect verbal overshadowing in another sense modality, taste, and 2) to determine if domain-related expertise modulates susceptibility to verbal overshadowing. Wine tasting was chosen as a domain in which to attempt to control subjects' relative levels of verbal and perceptual expertise. Based on suggestive data from previous face recognition studies, it was hypothesized that subjects whose perceptual expertise was greater than their domain-related verbal expertise (termed Intermediates) would show verbal overshadowing. On the other hand, subjects with relatively equal perceptual and verbal expertise, either low/low (Novices) or high/high (Experts) would not show verbalization effects. After tasting a target red wine Verbalization subjects wrote detailed taste descriptions from memory while controls participated in an unrelated verbal task. All subjects then attempted to identify the target wine from among three foils. As predicted, the verbalizing Intermediates performed significantly worse than the nonverbalizing controls on Trial 1. No-effect of verbalization was observed for either the novices or experts. The results are explained in terms of the differential development of perceptual and verbal skills in the course of becoming an expert.

There is currently a competition between two theories that propose to explain the cognitive phenomenon of analogy: Dedre Centner's Structure-Mapping Theory and Douglas Hofstadter's theory of Analogy as High-level Perception. W e argue that the competition between the two theories is ill-founded because they arc after two different aspects of analogy: structure-mapping seeks a "horizontal" view of analogy where the phenomena is examined at the level of already existing psychological representations, and where the task is to identify what processes are common to all or most analogy function; High-level Perception, on the other hand, seeks a "vertical" view of analogy in which the goal is to explain the processes that make up the construction of represenUtions. An integrated theory of analogy should encompass both horizontal and vertical views.

Because meaning is both the common outcome and the typical goal of language processing, including reading, semantic processes have received a privileged position, especially in cognitive science accounts that emphasize semantic, goal driven components in language. Even in accounts of written word identification, a "low-level" process, it is typical to assume that semantic outputs are achieved with optional contributions of phonology. Our goal here is to present evidence for an alternative perspective, one that gives phonology a central rather than a peripheral, optional role in word identification. W e first briefly discuss a writing system comparison that is important to this perspective. W e then summarize recent published and unpublished research that gives definition to our conclusion that phonology is a central and universal component of word reading.

A mathematical basis is proposed for the distinction between associative and relational (symbolic) processing. Associations can be contrasted with relations in terms of ordered pairs versus general ordered N-tuples, and unidirectional access versus omnidirectional access. Relations also have additional properties: they can exhibit predicate-argument bindings, they can be arguments to higher-order structures, and they can participate in operations of selection, projection, join, union, intersection, and difference. Relations can be used to represent structures such as lists, trees and graphs, and relational insUnces can be thought of as propositions. Within neural net architectures, feedforward networks can be identified with associative processing, and tensor product networks with relational processing. Relations have the essential properties of symbolic processing; flexibility, accessibility, and utility for representing complex data structures.

We present a model of graph understanding and describe our implementation of the model in a computer program called SKETCHY . SKETCH Y uses a combination of general graph knowledge and domain knowledge to describe graphs, answer questions, perform comparative analyses, and detect contradictions in problem solving assumptions. SKETCH Y has generated reasonable graph summaries for 65 graphs from multiple domains. SKETCHY illustrates the robustness of our model of graph understanding.

I present a theory of discourse interpretation based on the hypothesis that the common ground of a conversation contains a record not only of complete speech acts, but, more in general, of each action of uttering a contribution to the conversation: single words, word fragments, and fillers. I call the action of uttering a "minimar contribution a MICRO CONVERSATIONAL EVENT. This model can serve as the basis for accounts of reference resolution in spoken conversations, as well as the interaction between parsing, repair, and reference resolution.

Letters are detected more efficiently among digits than among letters. This alphanumeric category effect suggests an architectural distinction between letter and number representation in human vision and dissociations between letter and number recognition following brain damage support this interpretation. Because letter and number recognition are not innate, this implies that experience can shape the functional architecture of vision. A possible explanation is that letters co-occur with letters in the environment while numbers co-occur with numbers; such statistics cause segregation of letter and number representations in artificial neural networks. To test the general hypothesis that environmental statistics affect the architecture of vision, and the specific hypothesis that within-category cooccurrence causes the alphanumeric category effect, we measured the effect in foreign mail sorters who process Canadian zip codes (which violate the co-occurrence statistics) and in control subjects. As predicted, foreign mail sorters showed a smaller category effect

The subject of this paper is the role of transferable commonsense principles in the acquisition of gameplaying expertise. W e argue that individuals skilled in a domain develop expertise because they know and apply these principles, and that most game-playing programs do not play like people. The paper describes Hoyle, a model of an expert game player that relies on the use of commonsense principles, limited memory, and useful knowledge to learn to play two-person, perfect information finite-board games expertly. W e then describe an experiment in which human subjects played three such games against a computer expert. After playing these games, the subjects evaluated Hoyle's game-playing principles in the context of their own behavior Verbal protocols and subjects' evaluations revealed considerable overlap between the principles preferred by our subjects and those preferred by Hoyle. Using learning time as a measure of difficulty, the subjects' performance and Hoyle's performance ordered the three games identically. This experiment also revealed differences in the use of gameplaying principles between skilled and unskilled players: skilled players judged the game-playing principles to be more effective than did unskilled players, skilled players used several different principles while unskilled players relied on one principle, and skilled players anticipated their opponent's moves while unskilled players merely reacted.

This paper describes a model that learns to pronounce English words. Learning occurs in two modules: 1) a rule-based module that constructs pronunciations by phonetic analysis of the letter string, and 2) a whole-word module that leams to associate subsets of letters to the pronunciation, without phonetic analysis. In a simulation on a corpus of over 300 words the model produced pronunciation latencies consistent with the effects of word frequency and orthographic regularity observed in human data. Imphcations of the model for theories of visual word processing and reading instruction are discussed.

An experiment is reported which uses a rephrasing task to investigate factors affecting the formation of initial mental models. It was found that both the syntax and the thematic content of the rule affect the initial model set^ formed: the syntax determines the form of the initial model set and the semantics add to this initial set through the representation of subjects' prior knowledge about the situation in question. Specifically, causal content invokes general knowledge about causal relationships which leads to the addition of models representing counterfactual situations in the initial model set. In comparison, familiar content invokes specific knowledge which leads to the completion of existing models in the initial set. Thus, our experiment enables an extension of mental models to be made that accounts for the diffoential effects of general and specific prior knowledge.

The paper describes a theory of temporal reasoning and its implementation in a computer program. The theory postulates that individuals construct mental models, and it predicts that inferences that call for only one model to be constructed, such as: a happens before b. b happens before c. d happens while b. e happens while c. What is the temporal relation between d and e? will be easier than those that call for multiple models, such as a problem identical to the previous one except for its first premise: a happens before c. Experiment 1 showed that subjects were faster and more accurate with one-model problems than with multiple-model problems. They look more time to read a premise leading to multiple models than the corresponding premise in a one-model problem. Experiment 2 showed that if the question came first and was presented with all the premises, then subjects can ignore an irrelevant premise. As predicted, the difference between one-model and multiple-model problems with valid conclusions then disappeared. Experiment 3 showed that the size of a model, i.e., the number of events in it, and the distance apart of the critical events, also affected performance.

This study has the descriptive aim of showing if and how epistemic procedures typical to mathematical reasoning can be practiced by children when they are in a social situation that supports their individual linguistic and cognitive activity. The present paper consists of a fine-grained analysis confronting argumentative skills and epistemic actions of a group of four students functioning in a Grade 9 mathematics class. The four students were presented with a mathematical problem-situation typical of a one year long experiment whose domain was an introductory course about functions. This activity was typical in the sense that: (i) it demanded inquiry; (ii) students worked in groups; (iii) they had computerized tools at their disposition; (iv) they were invited to discuss their work in a whole class forum. The role of the technological tools as a trigger for the application of argumentative skills is investigated.

Any designer of intelligent agents in a multiagent system is faced with the choice of encoding a strat?egy of interaction with other agents. If the nature of other agents are known in advance, a suitable strategy may be chosen from the continuum be?tween completely selfish behavior on one extreme and a philanthropic behavior on the other. In an open and dynamic system, however, it is unrealis?tic to assume that the nature of all other agents, possibly designed and used by users with very dif?ferent goals and motivations, are known precisely. In the presence of this uncertainty, is it possible to build agents that adapt their behavior to interact appropriately with the particular group of agents in the current scenario? W e address this question by borrowing on the simple yet powerful concept of re?ciprocal behavior. W e propose a stochastic decision making scheme which promotes reciprocity among agents. Using a package delivery problem we show that reciprocal behavior can lead to system-wide co?operation, and hence close to optimal global perfor?mance can be achieved even though each individued agent chooses actions to benefit itself. More inter?estingly, we show that agents who do not help others perform worse in the long run when compared with reciprocal agents. Thus it is to the best interest of every individual agent to help other agents.

Newstead (1989) reports both graphically and sententially elicited data on the interpretation of quantifiers by logically naive undergraduate students. The sentential elicitation method fails to make the critical distinction between entailment relations between sentences, and mith-value-in-a-model relations between sentences and diagrams. The present study modifies the elicitation technique and shows that die resulting sen?tential data can be insightftiUy described in terms of broad ten?dencies of response (to over- or under-infer) interacting with highly specific grammatical sU^ctures (subject/predicate rela?tionship). The resulting categorisation of subjects into four groups is then predictive of graphically elicited behaviour These results are interpreted by conu-asting expository and deductive discourse, and proposing that students initially assimilate the latter to the former

An explanatory model of ambiguity resolution in human parsing must denve a multitude of preference behaviors from a concise computational framework. One behavior that has been difficult to account for concisely is the preference to interpret an ambiguous phrase as an argument of a predicate, rather than as a modifier that is less integrally related to a phrase (an adjunct). Previous accounts of the argument preference have rehed on assumptions about adjuncts requiring a more complex structure or entaiJing a delay in their mterpretation. This paper explores a more fundamental distinction between arguments and adjuncts—that the numberof potential arguments of a predicate is fixed, while the number of adjuncts for a phrase is unpredictable. This simple difference has important computational consequences withm the competitive attachment model of human parsing. The model exhibits a preference for arguments over adjuncts due to the necessary differences in competitive properties of the two types of attachment site. The competitive differences also entail that adjuncts accommodate more easily than arguments to contextual effects. The model thus provides a concise and explanatory account of these argument/adjunct asymmetries, avoiding the unnecessary structural or interpretive assumptions made within other approaches.

This paper claims that, contrary to the "Theory oriented" approach to cognitive development and instruction, children's informal concepts play important roles in learning, and reports two cases that support the constructivist view of learning. It is widely believed that children's knowledge about various domains is organized into coherent systems, i.e., theories. Although this approach provides a new perspective on knowledge organization, too much emphasis on the conceptual difference makes the interaction of prior knowledge and learning materials impossible. Without the interaction, learned rules remained uninterpreted. Consequently they can be applied only to a restricted set of problems. A case from mathematics revejJed that students' informal concept of concentration can be bridged to the formad one, by rewording quantitative terms in problems with quzditative terms. A case from physics showed that by combining fragmentJiry understandings, students could acquire the concept of force decomposition which is difficult to learn by formal instruction. Finally, instructional techniques are proposed that make use of informal concepts as a partial base analog to enrich students' understanding.

A review of empirical work suggests that the lexical representations of a bilingual's two languages are independent (Smith, 1991), but may also be sensitive to between language similarity patterns (e.g. Cristoffanini, Kirsner, and Mi lech, 1986). Some researchers hold that infant bilinguals do not initially differentiate between their two languages (e.g. Redlinger & Park, 1980). Yet by the age of two they appear to have acquired separate linguistic systems for each language (Lanza, 1992). This paper explores the hypothesis that the separation of lexical representations in biUnguals is a functional rather than an architectural one. It suggests that the separation may be driven by differences in the structure of the input to a comm o n architectural system. Connectionist simulations are presented modelling the representation of two sets of lexical information. These simulations explore the conditions required to create functionally independent lexical representations in a single neural network. It is shown that a single network may acquire a second language after learning a first (avoiding the traditional problem of catastrophic interference in these networks). Further it is shown that in a single network, the functional independence of representations is dependent on inter-language similarity patterns. The latter finding is difficult to account for in a model that postulates architecturally separate lexical representations.

This paper examines the tremendous similarities between the Parallel Constraint Satisfaction Processes that are a central part of many connectionist models and the Gestalt principles that played a central role in the history of Psychology. Gestalt Psychology played a major role in a number of areas in psychology, such as perception, reasoning and problem solving, causal reasoning, and many key aspects of social psychology, such as social perception, group interaction, and belief consistency. Many of the key assumptions of Gestalt Psychology have resurfaced in recent connectionist models. W e propose that Parallel Constraint Satisfaction Processes provide a computational implementation of many of the central principles of Gestalt Psychology. In this paper we discuss the clear parallels between each of five key assumptions of Gestalt Psychology and aspects of Parallel Constraint Satisfaction Processes. The five assumptions we examine are: (1) psychological processing can be treated as interactions in fields of forces, (2) psychological processing is holistic, (3) the whole is greater than the sum its parts, (4) the importance of the structure of cognitive elements; how things are connected and related, and (5) the emphasis on cognitive dynamics, and such concepts as change, equilibrium, and tension.

Dual space models of problem solving (e.g., Simon & Lea, 1974; Klahr & Dunbar, 1988) assume that the problem space for a task consists of two spaces: an hypothesis space and an experiment space. In hypothesis space, hypotheses about rules governing the task are generated, which can then be tested in experiment space. However, experiment space can be searched by applying the operators even without knowledge about the task. W e predicted that people searching hypothesis space would learn more about the task. To test this claim, two experiments were performed in which subjects had to learn to control a system consisting of three input variables that had unknown links to three output variables. Subjects first explored the task, then they had to reach goal states for the output variables. In both experiments subjects were presented with an hypothesis about one of the links, which should foster search of hypothesis space. In Experiment 1, hypothesis instruction improved performance and we showed that it had a similar effect to a manipulation of goal specificity, suggesting that both factors improve learning by encouraging search in hypothesis space. In Experiment 2 subjects were given a correct hypothesis or an incorrect hypothesis. Both groups performed better than an appropriate control. Thus instructions that encourage hypothesis testing appear to improve learning in problem solving.

A central issue in cognitive science is whether learning and processing constraints are particular to domains or whether they generalize across domains. In this paper the domain-generality of a particular type of constraint, linear separability, was examined. Prior research has found that decisions in the social domain are often consistent with linear separability but this is rarely true of decisions in the object domain. Two experiments were conducted to examine the generality of this result by using fiindamentally different types of social and object materials than have been used in previous research. In both experiments different integration strategies were observed in social and object domains, and as in prior research many more Summation sorts occurred with social materials. These results indicate that previous differences that have been observed between object and social domains generalize to very different types of object and social materials. At a general level the results indicate that the structure of knowledge varies with domain, and consequently it will be difficult to formulate domaingeneral constraints in terms of abstract structural properties such as linear separability.

Previous computational theories of problem solving have not accounted for the occasional display of accelerated problem solving by humans working on conceptually hard problems. Researchers refer to this behavior as insight. Kaplan and Simon describe insight as the selection of a good representation of the problem by the problem solver. They propose a dual-state space theory, meta-space theory, to explain insight (lO^lan and Simon, 1990). W e show that meta-space theory is unfalsifiable. W e then show that the nature of meta-space theory makes it superfluous for the study of human problem solving.

It is widely acknowledged that categories have many functions, but few studies have actually addressed the impact of these functions on the way categories are learned. For instance, many categorization experiments predominantly rely on classification-based incremental learning. The problem with this approach is that it implicitly assumes that the function of categorization is separable from the way that categories are learned. In this study, we examined the relation between learning and the subsequent use of categories by contrasting three types of category-learning methods — inference-based, classification-based, and a combination of these methods. The results of the experiment indicate that there is an intricate relationship between category-learning and subsequent use of the category. The results further suggest that different processing modes may have been adopted by subjects in the different learning conditions.

Numerical tasks with Arabic numerals involve the integration of internal and external information and the interaction between perception and cognition. 2-digit number comparison task was selected to study these integration and interaction processes. To compare the magnitudes of two 2digit Arabic numerals, we can (1) compare them digitbydigit sequentially, (2) compare corresponding digits in parallel, or (3) encode them as an integrated representation and compare the whole numerical values. Previous studies showed that 2- digit comparison was holistic when target numerals were compared with a standard held in memory. In our experiment target numerals and standards were presented on the same external display at the same time. Instead of a holistic comparison, we found that 2-digit comparison was a combination of sequential and parallel comparisons. The implications of this discrepancy were discussed in terms of the interplay between perception and cognition.