Proceedings of the Annual Meeting of the Cognitive Science Society

The dominant model of cognition is based upon amodal symbol systems. We support an alternative model that places embodiment at the center of cognition. On this view, sensorimotor experiences of actions form the basis of linguistic and nonlinguistic understanding.

As is well known, teaching cognitive science presents some special challenges, whether to undergraduates or to graduate students. In this symposium, we aim to explore some of the problems that teaching cognitive science presents in the class- or seminar room and identify some of the special opportunities.

This symposium reports research concerned with developing tools and models for the analysis of cognitive processes in interactive learning situations. Results of several studies of natural, interactive task-oriented learning situations including face-to-face learning situations and learning over a network are compared. The results show how interactive, situated discourse is used to support processes of knowledge construction and problem solving within these learning environments.

Instead of subscribing to the view that people are unable to perform Bayesian probabilistic inference, recent research suggests that the algorithms people naturally use to perform Bayesian inference are better adapted for information presented in a natural frequency format than in the common probability format. We tested this hypothesis on the notoriously difficult three doors problem, inducing subjects to consider the likelihoods involved in terms of natural frequencies or in terms of probabilities. We then examined their ability to perform the mathematics underlying the problem, a stronger indication of Bayesian inferential performance than merely whether they gave the correct answer to the problem. With a robustness that may surprise people unfamiliar with the effects of information formats, the natural frequency group demonstrated dramatically greater normative mathematical performance than the probability group. This supports the importance of information formats in a more complex context than in previous studies.

Instead of subscribing to the view that people are unable to perform Bayesian probabilistic inference, recent research suggests that the algorithms people naturally use to perform Bayesian inference are better adapted for information presented in a natural frequency format than in the common probability format. We tested this hypothesis on the notoriously difficult three doors problem, inducing subjects to consider the likelihoods involved in terms of natural frequencies or in terms of probabilities. We then examined their ability to perform the mathematics underlying the problem, a stronger indication of Bayesian inferential performance than merely whether they gave the correct answer to the problem. With a robustness that may surprise people unfamiliar with the effects of information formats, the natural frequency group demonstrated dramatically greater normative mathematical performance than the probability group. This supports the importance of information formats in a more complex context than in previous studies.

Two experiments investigate the conjunction fallacy (judging that conjunctive probabilities are higher than the probabilities of the constituents). The conjunction fallacy was much less for P(E|C) tasks than for P(C|E) tasks. The results are explained in terms of the way people interpret the conditional probabilities. We argue that people prefer to reason from cause to effect (cause-to-effect reasoning heuristic), and for that reason, the instructions given for P(C|E) tasks were misinterpreted, resulting in apparent fallacy. In addition, we provide evidence showing that likelihood judgments are higher with more evidence (more-is-better heuristic).

This paper describes an approach to cognitive engineering which promotes a symbiosis between the theories and methodologies of cognitive psychology and the practices of human-computer interaction design. We ground the description of our approach in a particular design problem: the development of computerised decision support in medical intensive care. We review the psychological literature of medical reasoning and decision making, highlighting its potential to inform the design of medical computerised aids. We also discuss how addressing this design problem may in turn benefit cognitive theory. This is followed by a brief description of our proposed methodology.

This study examined how changes in children's problem encoding influenced their strategy selection. Fourth-grade students {N=51} solved six mathematical equivalence problems (e.g., 3+4+5=_+5) in a pretest. Children's problem encoding was then manipulated in one of two ways, or was not manipulated in a Control group. In the Subtle group, children solved four additional problems with the equal sign highlighted in red. In the Direct group, children solved the same four problems, and were directed to notice the equal sign in each problem. Children then solved six problems in a posttest, and did so again four weeks later in a follow-up test. The strategies children conveyed in their spoken and gestured explanations were assessed. Children in the Direct group considered multiple strategies for the posttest problems more often than children in the other groups, as reflected in their spoken and gestured explanations. Children in the Direct group were also most likely to generate gestured strategies and to abandon verbal strategies over the course of the study. These findings suggest that changes in problem encoding lead to changes in strategy selection.

Events appear to be represented distinctly in memory in large numbers at a fine grain, even in tasks in which memory retention is not a primary performance measure. In Experiment 1, participants classified character strings in sequences governed by randomly-alternating instructions. Response times were fastest near the start of a sequence, slowed gradually throughout the sequence, then sped up again near the start of the next sequence. This speedup and gradual slowdown were modeled in the ACT-R architecture as a combination of priming and interference effects in episodic memory. The model correctly predicts the absence of these effects in Experiment 2. in which the instruction must be inferred from the trial stimulus and hence is not a source of priming. These findings suggest (a) that episodic encoding is a pervasive side effect of cognitive performance; (b) that elements of episodic memory interact through priming and interference—effects traditionally associated with semantic memory; and (c) that brief interruptions of task performance have more complex effects than previously documented.

In this paper, we make a methodological point concerning the contribution of the representation of the output of a neural network model when using the model to compare to human error performance. We replicate part of Dell, Juliano & Govindjee's work on modeling speech errors using recurrent networks (Dell et al., 1993). We find that 1) the error patterns reported by Dell et al. do not appear to remain when more networks are used; and 2) some components of the error patterns that are found can be accounted for by simply adding Gaussian noise to the output representation they used. We suggest that when modeling error behavior, the technique of adding noise to the output representation of a network should be used as a control to assess to what degree errors may be attributed to the underlying network.

There is growing interest in exploring the interaction of semantics with outlier information sources in sentence processing. Tliis study explores the role of noun animacy in these processes in combination with syntactic (word order) and morphological (mmiber marking) factors by means of an on-line agency assignment task in Bulgarian. Unlike smiilar studies, the preparation of the stimuli follows a rigorous series of pro-tests which reveal the gradable nature of animacy and attempt to account for the interaction of verb semantics (agency reversibility) with noun semantics (anmiacy contrast). Results confirm expectations of tlie high significance of agreement in Bulgarian and present a challenge to the binary view of animacy. Some of tire predictions on cue interaction within the Competition Model are put to the test and generally confirmed.

Most models of object recognition assume that shape is the primary dimension of recognition and that color and texture play only a secondary role. One reason for this could be that color and texture are generally less diagnostic for recognition and so it would be comparatively more difficult to find evidence of their usage. Another, but as yet unexplored reason for their secondary role, is that color and texture differences are not as well perceived at short exposures of stimuli. We report two experiments that address the perception (as opposed to the usage) of dimensions over the time course of visual processing.

Levin (1993) has proposed a semantic distinction between two types of motion verbs: manner of motion verbs and inherently directed motion verbs. In contrast, Jones (1995) has argued that this distinction is better accounted for by syntactic principles. Two simulations are presented that demonstrate that verb representations from the Hyperspace Analogue to Language (HAL) model of memory (Burgess & Lund, 1997a; Lund & Burgess, 1996) are sensitive to the distinction between these two verb classes. The second simulation shows that this effect is not due to word frequency differences. The final experiment uses human judgments of concreteness, imageability, and familiarity on these verbs to provide further data on the particular semantic characteristics that may be salient to the language user. We argue that these results provide empirical support for Levin's semantic distinctions.

This study examined the problem solving strategies used by staff radiologists and radiology residents during the interpretation of difficult mammograms. Ten radiologists and ten residents diagnosed 10 cases under two experimental conditions (authentic and augmented). In the authentic condition, standard unmarked mammograms were used. Mammographic findings were highlighted on a second set of the same cases for the augmented condition. Verbal protocols were analyzed and revealed that mammography interpretation was characterized by a predominant use of data-driven or mixed-strategies depending on case typicality and clinical experience. Repeated measures ANOVAs revealed that the radiologists scanned the cases significantly faster than the residents. No group differences were found in the number of radiological findings, radiological observations, and number of diagnoses across experimental conditions. Frequency analyses revealed that regardless of experimental condition both groups (a) used the same types of operators, control processes, diagnostic plans, (b) committed the same number of errors, and (c) committed case-dependent errors. Overall, the fact that few differences were found between the groups on the various measures may be due to the fact that mammogram interpretation is a well-constrained visual cognitive task. The results have been applied to the design of a computer-based tutor for training residents to interpret mammograms. Future empirical directions include building a more comprehensive model of the perceptual and cognitive processes underlying mammogram interpretation by converging eye-movement, cortical activation (e.g., fMRI) and verbal protocol data.

The purpose of this paper is to present a cognltively-based and empirically-derived approach for the design of the RadTutor, a prototype computerized tutor to train radiology residents in diagnosing mammograms exhibiting breast diseases. A multitude of computer-based radiology training environments have recently been developed with the objective of supporting the acquisition of radiological expertise. In general, however, these systems have failed in several aspects including a failure to incorporate theoretical perspectives and empirical findings to the design of these systems. This paper outlines the conceptual framework for the development of the prototype which includes; (1) a discussion of the objectives and goals of the radiology residency training program, (2) a review and critique of existing computer-based radiology training environments, (3) a synthesis of an expert-novice study aimed at attaining a cognitive model of problem solving in mammogram interpretation (Azevedo, 1997), (4) a description of the results of analyses of authentic radiology resident teaching rounds, and (5) deriving instructional principles for the design of the mammography tutor.

This paper describes an implemented computational model of lexical development for the case of action verbs. A simulated agent is trained by an informant labeling the agent's actions (here hand motions), and the system learns to both label and carry out similar actions. The verb learning model is placed in the broader context of the NTL project on embodied natural language and its acquisition. Based on experimental results and projections to the full range of early lexemes, a significantly enriched model is proposed and its properties discussed.

Wordlikeness, which is generally equated with phonotactics, is becoming an increasingly important variable in the study of language acquisition and processing as well as in the context of verbal short term memory. Past research has sought to establish phonotactic knowledge (knowledge of the possible sequences of sounds within a language) as a distinct kind of knowledge above and beyond knowledge of individual lexical items. It is unclear, however, how separate phonotactic and lexical knowledge really are; conceivably there could be effects of similar sounding lexical neighbors on perceived wordlikeness. We report empirical evidence and analysis demonstrating independent contributions of phonotactics and of lexical neighbors in accounting for wordlikeness ratings, a finding with both methodological and cognitive implications.

Artificial grammar learning provides a principled experimental framework to investigate the roles of similarity and rule-induction mechanisms in category generalisation. Past attempts to disentangle these two mechanisms may be criticised for employing insensitive measures of similarity with little theoretical or empirical motivation, for failing to achieve independent measures of the effects of similarity and rule-induction components, and, with several notable exceptions, for confining stimuli to the domain of letter strings. The present work reports on two studies of artificial grammar learning using a standard grammar to arrange nested geometric shapes (Experiment 1) and angles between connected lines (Experiment 2). Grammaticality judgements for novel items are significantly above chance in both experiments. Similarity judgements for pairs of stimuli are used as the basis for modelling grammaticality judgements, using an exemplar-based model of categorisation. We test for independent contributions of similarity and rule-induction mechanisms by fitting nested regression models. Similarity is significant in accounting for grammaticality judgements in both experiments. Rule-induction has an additional, independent effect in Experiment 2, but not in Experiment 1. We discuss the implications of these results and their relationship to previous studies.

This paper describes a computational mode! of spatial learning and localization in rodents. The model is based on the suggestion (based on a large body of experimental data) that rodents learn metric spatial representations of their environments by associating sensory inputs with dead-reckoning based position estimates in the hippocampal place cells. Both these sources of information have some uncertainty associated with them because of errors in sensing, range estimation, and path integration. The proposed model incorporates explicit mechanisms for information fusion from uncertain sources. We demonstrate that the proposed model adequately reproduces several key results of behavioral experiments with animals.

Models of perceptual systems customarily characterize their maximally efficient operation in optimal circumstances. Another engineering consideration - graceful failure - is usually ignored. Three experiments on spontaneous speech show that on-line speech recognition fails gracefully by making us deaf to the words in reparanda. the items which must be expunged to restore disfluent utterances to fluency. Experiment 1 uses word-level gating of fluent and disfluent utterances to show that disfluencies principally disrupt normal late recognition (Bard, Shillcock & Altmann, 1988) of words in reparanda. Experiment 2 shows that in more natural listening conditions, attention to continuing material and additional effects of repetition deafness (Miller & Mackay, 1996) make recall of the same words even more unlikely. Experiment 3 shows that the results are not attributable to the clarity of the lost words. Finally the relationships among late recognition and various kinds of disfluency deafness are discussed.

An implemented AI reasoning system called ATT-Meta is sketched. It addresses not only AI issues but also ones that are salient in psychology, philosophy, cognitive linguistics, discourse pragmatics and other disciplines. These issues include the Simulation-Theory/Theory-Theory debate and Fauconnier and Turner's notion of concepmal blending. The system performs metaphor-based reasoning and reasoning about mental states of agents; in particular, it performs metaphor-based reasoning about mental states. Although it relies on built-in knowledge of specific conceptual metaphors, it is flexible in allowing novel discourse manifestations of those metaphors. The metaphorical reasoning and mental-state reasoning facilities are fully integrated into a general framework for uncertain reasoning. A special result of the overall approach is that it enables a unified handling of certain apparently separate discourse phenomena: chained metaphor, personification metaphor, and reports of agents' own metaphorical thoughts.

Prelinguistic infants must find a way to isolate meaningful chunks from the continuous streams of speech that they hear. This bootstrapping problem has recently been the focus of several attempts to model the cognitive problem computationally. How can we evaluate whether this kind of simulation is relevant to the cognitive situation, and how can we compare different computational approaches? I discuss my O-B algorithm, a variable-length clustering procedure, and compare it with five other models—three connectionist ones and two statistical programs which use Minimum Description Length as a decision metric. I show that the models differ in their similarity to cognitive processes with respect to: a) the timing of inputs and outputs; b) constraints on the incremental learning process; c) clustering vs. dividing strategy; and d) whether the goal is to find words or to learn word-finding rules.

The relationship of different learning environments (traditional versus constructivist/collaborative) to students' psychological sense of community in the classroom was examined in this study. In addition to students' sense of community, students' social skills and social behavior were also examined. Measures of students' psychological sense of community in the classroom, social problem-solving skills, one's own social behavior, and social behavior of the class were collected. Results from this study suggest that constructivist/collaborative learning environments support students' psychological sense of community in the classroom and social problem-solving skills better than traditional learning environments, and that psychological sense of community in the classroom is an important factor in students' social skills and social behavior in the classroom setting.

A thorough account of how grammar is acquired must handle the problem of how learners deal with covert grammatical elements. In particular, there is cross-linguistic evidence that languages contain verbs that are formed by incorporating a silent grammatical element (a head, in GB terms). Assuming this to be a possibility in natural grammar, this paper investigates what type of input would enable a learner to identify a verb with covert head incorporation, and thus to identify a grammar that contains such a verb. I show that such a grammar cannot be learned from input that does not give the locations of empty heads in sentential structure.

Is human domain specific reasoning illogical? The effect of perspective change on reasoning about social contracts is one of the puzzling phenomena known from research on Wason's selection task that seems to corroborate an affirmative answer to this question. Therefore, some authors postulated non-logical cognitive processes specialized for reasoning about social contracts. In contrast to this view, we argue that such effects reflect the influence of domain specific knowledge on logical reasoning. This knowledge must not be ignored when checking the deductive validity of subjects' inferences. Taking it into account sheds a new light on individuals' deductive competence. Further, it becomes possible to predict such effects not only for the domain of social contracts. We present a model of causal reasoning that allows us to derive new effects of perspective change. W e argue that these effects do not show that people make illogical inferences but, on the contrary, that subjects validly reason deductively from their causal knowledge. Finally, we present empirical results that strongly support our arguments.

Adaptivity is examined within a complex task environment: the Kanfer-Ackerman Air Traffic Controller Task. A computational model is developed in ACT-R to account for such adaptivity using an implicit learning mechanism.

Cognitive Science is founded on notions of representation, and shifts in models of representation have constituted the major internal revolutions in the field. Symbol System and related conceptions were long dominant, but the frontiers passed first to connectionism and more recently to autonomous agent orientations. In spite of its foundational role, representation has never received a consensual or adequate characterization within cognitive science. This is not surprising, given that millennia of effort in philosophy have also failed to achieve consensus or adequacy, but the situation nevertheless constitutes something of a scandal or impasse in a field in which representation is so central. More recently, workers in dynamicist and autonomous agent approaches have argued that representation is not even a useful notion. I argue that this confusion and impasse conceming representation is due to a fundamental misconception about the nature of representation, and offer an alternative model.

The question of how and whether language influences thought is an important one in many of the cognitive sciences. Our work integrates linguistic analysis on lexical semantics with psychological work on memory. It is motivated by neoWhorfian question whether differences in language use will produce corresponding differences in nonlingusitic cognition. The research reported here asks how memory for familiar, unambiguous, "verb-sized" events presented on video might be influenced by an accompanying verb. Our verb choices of Path versus Manner Verbs were guided by cross-linguistic variation in which aspects of an event are highlighted by the verb. W e find a predicted interaction: the verb altered recognition memory of familiar, unambiguous events.

From a cognitive perspective, intentional communication may be viewed as an agent's activity overtly aimed at modifying a partner's mental states. According to standard Gricean definitions, this requires each party to be able to ascribe mental states to the other, i.e.. to entertain a so-called theory of mind. According to the relevant experimental literature, however, such capability does not appear before the third or fourth birthday; it would follow that children under that age should not be viewed as communicating agents. In order to solve the resulting dilemma, we propose that certain specific components of an agent's cognitive architecture (namely, a peculiar version of sharedness and communicative intention), are necessary and sufficient to explain infant communication in a mentalist framework. We also argue that these components are innate in the human species.

Constant interaction with a dynamic environment — from riding a bicycle to segmenting speech — makes sensitivity to the sequential structure of the world a crucial dimension of the cognitive system. Accounts of sequence learning vary widely, with some authors arguing that parsing and segmentation processes are central, and others defending the notion that sequence learning involves mere memorization. In this paper, we argue that sequence knowledge is essentially statistical in nature and that sequence learning involves simple associative prediction mechanisms. We focus on a choice reaction situation introduced by Lee (1997), in which participants were exposed to material that follows an extremely simple rule, namely that stimuli are selected randomly but never appear more than once in a legal sequence. Perhaps surprisingly, people can learn this rule very well. Or do they? We offer a conceptual replication of the original finding, but a very different interpretation of the results, as well as simulation work that makes it clear how highly abstract dimensions of the stimulus material can in fact be learned based on elementary associative mechanisms.

According to the received view in the psychology of reasoning, Piaget's view that: F Humans naturally develop a context-free deductive reasoning scheme at the level of elementary first-order logic. has been overthrown by the poor performance of educated adult subjects on specific logic problems (e.g., Wason's selection task). I propose that Piaget's F (or at least a variant) is alive and well, because the subjects in question are simply victims of a defective education. With a modicum of the right sort of logic training, humans reason deductively on logic problems well enough to vindicate Piaget.

This study deals with the interpretation of verbal commands for action. After an experimental study of human interpretation of instructions for drawing geometrical figures, we have devised a model whose computerized version is called SIROCO. This model represents an attempt to simulate differents mechanisms implied in interpretation of verbal commands. These mechanisms exploiting contextual informations allow clarifying and completing propositions expressed in natural language. In the model, first, the constraints expressed in present command, in environnement (already present figures), and in background communication, are represented with fuzzy subsets and circumstantial semantic networks (well suited for flexible and dynamical representations). Subsequently, an optimization procedure integrating all this constraints allows finding a relevant response to the command. Finally a simulation which consists in translating instructor's verbal commands in a defined minimal language and making it interpreted by the system shows quite good results for the model.

A series of simulations using the HAL model of memory demonstrates that word representations from the model can be used to categorize a variety of common and famous proper nouns, cities, and states. The internal semantics of famous proper names provides a richer set of meaning constraints than do the neighborhoods of common proper names. Retrieval errors with names may be due to this difference in the neighbors and the density of these neighborhoods. A very salient constraint in common proper name semantics is gender.

Competition can be seen as Adversarial Problem Solving (APS), thus ideas from problem solving research can be applied to it. We tested if better modeling of the opponent led to better performance in APS using a zero-sum game played by pairs, but with no obvious skill component. We replicated earlier results that showed that third-order modeling (i.e., what I think my opponent thinks of me: R3MA), but not second-order model (i.e., what I think about my opponent: R2MA) correlated with performance. We also manipulated who was played (same person as in an earlier game, or a predetermined sequence) and who players were told their opponent was (same or different). Players performed better when they could apply the appropriate model (i.e., what they were told matched the opponent). Therefore, we showed that more accurate modeling of an opponent can lead to better APS. However, the critical aspect of modeling may be third-order modeling accuracy. We also found support for a game theory analysis of the task.

It is generally accepted that the process of forming a new category is biased by the learner's prior knowledge. In this context, numerous studies and models have paid attention to the effects of prior domain theories on the process of forming new categories. What is yet to be understood is how this process of acquiring new knowledge might be affected by background knowledge of the very same type, i.e. by prior categories. This paper presents a few experiments showing how the formation of new categories might be facilitated by a high overlap between the new and the old categories, where overlap is operationalized as the mutual entropy between the two.

This paper investigates the relationship readers with different levels of prior knowledge construct among procedural text elements, specifically, among the goal, the actions and the outcome of a procedural text. Readers were either beginners, intermediates, or experts in using a particular software. Our hypothesis was that the main difference between the prior knowledge organization of beginner, intermediate, and advanced subjects was due to the relationship among a goal, the necessary actions to attain this goal, and the obtained outcome. An experiment using a primed recognition task with the goal as prime and both the outcome and the actions as targets confirmed this hypothesis. The primed recognition results were simulated with the Construction-Integration model of comprehension (Kintsch, 1998).

The well-attested friction between linguistics and psychology is not a superficial phenomenon. No conception of language has had more influence on psychology and Cognitive Science than the linguistics of Noam Chomsky. Yet Chomskyan linguistics is radically incompatible with viable accounts of knowledge, and of the development or evolution of knowledge. This incompatibility is strongly manifested in two characteristic Chomskyan doctrines: linguistic competence and the autonomy of syntax. The fallacious arguments on which Chomsky relies are analyzed, and their deep implications for Cognitive Science are traced.

Two views of problem solving procedure generalization are compared in an experiment: the Generalization by Applying Examples (GenApp) and Generalization by Studying Examples (GenStudy) views. The results suggest that learners can acquire a sufficiently general approach for solving novel problems by studying appropriately-designed examples that encourage one to form subgoals to represent a solution procedure. Learners who are led to form a more rote procedure show much less transfer. No evidence was found for generalization through application.

Two eyetracking experiments were conducted to investigate how the domain of interpretation for referential expressions is constructed and coordinated during utterance processing. Of particular interest was how the uniqueness requirement for a definite noun phrase (e.g. 'the book') could be satisfied given a particular array of candidate referents. Experiment 1 demonstrated that the conceptual relation expressed by a preposition immediately limits attention to compatible referents and in turn facilitates definite reference to these objects. Experiment 2 showed that domains are further constrained by the judgment of which referents are compatible with an intended action, and that uniqueness can be established by such factors even when several objects compatible with the noun phrase are present in the perceptual field. The results demonstrate that domains of interpretation for referential expressions are constructed and updated dynamically as an utterance unfolds in time, taking both linguistic and nonlinguistic factors into account.

This paper describes results of a dynamic model of aspectual composition that demonstrates how features necessary for planning and controlling actions can also motivate and ground simple analyses of a number of aspectual phenomena. A novel feature of the model is an active computational representation for verb semantics called x-schemas, an extension of the Petri net formalism that can encode goals, resources and other features affecting aspect. Vexing problems of aspectual composition lend themselves to simple analyses in terms of the context-sensitive interaction between verb-specific x-schemas and a controller x-schema that captures important regularities in the evolution of events. The resulting x-schemas can be elaborated and constrained by such factors as tense, temporal modifiers, nominals and pragmatic context, providing a rich representation that supports simulative inference in language understanding.

Law Encoding Diagrams (LEDs) are classes of representations for problem solving and learning in science. A framework consisting of four schemas has been proposed to account for problem solving and learning with LEDs. This paper assesses the utility of this framework by using it to analyse verbal and behavioural protocols of a subject involved in problem solving with a class of LEDs for electricity.

In this paper we investigate representational and methodological issues in a attractor network model of the mapping from orthography to semantics based on [Plaut, 1995]. We find that, contrary to psycholinguistic studies, the response time to concrete words (represented by more 1 bits in the output pattern) is slower than for abstract words. This model also predicts that response times to words in a dense semantic neighborhood will be faster than words which have few semantically similar neighbors in the language. This is conceptually consistent with the neighborhood effect seen in the mapping from orthography to phonology [Seidenberg & McClelland, 1989, Plaut etal., 1996] in that patterns with many neighbors are faster in both pathways, but since there is no regularity in the random mapping used here, it is clear that the cause of this effect is different than that of previous experiments. We also report a rather distressing finding. Reaction time in this model is measured by the time it takes the network to settle after being presented with a new input. When the criterion used to determine when the network is "settled" is changed to include testing of the hidden units, each of the results reported above change the direction of effect - abstract words are now slower, as are words in dense semantic neighborhoods. Since there are independent reasons to exclude hidden units from the stopping criterion, and this is what is done in common practice, we believe this phenomenon to be of interest mostly to neural network practitioners. However, it does provide some insight into the interaction between the hidden and output units during settling.

Language plays a pervasive role in our day-to-day experience and is likely to have an effect on other non-linguistic aspects of life. At the same time, language is itself constrained by the world. In this paper we study this interaction using Playpen, a connectionist model of the acquisition of word meaning. We argue that the interaction between linguistic and non-linguistic categories depends on the pattern of correlations in the world and on their relation to the correlations defined by words. We then discuss three kinds of possible interactions and present simulations of each using Playpen, a neural-network model of the acquisition of word meaning.

Visual dominance refers to the tendency for visual stimuli to dominate awareness of stimuli of similar or lesser intensity presented simultaneously in other modalities. The effect may be seen in simple and choice reaction time studies. When visual and auditory stimuli are presented separately, visual reaction time is typically slower than auditory reaction time. However, when visual and auditory stimuli are presented simultaneously the visual stimulus generally provokes the first response. In this paper we provide a computational account of such visual dominance effects. The simulation extends an existing computational model of routine action selection, accounting for the counterintuitive visual dominance findings whilst providing further support for the original action selection model.

Here, the fundamental question is the existence of a mental lexicon for pictures, as it exists one for words. And so, what is the structure of the semantic memory and its relationships with access modalities? In the first experiment we determine if a picture or a word can activate the same semantic structure by using the semantic priming paradigm. The classical effect of facilitation was found for word priming, but not for picture as prime. An hypothesis is that the semantic network activated by pictures is different from that activated by words. In the second experiment, we demonstrate that the semantic association structure for pictures and for words is different. So on the basis of these data, we replayed the experiment I. In this case, pictures as words could induce semantic priming of a word. And so we discuss the issue of multiple vs unitary semantic storage.

One large experiment is reported which examined the role of geometry and functional relations on the comprehension of the spatial prepositions over, under, above and below. The task consisted of rating how appropriate a sentence (containing one of these repositions) was to describe a picture. The results show a significant effect of functional relations on the ratings given, demonstrating the importance of functional relations as a determinant of the comprehension of spatial prepositions. However, while over and under were very sensitive to functional relations, above and below were more influenced by geometric relations. Thus these results indicate for the first time that spatial prepositions are differentially influenced by geometric and functional relations, and that geometry and functional relations are distinct factors.

Emotions are a special class of Intentional states with structural components and properties similar to those of the traditional somatic appetites of thirst, hunger and sex. These were originally part of a hardwired, phylogenetically adapted, nonverbal information system for implicitly conveying information about these states both among and within individual members of the species. A classification system provides two major functional classes of emotions, (1) those serving as Appetitive Wishes toward objects, and (2) those serving as Beliefs about the status of fulfillment of those and other significant wishes. Thus, emotions such as Anger and Fear indicate a wish to attack or escape from some object or situation, while Love and Surprise indicate wishes to care about or explore an object or situation. Emotional wishes, like their somatic brethren, require Consummatory Acts for their fulfillment. The result of these acts are emotions such Anxiety and Depression, which indicate Beliefs that the relevant wishes will be hard or impossible to satisfy, or Contentment and Elation, which function as Beliefs that the wishes have been or are being fulfilled. Together, emotional wishes and beliefs form a comprehensive wish-belief information feedback system with manifold causal consequences.

A previous experiment tested subjects' new/old judgments of previously-studied faces, distractors, and morphs between pairs of studied parents. We examine the extent to which models based on principal component analysis (eigenfaces) can predict human recognition of studied faces and false alarms to the distractors and morphs. We also compare eigenface models to the predictions of previous models based on the positions of faces in a multidimensional "face space" derived from a multidimensional scaling (MDS) of human similarity ratings. We find that the error in reconstructing a test face from its position in an "eigenface space" provides a good overall prediction of human familiarity ratings. However, the model has difficulty accounting for the fact that humans false alarm to morphs with similar parents more frequently than they false alarm to morphs with dissimilar parents. We ascribe this to the limitations of the simple reconstruction error-based model. We then outline preliminary work to improve the fine-grained fit within the eigenface-based modeling framework, and discuss the results' implicadons for exemplar- and face space-based models of face processing.

Analogy has traditionally been defined in terms of a contrast definition: analogies represent connections between things which are distinct from the 'normal' connections determined by our 'ordinary' concepts and categories. In this paper we present empirical evidence which, when added to other findings, supports our argument that in the light of current knowledge, the distinction between the two is based more on folk-psychology than on empirically based theory. Research into analogy is however, distinct from research into categorisation when it comes to the richness of its process models. A number of detailed, plausible models of the analogical process exist (Forbus, Centner and Law, 1995; Holyoak and Thagard. 1995): the same cannot be said of categorisation. On the other hand, these analogical process models make a number of explicit and implicit assumptions regarding an 'extemal' categorical process. Whilst treating these processes as separate has been useful in constraining the scope of cognitive investigations, we argue that it ultimately confuses the relationship between analogy and categorisation and is hampering the progress towards further understanding of both.

We present an unsupervised learning methodology derived from compact data encoding and demonstrate how to construct models of polysemy, priming, semantic disambiguation and learning using this theoretical basis. The model is capable of simulating human-like performance on artificial grammar learning.

This article is about nooplasis. That is, the article outlines a general model about the dynamic organization and development of mind. This is done in terms of a number of postulates concerned with the architecture of mind, its development and dynamics, and the nature of learning. Specifically, the model postulates that the mind involves systems oriented to the understanding of the environment and of itself, in addition to general processing functions. It is also postulated that the development of each of the systems is partially autonomous and partially constrained by the development of the other systems, and that it involves both system-specific and system-wide mechanisms of development and learning. Finally, it is argued that these postulates suggest a model of constrained constructivism which differs considerably from what is suggested by the Piagetian or the Vygotskian conception of constructivism.

The Process Dissociation Procedure as applied to episodic recognition requires subjects to study two lists and then determine which of the words in a test list appeared in the second list (Exclusion condition) or on either list (Inclusion condition). We demonstrate that the dual processing account of episodic recognition (Jacoby 1991) does not account for the effects of manipulating the amount of time between the study lists. In contrast, the Bind Cue Decide Model of Episodic Memory (BCDMEM) is fit to the list separation data.

The Operator Model Architecture (OMAR) provides a computational framework in which to develop human performance models that generate reasonable multiple-task behaviors. An interdisciplinary foundation that reached beyond the experimental psychology and artificial intelligence literatures was considered essential to the construction of successful models. Brain imaging and clinical studies suggest that tasks are accomplished through the coordinated execution of function-specific perceptual, cognitive and motor capabilities. These studies together with philosophically grounded cautions, further suggest that the mediation of task contention be accomplished in a framework that does not require an executive that manages task execution. The computational framework for building models sensitive to these considerations is described. Examples from a commercial air traffic control domain are used to illustrate OMAR modeling capabilities.

This study analyzed how a working group used an analogy of "education as a pathway" as a tool for conceptual understanding and teamwork. The team's definition of the analogical structure remained consistent over the course of team work, but they applied the analogy to different targets. Ambiguities in the analogy contributed to its limited application as a tool in detailed analysis. Implications are noted with respect to how analogies are evaluated in team work, how the nature of a metaphor influences its applicability in analytical work, and how group process may affect analogy use in team work.

This study discusses belief-change as the problem of deciding which previously-accepted belief, or premise, to abandon, when an inference from an initial belief set is subsequently contradicted. The data concern how "disbelieving" a previously-accepted conditional premise is realized as a particular modification to that premise. The types of revisions that are made are influenced by the kind of knowledge expressed in the conditional. The results and the broader issues of belief-revision are related to other concerns that have emerged in the literature on propositional inference, such as the reported reluctance of people to make simple valid modus ponens inferences in some circumstances and the general interest in incorporating subjective belief into accounts of deductive inference.

This paper provides empirical evidence that multimodal signals are produced and understood as integrated units of communication called composite signals, rather than being independently interpretable "channels" of communication. I propose that using composite signals relies on two communicative norms, co-expressivity and consistency: - co-expressivity: each element of a composite signal refers to the same underlying referent • consistency: elements of the same composite do not contradict each other. This paper will show that these norms are consistent with data comprising a set of explanations of how locks work in which participants spoke while gesturing, drawing diagrams, and manipulating a sample lock. Co-expressivity is supported by the fact that co-expressive speech segments can be found in nearby speech for communicative nonverbal behaviors but not for non-communicative nonverbal behaviors. Consistency is evidenced in inferences that maintain number and modality consistency in cases of apparent contradiction.

The effects of repeating a task-irrelevant element and inserting intervening trials between the last prime and the probe trial in a negative priming study were compared with a standard prime/probe pair. An associative model based on SAC (e.g. Reder & Schunn, 1996; Schunn, Reder, Nhouyvanisvong, Richards, & Stroffolino, 1997) was able to account for both the decrease in response times across the repeated primes and the increase in response times when the task-irrelevant element became relevant.

We examine how context affects the accessibility of features of combined concepts. A 'contrast hypothesis' suggests that contrasting a to-be-verified feature in the context hinders its later verification. Results of Experiment 1 instead support a priming hypothesis whereby features are differentially activated by contexts. Experiment 2 demonstrates that this priming effect is positive rather than negative, even when feature verification follows a contextual combined concept that is inconsistent with the to-be-verified feature. We conclude that context can differentially activate features of combined concepts, and that it may do so by way of semantic priming.

We report the results of two experiments designed to investigate the role of inference in Wason's selection task. In Experiment 1 participants received either a standard one rule problem or a task which contained an additional rule. This additional rule specified an alternative antecedent. Both groups of participants were asked to select those cards they considered necessary to test whether the rule common to both problems was true or false. The results showed a significant suppression of "q" card selections. In addition there was weak evidence for increased "not-q" selection. In Experiment 2 we manipulated number of rules, as before, and the presence or absence of explicit negation on the cards. Once again "q" card selections were suppressed, but there was no evidence of an increase in "not-q" selection. There was also no effect of type of negation. Our results suggest that inferences about the unseen side of the cards underlie participants' selections. We argue firstly that these findings are inconsistent with current views of selection task performance (Oaksford and Chater, 1994, Evans and Over, 1996) and secondly, that they support accounts which emphasise the role of inference in the task.

Previous researchers on the semantics of spatial relational terms have reported the importance of geometric factors (e.g., Bennett. 1975; Talmy, 1983), the importance of functional factors {e.g., Coventry, Carmichael, and Garrod, 1994; Vandeloise, 1991), and the lack of importance of the nature of the Figure, or object located (e.g., Landau and Stecker, 1990; Talmy, 1983). In this paper, we present the results of an experiment testing each of these claims for the English spatial prepositions IN and ON. Our findings confirm that geometric and functional factors are indeed important. In addition, our results suggest that the nature of the Figure contributes to the selection of spatial prepositions.

Children face an enormously difficult task in learning their native language. It is widely believed that they do not receive or make little use of negative evidence (Marcus, 1993), and yet it has been proven that many classes of languages less powerful than natural languages cannot be learned in the absence of negative evidence (Gold, 1964). In this paper we present an approach to learning good approximations to members of one such class of languages, the regular languages, based on positive evidence alone.

We present a GOMS-MHP style model-based approach to the problem of predicting human habit capture errors. Habit captures occur when the model fails to allocate limited cognitive resources to retrieve task-relevant information from memory. Lacking the unretrieved information, decision mechanisms act in accordance with implicit default assumptions, resulting in error when relied upon assumptions prove incorrect. The model helps interface designers identify situations in which such failures are especially likely.

We used the Quinlan's C4.5 machine learning algorithm to analyze tutorial dialogues as part of the derivation of planning rules for CIRCSIM-Tutor v. 3, a natural-language based intelligent tutoring system. We annotated a corpus of tutoring dialogues with an SGML-based representation of tutorial goals in order to make mechanical processing possible. We looked for rules of the form "under what conditions is goal x implemented with plan y?". We discovered rules for high-level plaiming of the tutoring session and dynamic modification of the tutorial agenda. At a lower level of planning, we looked at rules for generating sections of the tutor's utterance. The use of the rule induction algorithm has helped us discover which knowledge available to the planner is significant in making these decisions, as well as producing some decision trees we can actually use in CIRCSIM-Tutor.

This paper draws on previous research that strongly suggests that bilingual memory is organized as a single distributed lexicon rather than as two separately accessible lexicons corresponding to each language. Interactive-activation models provide an effective means of modeling many of the cross-language priming and interference effects that have been observed. However, one difficulty with these models is that they do not provide a plausible way of actually acquiring such an organization. This paper shows that a simple recurrent connectionist network (SRN) (Ehnan, 1990) might provide an insight into this problem. An SRN is first trained on two micro-languages and the hidden-unit representations corresponding to those languages are studied. A cluster analysis of these highly distributed, overlapping representations shows that they accurately reflect the overall separation of the two languages, as well as the word categories in each language. In addition, random and extensive lesioning of the SRN hidden layer is shown, in general, to have little effect on this organization. This is in general agreement with the observation that most bilinguals who suffer brain damage do not lose their ability to distinguish their two languages. On the other hand, an example is given where the removal of a single node does radically disrupts this internal representational organization, similar to rare clinical cases of bilingual language mixing and bilingual aphasia following bram trauma. The issue of scaling-up is also discussed briefly.

Category-specific semantic deficits refers to the inability to name objects from a particular category while the naming of words outside that category is relatively unimpaired. We suggest that such semantic deficits arise from the random lesioning of a unified semantic network in which internal category representations reflect the variability of the categories themselves. This is demonstrated by lesioning networks that have learned to categorise butterfiies and chairs. The model shows category-specific semantic deficits of the narrower (butterfly) category with the occasional reverse semantic deficits of relatively impaired chair category.

This paper examines the naive Bayesian model and extensions of it to account for the effects of base rate neglect and inverse base rates. These are human categorization phenomena in which base rate information appears to be ignored. The naive Bayesian classifier accounts for a subset of the phenomena observed in base rate experiments. An extension to the model is examined that uses structure in the data sets resulting from features shared between categories.

Deduction is a high level cognitive ability which has not been much analyzed in neuropsychology. Cognitive psychologists and cognitive scientists strongly debate the nature of the mental processes involved in deductive reasoning. A theory particularly pertinent to the neuropsychology of thinking is Mental Model Theory, which postulates the use of analogical representations in reasoning. Studies on unilateral neglect in neuropsychology show that the right hemisphere is involved in analogical representations. On these theoretical bases we make a critical prediction about the role of the right hemisphere in reasoning. This paper investigates the ability of right-brain damaged patients to deal with two main sorts of deduction: syllogistic and relational reasoning. Our results suggest a significant involvement of the right hemisphere in reasoning. Also, as far as syllogistic reasoning is concerned, the results allow for the existence of a verbal component, beside the analogical one.

In this paper we present a context-based family of formal systems (called MultiContext Systems) which we propose to use as a formal framework for a theory of mental representation. We start with an intuitive notion of context as a subset of the complete (cognitive) state of an individual. Then we introduce two general principles which we believe are at the core of any logic of context, namely the principles of locality and compatibility. We show how these principles can be formalized in the framework of MultiContext Systems, and argue that this conceptual/logical framework can be used to account for a variety of phenomena in a theory of mental representation. Finally, we compare our framework with previous work.

It is a classic result in cognitive science that chess masters can recall briefly presented positions better than weaker players when these positions are meaningful, but that their superiority disappears with random positions. However, Gobet and Simon (1996a) have recently shown that there is a skill effect with random chess positions as well. The impact of this result for theories of expert memory is discussed. CHREST, a computational, chunking model of chess expertise based on EPAM (Feigenbaum & Simon, 1984) accounts for this skill difference. The model is also compared with human data from an experiment where the role of presentation time for random positions was systematically varied from 1 second to 60 seconds. Simulations show that the model captures the main features of the human data, thus adding support to the EPAM theory. They also corroborate earlier estimates that visual short-term memory may contain three or four chunks.

A new explanation is offered for the phenomenon of dreaming, based on the findings about brain activits during sleep reported in McClelland. McNaughton, and O'Reilly (1995). Many of the phenomena that make dreams seem so strange to us are explained as a byproduct of the process of storing temporary memones into permanent memory during sleep, as it occurs in the connectionist networks of the brain. This explanation provides physiological support for for Malcolm's (1962) criticism of Dement and KIcitman's (1957) interpretations of their findings about the correlation between REM sleep and dreaming, suggesting that the sense of having had a dream is an artifact of being awakened during the process of memory storage.

While a hindrance to statistical and computational models of inference, missing knowledge can be exploited by organisms in their natural environments. The recognition heuristic utilizes missing knowledge to make accurate inferences about the real world. A consequence of applying this heuristic is a counterintuitive less-is-more effect where less knowledge is better than more for inferential accuracy. Theoretical arguments and experimental evidence supporting the less-is-more effect are given.

Contemporary dynamic theories of cognition and functional theories of linguistics fall into two general camps: "traditional" and "emergent" approaches. Building on work of the linguist Paul Hopper, I identify four characteristics of emergent phenomena: feedback properties; sociohistorical embeddedness; language and language-like "structures"; and what I call "recursivity," the feedback-based presence of system-analytic elements within the cognitive systems they seem to explain. This latter feature, especially, raises questions about whether "emergence" is a phenomenon, a theory, an approach, etc. I suggest that emergence offers at least a refreshingly ordinary framework for theories of empirical cognition, which nevertheless flow to the "deep" levels claimed by rule-based cognitive explanations.

Our goal is to develop a cognitive model of how humans acquire skills on complex, sensorimotor tasks. To achieve this goai, we collected data from subjects learning the NRL Navigation task, then used the data to construct a model that reflects the bcisic, cognitive elements required to learn and thereby succeed at this task (Gordon & Subramzuiian, 1997). This paper describes a new experiment with humcin subjects on the task. Data from this experiment not only confirms the key cognitive element of our model, but also helps us better understand individual differences in learning this task. Four evaluation metrics indicate that we are able to model important trends in the evolution of action choice.

The correlational structure of judgments of grammatical coreference is examined using factor analysis and the results are used to identify the dimensions of grammatical variation in competent speakers of English. The dimensions that are discovered do not correspond to those typically discussed in generative linguistics but they can be explained very naturally by a model in which coreference is achieved through a process in which linguistic expressions are mapped onto a model of discourse.

Causality plays an important role in giving discourse its characteristic coherence. This paper examines how causality implicit in an utterance helps to organize dynamically constructed mental models of discourse. Experiments are reported suggesting that the linguistic form of utterances contributes significant semantic information about causality to a discourse representation. This view is contrasted with competing claims in the literature that causality only emerges from social psychological inferences or optional inferences on background knowledge.

To produce effective advice several sources of knowledge are needed. Knowledge about the application domain the advice is concerned with is of course necessary, but not sufficient. If the aim of the intervention is inducing people to modify their habits, we also need specific theories of how and why people change behaviour to guide the advising process. In some cases, however, it still does not suffice: when suggesting a change in a well established habit, several factors have to be taken into account, and a good adviser might also need argumentative capabilities, in order to overcome possible personal and environmental barriers to the change. This paper presents a model of advice giving that integrates Artificial Intelligence with concepts and methods coming from different disciplines. The model has been implemented in Daphne, an advice giving system that operates in the nutrition education domain.

We present a view that conceives of conceptual learning as changes in discourse practices. This view focuses on interactions in which people construct understanding collaboratively, either as deliberate conceptual inquiry or to facilitate accomplishing something else. Our analysis combines concepts and methods from ethnography (e.g., Jordan & Henderson, 1995), linguistic discourse analysis (e.g., Lemke, 1990), cognitive analyses of conceptual growth (e.g., Keil, 1994), and theories of information structures in comprehension and reasoning (e.g., Kintsch & van Dijk, 1978). In this view, conceptual understanding is considered mainly as an interactional process. The view focuses on how concepts are created and built up when people engage in activity, especially when they communicate about the things they are doing and trying to understand. Participation in a community includes using its concepts according to practices in which members communicate, coordinate their action, and achieve mutual understanding. Our view of concepts is illustrated with examples drawn from a study of two FCL science classes.

In this paper we explore the idea of reflection in the context of scientific exploration. How does an agent reflect upon its behavior in order to enable productive exploration? We outline an abstract cognitive architecture for combining reflection and exploration. To achieve this we present a language for modeling cognition: the Task-Method-Knowledge (TMK) language. We further present a computational model based on this language, TORQUE2 (Griffith etal., 1997; Griffith, 1997). TORQUE2 is a model of exploratory reasoning in the domain of scientific problem solving. We claim that the TMK language supports both reflection and exploration, and enables them to benefit from one another.

Task analysis is a critical first step in understanding a new complex domain. Currently, tasic analysis is a mostly manual process with weak automation support. This paper introduces the first phase of the SAVVII prototype as a proof-of-concept for early validation of task analysis activities. Early validation is supported by the transference of semantics from data values to data structures. Rough estimations of discrepancies between tasks are used to focus the knowledge elicitor's attention on questionable areas, thereby reducing much of the tediousness and time-intensive nature of validation. SAVVII was shown to work on the developmental domain of parables. It is currently undergoing experimentation in two real-world knowledge acquisition activities.

A Hebbiein-inspired, competitive network is presented which learns to predict the typical semantic features of denoting terms in simple and moderately complex sentences. In addition, the network learns to predict the appearance of syntactically key words, such as prepositions and relative pronouns. Importantly, as a by-product of the network's semantic training, a strong form of syntactic systematicity emerges. Moreover, the network can integrate novel nouns and verbs into its training process. This is achieved by assigning predicted semantic features as a default meaning when a novel word is encountered. All network training is unsupervised with respect to error feedback. Issues addressed here have been the subject of debate by notable psychologists, philosophers, and linguists within the last decade.

The extent to which human cognition can be understood as rule-based is a classic issue in Cognitive Science and one which continues to provoke heated debate in a wide variety of areas, ranging from Implicit Learning through Inflectional Morphology to the acquisition of reading skills. Despite its centrality, the central notion of "rule" is far from well-defined. This paper examines a central feature of rule-based models, the concept of rule-following, and clarifies its role, its content, and some of the typical fallacies associated with its use.

Inflectional morphology has figured prominently not only in debate about the nature of linguistic knowledge, but also in the foundational debate between proponents of symbolic and of connectionist accounts of cognition. We present two experiments designed to test predictions of Tinker's (1991) dual-route account of inflection, the central component of which is a symbolic rule. Contrary to the predictions of the dual-route account, we find evidence of both frequency and similarity effects on the regularization of novel items (i.e., pseudo words).

We describe an experimental investigation of the development of children's knowledge stmctures which aims to provide data for connectionist modelling. 167 children between 5 and 11 years of age completed two category fluency tasks where they were asked to produce as many names of a) animals and b) parts of the body, as they could in one minute. Similarity scores were derived based on distances between concepts in the lists produced. These were analysed using the ADDtree algorithm (Sattath & Tversky, 1977) to build structures representing the organisation of the children's knowledge of animals and body parts. The results showed that animal knowledge was generally organised in terms of environmental context/habitat, however, there was evidence for subtle changes in knowledge organisation between age groups. More pronounced changes were observed in the organisation of knowledge of body parts which gave some support to the assertion that children progress from making coarser to making fmer distinctions between concepts (see Keil, 1979) and reflected the progression observed in knowledge structure development in a connectionist model of semantic memory discussed by McClelland, McNaughton and O'Reilly (1995). Our aim is to extend this work to provide data enabling connectionist modelling of semantic memory within a developmental framework.

Given that the single most important mathematical skill for students who wish to study beyond arithmetic is the ability to take a problem situation (usually stated in words) and formulate a mathematical model (usually an equation), we are working on a cognitive developmental model of this skill to be used in an intelligent tutoring system. We call this skill symbolization. High school students do it poorly and improve slowly. We are using a Difficulty Factors Assessment as an efficient methodology for identifying the critical cognitive factors that distinguish competent from less competent symbolizers. We present a developmental model identifying three major transitions through which a student must pass. Underlying the developmental model are empirical results which suggest, contrary to prior research and common belief, the difficulty in algebra word problem solving is less about the difficulties of comprehending the word problems, and more about the difficulty of speaking in the foreign language of algebra. Many of students' errors are analogous to the errors people make when learning to speak in a new language. While it may be that mathematically algebra symbolization is a generalization of arithmetic, cognitively it is more accurate to say algebra symbolization is the articulation of arithmetic.

A model of the productive sound-spelling mapping in English is described, based on previous work on the analogous problem for reading (Zorzi, Houghton & Butterworth, 1998a, 1998b). It is found that a two-layer network can robustly extract this mapping from a representative corpus of English monosyllabic sound-spelling pairs, but that good performance requires the use of graphemic representations. Performance of the model is discussed for both words and nonwords, direct comparison being made with the spelling of surface dysgraphic MP (Behrmann & Bub, 1992). The model shows appropriate contextual effects on spelling and exactly reproduces many of the subject's spellings. Effects of sound-spelling consistency are examined, and results arising from the interaction of this system with a lexical spelling system are compared with normal subject data.

This paper describes psychometric investigations that have been carried out as a prelude to developing new approaches to learning structures within education, based on connectionist concepts. In Experiment A, the ability of 15 subjects to produce different interpretations of an image fonned from abstract geometric shapes was studied over a 30 minute period of observing the diagram. The rate at which these subject produced ideas was shown to initially decline and then become constant. Experiment B investigated the effect of a strategy that encouraged 16 subjects to defocus their thinking before attempting to find another new interpretation. On returning to the problem, the average time taken to produce another interpretation was significantly reduced. Both sets of results are discussed m terms of connectiomst modelling, the need to broaden one's attention during creative problem-solving and the neural mechanism of 'lateral inhibition'. Further evidence for the potential effectiveness of 'chance' strategies is also referenced in the work, techniques and philosophy of well-known and recognised artists.

Several theorists in perception, attention, and memory have suggested that temporal correlation in neural firing patterns (synchrony) could play an important role in processing and learning. Recent neuropsychological evidence demonstrates the wide spread occurrence of synchrony and its stimulus specific nature. Numerous proofs and simulations have demonstrated the ease with which synchrony develops. However, ease of development could be a problem since synchrony is the mechanism behind abnormal processing in epileptic seizures. Previous modeling ignores the role of spatial propagation along the axon. Comparing simulations with and without propagation for a biologically plausible model of neural oscillations, I show that synchrony is far less liable to occur. Using a grid of fully activated cells, the extent of connectivity, impulse amplitude and duration, and natural frequency variability are examined: synchrony is substantially diminished when propagation is included.

A visuospatial reaction time task was used to gain an online measure of learning as subjects responded manually to strings of stimuli containing embedded transitional probabilities. We hypothesized that items within a stimulus sequence that have low transitional probabilities will be learned more slowly than items that have high transitional probabilities. Subjects were instructed to make button press responses to stimulus strings composed of sequences of lights. Items in the strings were organized into triplets, with a low average transitional probability for the first item in a triplet, and transitional probabilities of 1.0 for the second and third items. Results indicate that learning is poorer for stimulus items with low transitional probabilities than for stimulus items with high transitional probabilities. This work ties together a number of previous investigations of sequence learning, and has implications for how more complicated, hierarchically structured sequential input, such as language, may be learned.

In this paper we examine the behavior of a human-computer system for crisis response. As one instance of crisis management, we describe the task of responding to spills and fires involving hazardous materials. We then describe INCA, an intelligent assistant for planning and scheduling in this domain, and its relation to human users. We focus on INCA's strategy of retrieving a case from a case library, seeding the initial schedule, and then helping the user adapt this seed. We also present three hypotheses about the behavior of this mixed-initiative system and some experiments designed to test them. The results suggest that our approach leads to faster response development than user-generated or automatically-generated schedules but without sacrificing solution quality.

This paper reports a miniature language study conducted to examine the acquisition of an ergative verb system. The study is designed to allow the learner the choice of creating either a natural or unnatural system. The study uses a new approach to teaching miniature languages in which the learner is exposed to the language while playing a computer adventure game. The learner acquires the miniature language by determining its properties while seeing words used in context. After learning a set of transitive and intransitive verbs, each with its own set of subject clitics, the learner is required to create new words with object clitics. The situation is set up in such a way that the learner has three options: 1. Respond randomly, 2. use the subject clitics of intransitive verbs, creating a system typical of ergative languages, or 3. use the subject clitics of transitive verbs, a pattern not found in natural language. It was found that most subjects (93%) did either 2 or 3, demonstrating that they were performing language learning by forming two classes of subject clitics. Most subjects (78%) used the third option, the unnatural one. This result is interpreted as evidence against a modularity driven imiversal grammar view of language learning. Instead it supports a cognitive account in that the unnatural pattern required less cognitive processing.

The cognitive maps that humans compute as representations of the spatial environment they have visited are rarely even close approximations to what was actually experienced. When we experience the environment we seem to see it all so perfectly, yet rarely are we able to reproduce from memory an exact description of the places visited. Yet these vague, muddled descriptions of the places visited are adequate for many spatial reasoning tasks. But how is such an impoverished representation computed from what is initially delivered by one's senses? And what effect does this representation have on the construction of the cognitive map? W e present one method for computing a vague description of each local space visited. It is derived from the initially accurate description needed for the actions the viewer might perform within the local space. We show the effect of this representation on the structure of the cognitive map.

This paper addresses three hypotheses concerning the procedural/declarative distinction: 1) Procedural and declarative knowledge speed-up as separate, but parallel, power curves; 2) Procedural knowledge operates in one direction only—from condition to action—hereas declarative knowledge can be cued by any of its elements; and 3) Declarative knowledge is active—it can result in behavior independent of procedural knowledge. The paper presents a single Act-R model that closely fits the data of two learning and transfer experiments conducted by Rabinowitz and Goldberg (1995). The model provides a good fit to the data, further validating Act-R as a model of the human cognitive architecture. In addition, the model shows that the two experiments cannot be used to argue that declarative knowledge can be retrieved without any intervening procedural knowledge.

Development of powerful brain imaging techniques has revolutionised our knowledge of the patterns of cerebral activation which underlie the performance of cognitive tasks. Particularly striking is the extent to which cognitive performance has been shown to be accompanied by motor processing even in the absence of physical movement, consistent also with considerable behavioral evidence. By definition, left-handed and right-handed people exhibit systematic differences in motor processing. It is thus possible in principle that handedness-dependent differences in patterns of motor activation may exert observable effects upon cognitive performance. New evidence suggests that this is indeed the case. It has been shown that people's handedness can significantly influence the accuracy of what they remember. Cognitive Science thus needs a chiral component. The results of experiments support the hypothesis that handedness effects are linked directly to specific patterns of motor activation, rather than indirectly to general differences in hemispheric processing.

In order to ground our understanding of cognitive development we have started to create a model of how children and adults solve a well-studied three-dimensional puzzle. We started with a model that fits the adult behaviour on the puzzle. We then modified the model's cognitive architecture (ACT-R) and its perceptual/motor architecture (the Nottingham Interaction Architecture) in three ways to simulate a younger problem solver by: (a) reducing the accuracy of vision, (b) reducing working memory, and (c) doing both. The modifications, particularly reduced working memory (and its combination with reduced visual accuracy), allow the model to approximate, on some measures, the behaviour of seven year olds on the puzzle. The results suggest that cognitive models and their architectures can help answer the question of "What develops?".

Children acquire noun inflections before they acquire verb inflections. Noun inflections are dso less affected by language disorders than verb inflections. We describe a single-system connectionist model of English noun and verb inflection which captures these facts of acquisition and atrophy, as well as other well-established developmental characteristics such as U-shaped learning and the ability to generalise to novel forms. The model makes the novel experimental prediction that irregular nouns are less affected by damage than irregular verbs, even though irregular nouns are harder to learn.

Traditionally, regularity in a data set is assessed by fitting a model to the data and examining the extent to which the variance accounted for by the model is large compared to the overall variance in the system. Such approaches, however, do not address the complementary question of how much regularity is present in the data, in the first place, and how much work is expected to be required to capture a particular amount of regularity. In this work we use the notion of Kolmogorov complexity to derive a measure of system regularity independent of any particular model. Thus, in our framework, the explanatory adequacy of a model can be readily quantified, so that one can examine the extent to which the model is satisfactory, or whether additional mechanisms need be postulated.

The conventional interpretation of double dissociations is that they are almost irrefutable evidence of distinctions in both function and type of mental processes, or of separation of cognition into modules. We present a connectionist model that demonstrates apparent double dissociations within a single-route, single-mechanism network and argue that these apparent dissociations are simply the expected tails of a standard bell curve describing network performance. We conclude that within a connectionist model, the appearance of double dissociations may not be evidence for functional or mechanistic separation, and that similar caveats apply to the interpretation of double dissociations in human cognitive behaviour.

We present a computer model for the acquistion of programming languages from textbooks. Starting from a verbal description of the notational conventions that are used to describe the syntactic form of programming commands, a meta grammar is generated that parses concrete command descriptions and builds up grammar rules for that commands. These rules are realized as definite clause grammar rules that captures the syntax of these commands. They can be used to parse and generate syntactically correct examples of a command. However, to solve real programming problems also the semantics of a command and of its parameters needs to be acquired. This is accomplished by the natural language parsing of the explanations given in the text and the augmentation of the definite clause command grammars with semantic structures.

Two experiments on spatial relational inference investigated effects known from relational and syllogistic reasoning. (1) Continuity effect; n-term-series problems with continuous (W r1 X, X r2 Y, Y r3 Z) and semi-continuous (X r2 Y, Y r3 Z, W r1 X) premise order are easier than tasks with discontinuous order (Y r3 Z, W r1 X, X r2 Y). (2) Figurai bias: the order of terms in the premises (X r Y. Y r Z or Y r X, Z r Y) effects the order of terms in the conclusion (X r Z or Z r X). In the first experiment subjects made more errors and took more time to process the premises when in discontinuous order. In the second experiment subjects showed the general preference for the term order Z r X in the generated conclusions, modulated by a "figural bias": subjects used X r Z more often if the premise term order was X r Y , Y r Z, whereas Z r X was used most often for the premise term order Y r X, Z r Y. Results are discussed in the framework of mental model theory with special reference to computational models of spatial relational inference.

People tend to give subadditive probability judgments when asked to assess each in a set of three or more exclusive hypotheses. The degree of subadditivity in such judgments is determined in large part by the evidence upon which the judgments are based, but the characteristics of the evidence that influence subadditivity have yet to be fully specified. In the present experiments, this issue was addressed using a classification learning task, in which the relationship between the evidence and the hypotheses under consideration can be controlled experimentally. Two potential evidential influences on subadditivity--cue conflict and cue frequency--are distinguished and tested in three experiments. The results indicate that (a) people's probability judgments are systematically subadditive--in violation of standard probability theory--even when the judgments are based on cues learned within the experimental context, contrary to the predictions of "ecological" theories of human judgment which attribute such biases to nonrepresentative item selection; and (b) cue conflict has a reliable influence on the degree of subadditivity exhibited in probability judgments.

Differential calculus provides various ways to conceptualize change, any of which can be employed with applied problems. Experts associated with different academic disciplines (chemistry, physics, mathematics) were asked to think out loud while working on a problem requiring a differential equation for its exact solution. These experts used strikingly different representations in solving the problem. Comparisons between their protocols are based on a historical-cognitive approach that ties present-day representational practices of differential calculus to the history and conceptual development of the calculus. Agency, here defined as the task assigned to the problem solver by the representation, is at the heart of this link between past and present practices. Whereas the agency characteristic of the Leibnizian calculus is choice, the agency characteristic of Newtonian calculus is transformation, and that of the modern function-based calculus may, in applied contexts, be characterized as observation and manipulation.

How do my mental states compare to yours? We suggest that, while we may not be able to compare experiences, we can compare neural representations, and that the correct way to compare neural representations is through analysis of the distances between them. In this paper, we present a technique for measuring the similarities between representations at various layers of neural networks. We then use the measure to demonstrate empirically that different artificial neural networks trained by backpropagation on the same categorization task, even with different representational encodings of the input patterns and different numbers of hidden units, reach states in which representations at the hidden units are similar.

Cooper and Shallice (1995) raise many issues regarding the unified theories of cognition research program in general, and Soar in particular. In this paper, we examine one specific criticism of Newell's (1990) treatment of immediate behavior and use it to explain the notion of the modeling idiom within a cognitive architecture. We compare a dual-task model using Newell's architecture and idiom to two other models that use different architectures and idioms (EPIC and an experimental version of Soar). We also look at the models' dependency on their respective cognitive architectures, and the theory/implementation gap also identified by Cooper and Shallice (1995).

There has been a heated debate between connectionist and symbolic models on the task of learning the past tense of English verbs. Claims are often made, but not often justified, that a new model has a superior generalization ability to the previous ones. In this paper, we first set up a proper criterion for making comparisons between models. We point out a crucial issue in comparison which has been largely ignored in the past. Then we present results on the generalization ability of the symbolic pattern associator, SPA. We challenge connectionist researchers to design connectionist models with similar or better generalization ability.

A series of experiments are presented that replicate the mediated priming effect (e.g., lion-stripes) using a naming latency task, and demonstrates that mediated priming does not rely on weak, but direct, semantic relationships or lexical co-occurrence as suggested by McKoon and Ratchff (1992). The magnitude of mediated priming is not negatively correlated with either semantic relatedness or lexical co-occurrence as McKoon and Ratcliff would predict. Furthermore, we show that differences in the contextual nature of the prime - target pairs affects whether or not mediated priming occurs. These findings are discussed in the context of the HAL memory model suggesting a view of "mediated" priming that is more consistent with a distributed representational view.

One important function of language is to name relations. Preschool children performed a simple mapping task with and without hearing spatial prepositions calling attention to key relations. Children at 44 months were successful only if they were in the language condition. By 49 months, children were competent on the task regardless of condition, although there were still benefits of language. These results suggests that relational language can therefore be an important tool for highlighting relational commonalities children may otherwise fail to use.

The increasing lexicalization of syntactic theories poses new difficulties for incremental models of language processing. In this paper, we describe an incremental interpreter that makes use of knowledge on categories to keep the syntactic structure always connected. This, in turn, guarantees a fine-grained syntax-semantics interaction. The paper introduces the general problem of formalizing the notion of incremental interpretation, and analyzes the current approaches in the cognitive literature.

This study is designed to investigate the predictions of a connectionist model of the development of inductive inference (Loose & Mareschal, 1997). We demonstrate that adults sometimes use perceptual as opposed to label information when reasoning about a taxonomically structured domain (biological kinds). Thirty six participants were taught the names of a set of tropical seeds. Participants believed that they were learning about real seeds, however the stimuli were constructed after the predictions of the model. Participants were taught that one seed had a particular non-perceptual property, and that a second did not. The task was to infer whether a third seed would have this property. In some cases, the third seed was given the appearance of one seed type, but the name of another. The results supported the model's prediction that participants would make perceptually based inferences in this condition (N = 32, /=2.18, p<0.05). These results stand in contrast to previous work using this experimental paradigm (e.g. Gelman & Markman, 1986). The results challenge previous interpretations of inference behavior to recognize that the use of perceptual information as a guide depends in part on the perceptual structure of the category in question, and is not simply explained by an appeal to conceptual representation in terms of causal "theory" structures.

Children's spoken word recognition is little understood compared to our knowledge of the adult system. We present here a combined experimental and computational exploration of the development of lexical access. Three accounts of the way children represent lexical form (Full-Specification, Radical Underspecification and Gradual Segmentation) are rejected in favour of one which derives from a connectionist approach. It sheds light on the pattern of results from two experiments investigating the way children, aged 5- to 9-years-old, process regular and irregular variation in the surface form of speech, which suggested, whilst children's lexical representations are functionally underspecified from at least 5-years-oId, they are only beginning to track the viability of regular phonological variation at 9-years-old. The late acquisition of phonological inference is accounted for in a connectionist model in terms of the sparseness of the information relevant to learning this structural relationship in language.

SUSTAIN (Supervised and Unsupervised STratified Adaptive Incremental Network) is a network model of human category learning. SUSTAIN is a three layer model where learning between the first two layers is unsupervised, while learning between the top two layers is supervised. SUSTAIN clusters inputs in an unsupervised fashion until it groups input patterns inappropriately (as signaled by the supervised portion of the network). When such an error occurs, SUSTAIN alters its architecture, recruiting a new unit that is tuned to correctly classify the exception. Units recruited to capture exceptions can evolve into prototypes/attractors/rules in their own right. SUSTAIN's adaptive architecture allows it to master simple classification problems quickly, while still retaining the capacity to learn difficult mappings. SUSTAIN also adjusts its sensitivity to input dimensions during the course of learning, paying more attention to dimensions relevant to the classification task. SUSTAIN successfully fits item and category learning data from Medin, Dewey, and Murphy (1983). SUSTAIN's performance on other data sets is discussed. SUSTAIN is compared with other models of category learning.

We are interested in a functional account of how capacity constrains memory use in natural, ongoing behaviors, and in how visual memory demands can be reduced through the use of what we have called perceptual pointers, or deictic codes. Here, we ask whether, with experience, participants can restructure task representations such that single fixations can point to more and more complex chunks of information. We tracked eye movements as participants copied simple model patterns which were presented with different frequencies. At first, participants made multiple fixations to individual pattern components. As patterns were presented repeatedly, model inspections were reduced substantially. This suggests that participants formed more compact representations of the patterns with experience, allowing single fixations to point to larger chunks of information. We also propose that deictic codes provide a short-term store analogous to the visuo-spatial scratchpad or articulatory loop. When the task was structured such that a separate visual search was required for each model component, much less learning was observed than when fixations to known locations were required, suggesting deictic codes were disrupted by active visual search.

It is well known that the time course of lexical access is shaped by the number and nature of potential competitor items in the lexicon. While research has outlined the macrostructure of lexical processing (e.g., that during spoken word recognition, lexical candidates similar to the input are activated and compete for recognition), many questions remain about the microstructure (how exactly is the competitor set defined?) and dynamics (what is the time course of lexical competition?) of lexical processing, as well as their development as words are learned. Here, we begin to address these issues with a study in which participants learned to recognize words from a lexicon of novel names associated with novel shapes. Each item in the lexicon (e.g., /pibo/) had two potential competitors (e.g., /pibu/ and /dibo/). Half of the words were presented more frequently than the other half during training. This allowed us to examine the development of competition effects with experience. An eye tracker provided an on-line measure of the items being considered for recognition. The results indicate that lexical competition effects among newly-learned items develop quickly.

Perceptual cues clearly play a fundamental role in early categorization. Perceptual properties, however, are typically understood to be static shape cues. Some studies have suggested that dynamic perceptual cues, such as motion, may also be important in categorization. This study was an attempt to explore the role that motion plays in children's categorization of biological kinds as well as in more abstract concepts, such as geometric figures. Confronted with a choice between movement and shape, 4-year-old children were found to base their inductions about category membership primarily on motion cues, regardless of whether the objects were animals or geometric figures. This pattern of responses is also present in 7-year-olds for animals but not for geometric figures. Older children may begin to appreciate that motion is unique to animals and are therefore less likely to use motion cues to categorize geometric figures. The results support the view that children are initially guided by motion in categorization. Only as they grow older do they begin to constrain their inferences with respect to different motion cues. The present findings suggest that motion plays an overriding role that is central in the process of concept acquisition and in the mechanisms by which concepts are later structured.

Philosophers of science have long argued that when evaluating explanations, we do not consider ideas in isolation. Instead, we possess an integrated web of information that comprises the context we consider when weighing evidence about any component of this web. In this paper, we provide empirical evidence that theories are considered in context by demonstrating that non-scientists change the strength of their belief in both of two alternative theories, even when only given information about one of these hypotheses. In addition, we seek to identify and describe some of the types of information people use in evaluating theories. Information about mechanism, inferences that discriminate between two explanations, and information about closely related situations in which the target factor operates as a mechanism can all significantly affect ratings of two rival explanations.

Effective arguments must be presented in a cohesive manner: simple collections of believed premises and connecting inferences supporting a goal may not persuade the recipient if they are not well ordered. We use semantic activation and Bayesian propagation in a user model to simulate the effect of presenting an argument generated by our system, NAG, to the user. This simulation is used to select a strategy for presenting the argument to the user. The simulation also identifies superfluous lines of reasoning that may be removed, and enables NAG to determine how multiple subarguments for points should be presented, e.g., as multiple individual supports or collectively. A greedy algorithm is then used to apply probabilistic pruning and semantic suppression to further simplify the argument. Probabilistic pruning removes unnecessary premises from the argument. Semantic suppression is used to select portions of the argument which are within the user's focus of attention, and which are also readily inferred, and hence can be left implicit without damaging the effectiveness of the argument.

Using an auditory semantic priming paradigm. Moss, Ostrin, Tyler and Marslen-Wiison (1995, Experiment 2) demonstrated facilitation for category coordinates and functionally-related stimuli both with and without the additive effect of normative association strength. In this paper we replicate these results computationally using a corpus-derived Contextual Similarity measure. In Experiment 1 we consider the adequacy of the Contextual Similarity measure in accounting for Moss et al.'s results, and discuss how fiinctional and categorical semantic relations are represented in corpus-based approaches to lexical semantics. We also offer an explanation for how the Contextual Similarity measure succeeds in replicating the additive effect of association strength on semantic priming without postulating a qualitatively different mechanism for associative priming. We then investigate why previous corpus-based approaches (Lund, Burgess & Atchley, 1995) have failed to produce similar results. We argue that this is because vector representations partly encode temporal co-occurrence information. This explanation is tested in Experiment 2.

In an attractor model of semantic memory, semantic similarity is determined by degree of featural overlap. In contrast, in spreading activation theory, two concepts are similar if they share features or if they are linked to the same superordinate category node. We present an attractor network model of computing word meaning and use it to simulate the data of McRae and Boisvert (in press), who found that short SOA semantic similarity priming directly depends on degree of featural overlap. The two accounts of semantic similarity are then contrasted in a human experiment. In support of attractor networks, priming effects were determined by featural overlap, and no evidence was found for priming through a purported superordinate node. It is concluded that lexical concepts are not represented as static nodes in a hierarchical system.

We describe our work on developing a general purpose tutoring system that will allow students to practice their decision-making skills in a number of domains. The tutoring system, B2, supports mixed-initiative natural language interaction. The natural language processing and knowledge representation components aire also general purpose—which leads to a tradeoff between the limitations of superficial processing and syntactic representations and the difficulty of deeper methods and conceptual representations. Our solution is to use a mixed-depth representation, one that encodes syntactic and conceptual information in the same structure. As a result, we can use the same representation framework to produce a detailed representation of requests (which tend to be well-specified) and to produce a partial representation of questions (which tend to require more inference about the context). Moreover, the representations use the same knowledge representation framework that is used to reason about discourse processing and domain information—so that the system can reason with (and about) the utterances, if necessary.

Accident, infection, surgery, or stroke resulting in brain trauma can leave individuals with significant and pervasive cognitive disabilities. The need to increase fiinctional recovery for these individuals challenges the combined knowledge, skills, and vision across disciplines including neuropsychology, rehabilitation psychology, occupational therapy, speech pathology, and computer science. This paper reports such interdisciplinary research to develop an approach to computer-assisted retraining that can support and encourage patients' own efforts to take charge of their lives again and rebuild their cognitive skills and thereby enhance their vocational and social opportunities. The Adaptable Learning Environment for Rehabilitation Training (ALERT) will track user performance levels, interest, preferences, and progress within an environment that uses Virtual Reality for life-skill simulations and activities to functionally model cognitive task domains. A single standardized assessment method is being designed to collect information about cognitive variables in the context of mediating and support variables. The functional developmental model of recovery upon which ALERT is based will use the ongoing assessment as it updates the patient user model within the intelligent tutoring system to guide the suggestions for treatment at each successive stage.

When people read spatial descriptions they construct a mental model. When they attempt to remember the spatial description they may rely on memory for the description itself, memory for the constructed model, and/or memory for the operations used to construct the mental model an episodic construction trace (Payne, 1993). This paper reports an ACT-R simulation of this multiple-representation account of memory for spatial descriptions. The simulation shows that the idea of a remembered construction trace can arise naturally from ACT-R's treatment of goals as declarative memory elements. The simulation captures the most important experimental data in favour of the construction trace hypothesis.

Prior research reports that children up to 3-years-oId map novel adjectives to object properties only in very limited situations (Gelman & Marlcman, 1985; Taylor & Gelman, 1988; Hall, Waxman, & Hurwitz, 1993; Klibanoff & Waxman, 1997; Waxman & Markow, 1997). Yet we know by 24-months children use adjectives. In two experiments we provide 36-month-olds (Experiment 1) and 24-month-olds (Experiment 2) with rich cross-situational and syntactic information. We show that 24- & 36-month-olds learn adjective-to-property mappings when given multiple examples of the mapping, and when object names are used. We claim that previous experiments failed to find robust adjective acquisition because at least one of these sources of information (multiple exemplars) was excluded. We also suggest that children's initial learning about the meanings of adjectives is affected by syntactic properties of the noun phrase in which they appear.

I review the purported opposition between computational and dynamical approaches in cognitive science, i argue that both computational and dynamical notions will be necessary for a full explanatory account of cognition, and give a perspective on how recent research in complex systems can lead to a much needed rapprochement between computational and dynamical styles of explanation.

Individual differences in the abilities and preferences of students have an influence on their responses to information presented in alternative ways. Explanations may appeal to differences in representation or in strategy. This paper reports an experiment that compares the response of students to two computationally similar methods of teaching syllogisms that rely on different external representations of the premiss information. The use of both representations can be broken down into the same stages: translating-in; manipulating; and translating-out. We show that the ease of acquisition and the understanding of the methods relate to a measure of spatial ability and also to preferences for serialist/holist styles of learning. We find that spatial ability and learning style relate to different stages in the two teaching methods, and are therefore complementary contributors to effective learning. In addition, a further test that predicts diverse responses of students to learning the same information from different modalities was used. This is found to relate specifically to stages of translating-in and manipulation of representations. The results of this study support the view that providing a computational account of reasoning and learning requires an acknowledgement of individual differences in the 'starting state' of the individual. These differences can be explored through measures of ability and learning style. This study also supports accounts of problem-solving that distinguish modality and strategy of information processing.

A parallel distributed processing model was used to simulate a central bias in social reasoning, the correspondence bias (Jones & Harris, 1967). This bias is the tendency to overattribute social behavior to the actor's personality and to underestimate the impact of situations. Simulations indicate that the extent of the correspondence bias can be understood as due to differences in the accessibility of explanatory concepts and the strength of causal links between potential explanations and behavior.

When two orthographically siniilar words are briefly and successively displayed, the second word is often difficult to detect or recall, a deficit known as repetition blindness, or RB (Kanwisher, 1987). Two experiments used word-nonword pairs to test predictions of a computational model based on similarity inhibition (Bavelier & Jordan, 1992) vs. predictions of a sublexical model (Harris & Morris, 1996, 1997; Moms & Harris, 1997). One striking finding was of strong RB even for a single repeated letter (cope carn; hot hix). Results generally supported a sublexical model where only the shared letters are affected by RB, and each shared letter can be differentially affected in a probabilistic manner.

There is now very wide psychological evidence that some quantifiers license subsequent reference to subsets of the complement of the set normally open to subsequent reference. This has posed problems for some formal theories of the kinds of reference made possible by quantified sentences. This paper examines the phenomenon, its interpretation, and its limits. A process-model is suggested.

We investigate some aspects of cognition involved in invention, more precisely in the invention of the telephone by Alexander Graham Bell. We propose the use of the Structure-Behavior-Function (SBF) language for the representation of invention knowledge; we claim that because SBF has been shown to support a wide range of reasoning about physical devices, it constitutes a plausible account of how an inventor might represent knowledge of an invention. We further propose the use of the ACT-R architecture for the implementation of this model. ACT-R has been shown to very precisely model a wide range of human cognition. We draw upon the architecture for execution of productions and matching of declarative knowledge through spreading activation. Thus we present a model which combines the well-established cognitive validity of ACT-R with the powerful, specialized model-based reasoning methods facilitated by SBF.

How can we integrate interrelated theories of individual elements of cognition? Computational models of reasoning processes encode an understanding of reasoning. Consequently, a computational modeling language may be ideally suited to the presentation of theories of cognition. By representing theories of a variety of phenomena in a single modeling language, we can potentially explore how these theories might interact. The Task-Method-Knowledge (TMK) modeling language evolves from artificial intelligence research on the subject of multistrategy reasoning. TMK models provide a compositional account of reasoning processes; they describe not only what the elements of a process are, but also how the functional properties of these elements combine to form the functional properties of the process as a whole. This paper explores the composition of theories of cognition within the TMK framework, drawing on some existing theories within cognitive science as examples.

Human language processing relies on many kinds of linguistic knowledge, and is sensitive to their frequency, including lexical frequencies (Tyler, 1984; Salasoo & Pisoni, 1985; Marslen-Wilson, 1990; Zwiteerlood. 1989; Simpson & Burgess, 1985), idiom frequencies (d'Arcais, 1993). phonological neighborhood frequencies (Luce. Pisoni, & Goldfinger, 1990), subcategorization frequencies (Trueswell, Tanenhaus, & Kello, 1993), and thematic role frequencies (Trueswell, Tanenhaus, & Garnsey, 1994; Gamsey, Pearlmutter, Myers, & Lotocky, 1997). But while we know that each of these knowledge sources must be probabilistic, we know very little about exactly how these probabilistic knowledge sources are combined. This paper proposes the use of Bayesian decision trees in modeling the probabilistic, evidential nature of human sentence processing. Our method reifies conditional independence assertions implicit in sign-based linguistic theories and describes interactions among features without requiring additional assumptions about modularity. We show that our Bayesian approach successfully models psycholinguistic results on evidence combination in human lexical, idiomatic, and syntactic/semantic processing.

We have simulated the brain mechanisms involved in semantic processing tasks such as the differentiation of two sequential stimuli based upon recalled semantic features. In this paper we examine the relationship between perceptual inputs, working memory and semantic memory in these tasks. We propose that phase synchronous firing binds features in semantic memory with concepts in working memory, and that a phase comparison mechanism subserves the process of response selection. The model is consistent with the anatomy and physiology of the component brain circuits where known. This research is important because the relationship between working memory and long-term memory is a central component of many theories of cognition.

In this paper, we review the literature on the relation between solving nonspecific goal problems and learning. Research has shown that reduced goal-specificity facilitates learning of rules and principles of the target domain. Researchers have accounted for this effect using a cognitive load theory (Sweller, 1988) and a dual space theory of problem solving (Vollmeyer, Bums, & Holyoak, 1996). Other researchers have shown that learning can be both facilitated by nonspecific as well as specific goals and account for their findings using goal appropriateness theory (Miller, Lehman. & Koedinger, 1997). W e judge each theoretical account by evaluating their consistencies with unified theories of cognition and other empirical data. We note the shortcomings of the each theory and incorporate elements of each to explain all the data.

In this paper, we investigate students' use of the generate-and-test strategy to solve algebra word problems. This strategy involves first choosing an estimate for the answer and then checking whether the estimate satisfies the constraints of the problem. Based on verbal protocol data, we developed a production system model to simulate students' behavior when they apply this informal strategy. The model predicts problem features that should affect the difficulty of the problems. A large-scale experiment tested the predictions of the model. Verbal protocol data provided additional insights into how students use the generate-and-test strategy.

We describe a hybrid agent architecture capable of simulating emotional behaviour. Agents have certain fixed personality traits, which influence emotion levels as well as relationships with other agents. Changes in emotions are modelled through propagation in an emotion network. We extend a standard action representation to include emotional preconditions and effects. Results from a test-bed environment with two sample scenarios show how differences in environmental factors and personality affect the social behaviour and emotions of the agents in a group over time.

Hierarchical systems can adapt by adjusting the strengths of their components in response to environmental feedback. Regimens for propagating adjustments through a hierarchy are either cascading or distributional, depending on whether the sum of the adjustments is variable or fixed. Both types of regimens can be dampened, amplified or sustained, depending on whether nodes higher in the hierarchy are adjusted less, more or with the same amount as lower nodes. We show that a cascading regimen learns most efficiently with amplified propagation, while a distributional regimen learns most efficiently with sustained propagation. Cognitive scientists ought to explore a wider range of propagation regimens.

Three experiments are presented that examine the influence of Emotional Valence and Familiarity of visually presented lexical stimuli on low-level visual processing. The results provide support for the idea that an early process of automatic appraisal acts to preferentially direct attentional resources to Negative or Novel stimuli. The results are discussed with respect to evolutionary considerations.

One of the major questions in child language acquisition research is whether children and adults have the same mental organization for grammar. We consider the case of the acquisition of -ed and -ing. This has standardly been assumed to show that children organize their grammatical knowledge differently from adults. In contrast to models that claim that children's restriction on these morphemes argues for a noncontinuous view of grammar, we show that it is most parsimoniously accounted for by a privative aspect and tense model (Olsen, 1997), independently needed in the adult grammar. In particular, we show i) that one cannot attribute this pattern of development to children's simple modeling of restrictions in the adult data, and ii) that it is not necessary to assume initial hypotheses discontinuous with the adult state, or primitives not found therein. Rather, the data requires a strong innate component that both delimits possible adult grammars and defines early stages. Our model provides an account of why children show these restrictions, how they recover, and what cross-linguistic variation might occur in the emergence of adult competence.

This paper presents simulation results and network analysis of generative Cascade-Correlation (CC) networks which model the child's learning of English personal pronouns. The network analysis revealed that overheard speech is crucial in learning the correct semantic rules not only for first and second person pronouns but also for third person pronouns. In addition, in order to induce the fully correct semantic rules without error-correcting feedback, the networks need to learn all three personal pronouns. Network analysis techniques used in the present study proved to be a powerful tool for understanding of what the networks are actually learning.

The performance of a neural network that categorizes facial expressions is compared with human subjects over a set of experiments using interpolated imagery. The experiments for both the human subjects and neural networks make use of interpolations of facial expressions from the Pictures of Facial Affect Database [Ekman and Friesen, 1976]. The only difference in materials between those used in the human subjects experiments [Young et al., 1997] and our materials are the manner in which the interpolated images are constructed - image-quality morphs versus pixel averages. Nevertheless, the neural network accurately captures the categorical nature of the human responses, showing sharp transitions in labeling of images along the interpolated sequence. Crucially for a demonstration of categorical perception [Hamad, 1987], the model shows the highest discrimination between transition images at the crossover point. The model also captures the shape of the reaction time curves of the human subjects along the sequences. Finally, the network matches human subjects' judgements of which expressions are being mixed in the images. The main failing of the model is that there are intrusions of "neutral" responses in some transitions, which are not seen in the human subjects. We attribute this difference to the difference between the pixel average stimuli and the image quality morph stimuli. These results show that a simple neural network classifier, with no access to the biological constraints that are presumably imposed on the human emotion processor, and whose only access to the surrounding culture is the category labels placed by American subjects on the facial expressions, can nevertheless simulate fairiy well the human responses to emotional expressions.

Three experiments test the processing of compound predictors in contingency judgments. Participants judged the relation between compound predictors and an outcome, as well as the relation between their constituent elements and the outcome, under different predictor-outcome contingencies. In Experiment 1, the contingency of an AB compound predictor was judged as independent of the contingencies of its elements A and B. In Experiment 2, judgments of a compoimd predictor (ABC) remained similarly unaffected by changes in the contingencies of its elements, even though the similarity between the compound predictor and one of its constituent elements (AC) was high. In Experiment 3, compound predictors were perceived as unique, although the rate of acquisition of an A+, AB- discrimination did not differ from that of an AC+ , ABC- discrimination, contrary to the prediction of Pearce's (1994) contlgural model. Overall, the elemental associative view is rejected in favor of a modified, low generalization, configural model.

Advice-giving is improved by understanding and responding to user feedback. Previous models of this task treat user responses as misunderstandings or misconceptions and focus on generating altemate or corrective explanations. By and large, however, these models do not consider the possibility that the system's advice is inappropriate and may need revision during the course of an on-going dialog. This paper presents a model of the process of revising plan-oriented advice in response to user feedback. Our focus is on mechanisms for evaluating planning alternatives, for detemiining when to revise advice, and for dynamically generating explanations of these revisions.

As Simon (1981) has pointed out, coming to characterize phenomena as functional systems is fundamental for our understanding of the natural and man-made worlds. Yet little is known about people's propensities for making such characterizations. In contrast to previous research that has focused on unfamiliar, opaque systems, the study reported here investigated experts and novices relative use of structure-function relationships to understand a familiar, inspectable system--a bicycle. As the study shows, the experts, but not the novices, spontaneously and consistently utilized a systems approach to characterize this familiar object.

We address the question of how artificial systems and natural organisms develop spatial competence. Most artificial systems draw upon considerable sophisticated operator- or developer-originated knowledge about what in the world sensor signals represent. Natural systems do not have such sophisticated auxiliary sources of information. We are interested in how, despite this, they achieve perceptual organisation, and suspect that the methods they use will have generalisable effectiveness. We describe a process that creates coherent mappings between the physical world and the phenomenological realm, analogous to retinotopicity and sensory homuncularity in natural systems, and discuss its application to problems of higher dimensionality and higher levels of abstraction. Importantly, such a process, having proved successful in the perceptual robotics domain of our current interests, is likely to be found in other cognitive domains because its strengths lie in its ability to organise and implicitly summarise data in the absence of clues about what that data represents.

This paper starts out from two observations: firstly, that there are complex links between what we term intelligence and what we term creativity and, secondly, that the phenomenon of surprise has a significant role in both the genesis and evaluation of creativity, and is tightly coupled to perception. We argue that for machines to develop to the point where we attribute to them intelligence and, therefore, their own degree of creativity, they must first develop a sensibility of surprise. This, we show, is predicated upon a multi-level organisation of perception, and a method of representing the interest, or novelty, of events and actions taking place in the physical world. A sensibility of surprise further depends on an ability to recognise the novelty of actions the system itself is contemplating. We describe methods of encoding surprise in perceptual robots, and show how this enables them to focus on what is interesting in their environment - a prerequisite to the production of behaviour both creative and intelligent.

Three important findings of research on expertise in formal sciences such as physics are that novices and experts differ with respect to (1) how they structure their domain knowledge, (2) how they mentally represent problems and (3) how they approach problems. Though first attempts have been made to account for the acquisition of knowledge structures as possessed by experts, the reconstruction of the involved learning mechanisms in a psychologically plausible and instructionally fruitful way still remains a challenge. In an experimental study, we investigated how tenth graders acquire and successively relate qualitative and quantitative problem representations in classical mechanics. Initially, subjects were taught either qualitative or quantitative aspects of classical mechanics. Afterwards, two subjects, who were taught differently, collaborated on problems which were beyond the competence of each of them separately. Before and after the collaboration subjects had to work on multi-component tests. In addition, protocols were taken of the subjects' verbal exchange of information during collaborative problem solving. An analysis of variance of the multi-component tests revealed that the subjects successfully learned to interrelate qualitative and quantitative problem representations. A protocol analysis further indicated that subjects gradually shifted their focus from quantitative problem representations to qualitative problem representations during collaborative problem solving.

We adopt the interpretation of rationality according to which an organism's behavior is rational if it is optimally adapted to its environment (Anderson, 1990, 1991a, 1991b). Rationality, according to this view, often implies mechanisms that are as informationally efficient as possible. We interpret the problem of basic-level categorization (Rosch & Mervis, 1975) as one of data compression within an information theory framework, to define a framework whereby the best classification on a set of items is the one that maximally compresses the description of the similarity structure of these items. This framework is then used to examine whether participants in two experiments classified meaningless items in a way that reflected such a compression bias. In addition to the implications for human basic-level categorization, an objective criterion is established for assessing the relative merits of alternative clustering solutions on the same domain.

Artificial grammar learning (AGL; Reber, 1989) has been a major experimental paradigm for the study of human induction processes. In this work we investigate the extent to which the learning mechanisms involved in AGL are general, an issue important to the ecological validity of AGL research. We have used three kinds of stimuli; Letter strings (the standard in AGL work), city sequences that corresponded to routes of an airline company, and shapes that were presented so that later shapes in a sequence contained all previous ones. We compared overall accuracy and patterns of error in these domains to find that performance was not different. The implications of this finding for existing theories of AGL and proposed relations to other cognitive mechanisms are discussed.

We argue that the key question in conceptual storage is best viewed not as a question of instances versus generalisations, but rather one of unitary versus multiple representation accounts of conceptual storage. On previous evidence, it has been difficult to determine whether a particular result stems from stored information regarding the concept or from the processes that operate in invoking a particular concept (Komatsu, 1992). In this paper, we attempt to shed some light on the nature of stored conceptual structure using the different influences that surface and stmctural features have been shown to have on the recall of a particular representation (Centner, Ratterman and Forbus, 1993). We conclude that at least some concepts may not be stored using a unitary representation.

We argue that common interpretations of Wittgenstein's Philosophical Investigations within Cognitive Science misrepresent his account, underplaying its radical content. Appropriately interpreted, this account continues to challenge contemporary theories of concepts and categorisation. We illustrate the continued relevance of his position by directly applying its critique to current approaches to categorisation.

Lexical and syntactic processes are usually regarded as separate sub-systems of the language processing system. We re-examine the autonomy of these processes, given a mental lexicon that is morphemically decomposed, in 3 self-paced reading experiments. Although inflectional affixes have a syntactic role and derivational affixes have a lexical role, there were similar patterns of processing for both types of affix (Experiments 1 and 3). This suggests that there is a common combinatorial process at both levels of the system. Using novel and established morphologically complex words, we varied word-internal factors together with sentence level constraints (Experiment 2). Both sentence-level constraints and word-internal factors had parallel effects on the processing of novel and established words. Overall, the results indicate that the relationship between lexical and syntactic processing may be non-autonomous when morphological composition is taken into consideration.

How do people represent causally complex situations? A real-world case was used to investigate whether single-cause explanations are preferred, and to assess whether goals facilitate causal discounting. Participants were asked to think about the causes of Princess Diana's death and were assigned the goal to show that either the driver or the photographers were not responsible. Participants drew a causal diagram depicting their theory, and rated the importance of the causal factors mentioned. In general, people did not seek a unique cause for the event and generated multicausal explanations with no explicit links between causes. Those given the goal to defend one party included fewer causal factors related to the defended party and rated them as less important, but did not over-emphasize the importance of other factors. The results differ from those found in typical attribution tasks.

In this paper wc present an approach to modeling emergent syllable systems using simulated evolution of a "vocabulary" of "words." The model is aimed at testing the general hypothesis that language-universal sound patterns emerge from selection pressures exerted on the system by the perceptual and articuatory constraints of language users. The model is able to distinguish between hypotheses about how specific, biologically-motivated constraints affect the sound structure of language. For example, it is shown that mandibular oscillation provides a strong constraint on the sequential organization of phonemes into words. Future work will explore the potential of other constraints that, with mandibular oscillation, will be sufficient to describe the emergence of syllable systems.

The meanings expressed by reduplication, or linguistic doubling, are similar across a wide array of languages. Interestingly, some of these shared meanings do not concern doubling, repetition, or plurality. This non-arbitrariness of the sign may be attributable to the interplay of two forces: iconicity, and conceptually-based semantic extension. Cross-linguistic evidence supporting this account is presented. More generally, this paper argues that the interaction of iconicity and semantic extension constitutes a potentially powerful source of nonarbitrariness in the mapping between sound and meaning.

Does categorization involve more than the similarity of an item to a category prototype or other category members? Rips (1989) argues yes, because categorization and similarity ratings sometimes diverge, indicating that they are based on different factors. However, Smith and Sloman (1994) suggest that such categorization/similarity dissociations may be limited to special conditions. We examined the effect of causal relationships between category attributes on categorization and similarity, and found that causal knowledge had a much larger effect on categorization than on similarity. This result was obtained with stimuli rich in characteristic attributes and without participants thinking aloud, that is, in just those conditions where Smith and Sloman found categorization to be solely similarity-based. Thus, the categorization/similarity dissociation demonstrated by Rips is alive and well, and the need for an account of categorization that goes beyond similarity is again highlighted.

In this paper we review three kinds of proposed solutions to the catastrophic forgetting problem in neural networks. The solutions are based on reducing hidden unit overlap, rehearsal, and pseudorehearsal mechanisms. We compare the methods and identify some underlying similarities. We then briefly note some potential implications of the rehearsal/pseudorehearsal based methods, including their application to sequential learning tasks.

When naming the ink color of color words, adults and children show Stroop interference (Stroop 1935). Cohen, Dunbar and McClelland (1990) produced a connectionist model that accounted for many of the Stroop phenomena within adults. This paper shows how the paradigm can be extended to show the development of the interference in children as they leam to read. We train a network taking into account the amount of reading practice and attentional skills that would befit a young child to give a prediction of the development of the Stroop effect. These predictions are then tested using a picture-naming Stroop study with two groups of 8 year olds of differing reading ability. The results support the model, suggesting children initially show reverse Stroop interference that with practice becomes normal Stroop interference.

How does time pressure affect cognitive behavior when solving problems in an uncertain environment? We found substantial evidence that, under time pressure, decision makers can not apply knowledge-based action, even if that approach is absolutely necessary for solving the problem. The present study aims to explain this phenomenon in terms of the subjective probability of the uncertain events associated with the problem. Our model insists that overestimating the possibility of getting correct answer with rule-based action, affected by time pressure and the attitude of decision makers, leads to the persistence of rule-based action. The experiment's results supported the proposed model.

We consider the process by which the syntactic parameters of human language are set. Previous work has shown that for natural languages there can be no instant "automatic" triggering of parameters because the trigger properties in natural languages are often deep properties, not recognizable without parsing the input sentence. There are parametric algorithms that learn by parsing, but they are inefficient because they do not respect the Parametric Principle, they evaluate millions of grammars, rather than establishing the values of a few dozen parameters. They do so because they cannot tell in advance which input sentences are pertinent to which parameters, and because they have no protection against misleaming due to parametric ambiguity of the input. There is one model that does implement the Parametric Principle. This is the Structural Triggers Learner (STL). For an STL, a parameter value and its trigger are one and the same thing; they are what we call a structural trigger or treelet (a subtree or in the limiting case a single feature). These structural triggers are made available by UG and adopted into the learner's grammar just in case they prove essential for parsing input sentences. This permits efficient recognition of the parameter values entailed by input sentences and allows the learner to avoid errors by discarding ambiguous input. However, the high degree of ambiguity inlierent in natural language impedes learning even for this efTicient system. An STL must wait a long time between unambiguous inputs. As we explain, this problem is particularly acute in the early stages of learning. In this paper we give a computational analysis of the performance of an STL. We then identify an important factor - the parametric expression rate - that holds promise of a solution to this early learning problem.

In using eye movements to develop cognitive models, researchers typically analyze eye movement protocols with aggregate measures and test models with respect to these measures. Because aggregate analyses sometimes conceal informative low-level behavior, protocol analyses comparing model predictions to individual trial protocols are frequently desirable; however, protocol analysis for eye movement data is often tedious and time-consuming. We describe how to automate the protocol analysis of eye movements using hidden Markov models. Working with data from an equation-solving task, we demonstrate two methods of tracing eye movement data—that is, mapping eye movements to the sequential predictions of a cognitive process model. W e evaluated these tracing methods in an experiment where participants were instructed to execute given equation-solving strategies. When coding the experimental protocols in terms of the given strategies, the automated tracing methods performed as well as human expert coders in a fraction of the time.

Two experiments are reported that investigated the influence of linear spatial organization on transitive inference performance. RewardAio-reward relations between overlapping pairs of elements were presented in a context of linear spatial order or random spatial order. Participants in the linear arrangement condition showed evidence for visual reasoning: They systematically mapped spatial relations to conceptual relation and used the spatial relations to make inferences on a reasoning task in a new spatial context. We suggest that linear ordering may be a "good figure", by constituting a parsinuxiious representation for the integration of premises, as well as for the inferencing process. The late emergence of transitive inference in children may be the result of limited cognitive capacity, which -- unless an extemal spatial array is available -- constrains the construction of an internal spatial array.

This paper derives a number of logically necessary principles that govern cognitive functioning, and reviews empirical evidence supporting the validity of these principles. It advances an argument in which mental representations are conceived as indexical signs, in that they are causally related to the objects they represent. This indexicality gives rise to four general principles of cognitive functioning. First, mental activity is strongly influenced by that which is present. Second, mental activity exhibits relative insensitivity to absence. Third, minds exhibit difficulty representing negation, because representing negation entails representing the absence of that which is negated. Fourth, thinking is believing, in that representing a proposition implicitly entails accepting the truth of the proposition.

Recent alternatives to the polygraph-based Guilty Knowledge Test by (Farwell & Donchin, 1991; Seymour, Mosmann, & Seifert 1997) raise important questions about automatic access to knowledge already in memory. Despite subjects' intentions, "guilty" knowledge in memory can be detected because its automatic access interferes with other recognition tasks (Seymour, et al., 1997). To account for this finding, we present a model based on classic models of recognition (e.g. Kintch 1970; Anderson & Bower 1972). We posit that 'recognition' is a dual process involving & familiarity component where recent occurrence is quickly assessed, and a slower source resolution component, where the source of the familiar information is identified. Our model of the Guilty Knowledge Effect can account for patterns of response time and accuracy used to measure access to guilty knowledge (Seymour, et al., 1997). We also explain why strategies used to mask the Guilty Knowledge Effect are likely to fail given constraints on the recognition process, and discuss potentially successful strategies suggested by the model.

We argue that the connectionist modelling of visual word recognition can be made more explicit and more accurate by the incorporation of information concerning the initial projection of the visual field to the visual cortex. We show that this initial projection involves the precise splitting of the visual field into two hemifields, even in the case of the foveal projection: when a single word is fixated, that part of the word to the right of the fixation point initially goes to the left hemisphere and that part to the left initially goes to the right hemisphere. We present a number of reasons why this initial splitting should be assumed to persist into the higher cognitive processing concerned with visual word recognition. We explore three different phenomena - the processing priority given to the exterior letters of words, the optimum and preferred viewing positions for word recognition, and the core phenomena of dyslexia - and show that in each case a model based on a split architecture makes the correct predictions and captures the human data in a parsimonious and natural way. We conclude that anatomical information concerning the initial visual projection can enrich the modelling of visual word recognition.

Edioic responses abound in dialogues, where a speaker reuses a portion of the text uttered by another in a preceding turm, though semantically they contribute little if any new information. The phenomenon has attracted the attention of researchers from diverse academic fields, ranging from sociolinguistics and developmental psychology, to computational linguistics and human-computer interfaces. This study reports an empirical investigation on echoic responses from an informational perspective. Drawing on statistical analyses of instances extracted from corpora of spoken dialogues in Japanese, we show that echoic responses with different timings, lengths, intonations, pitches, and speeds signal different degrees in which the speakers have integrated the repeated information into their prior knowledge. We further consider dialogue-coordination functions enacted by this informational potential of echoic responses, and identify the function of display as distinguished from the functions of acknowledgment and repair-initiation.

The consonance constraint-satisfaction model is applied to Machiavellianism self-concept effects in cognitive dissonance. Networks parameterized for low Machiavellian traits showed the usual dissonance effect, i. e., more attitude change after giving a counterattitudinal speech than after not giving such a speech, whereas networks parameterized for high Machiavellian traits showed the reverse, thus capturing human data. Classical dissonance theory had not accounted for the fact that people with high Machiavellian traits showed less attitude change after giving a counter-attitudinal speech than after not giving such a speech. The model predicts initial dissonance and the course of dissonance reduction in the various experimental conditions. The results underscore the point that cognitive dissonance operates according to the same constraint-satisfaction principles that govern a variety of other psychological phenomena.

Bailenson, Shum. and Uttal (1998) showed that when people are asked to select from potential routes on a map, their decision relied heavily on the initial attractiveness of the routes. Specifically, people preferred routes that were initially long and straight and headed in the general direction of the destination, even if that route was not the optimal (shortest) route. This paper extends this road climbing theory to route choice on maps of college campuses and to actual navigation around a college campus. Both experiments confirm that when given a choice among routes, people often resort to choosing the one that is most initially attractive. The road climbing model provides an explanation for both people's navigational decisions and also the path asymmetries that have been discovered by previous researchers studying route choice.

Current theories concerning the comprehension of noun-noun combinations propose that relational interpretations are the result of tlie modifier filling some role within the head noun (Murphy. 1988; Wisniewski, 1997), whereas property interpretations involve the structural alignment of the head noun and modifier concepts (Wisniewski, 1997). In this paper we argue that structural alignment underlies the formation of both relational and property interpretations of noun-noun combinations. Property interpretations result from the alignment of the modifier with the head noun, whereas relational interpretations result from the alignment of the modifier with a filler in the head noun. Modifiers for noun-noun combinations were chosen based on their similarity to relational fillers in the head noun concept. Results indicated that frequency of instantiation of relational interpretations was positively correlated with the similarity of the modifier to fillers in the head noun. Similarity of modifier to fillers in the head noun predicted frequency of instantiation of relational interpretations to a greater degree than the rated ability of modifiers to fill roles in the head noun concept.

This paper identifies goal handling processes that begin to account for the kind of processes involved in invention. We identify new goal properties and mechanisms for processing goals, as well as means of integrating opportunism, deliberation, and social interaction into goal/plan processes. We focus on enterprise goals, which extend traditional design goals and knowledge goals to address significant enterprises associated with an inventor. Enterprise goals represent "seed" goals of an expert, around which the whole knowledge of an expert gets reorganized and grows more or less opportunistically. Enterprise goals reflect the idiosyncrasy of thematic goals among experts. They constantly increase the sensitivity of individuals for particular events that might contribute to their satisfaction. Our exploration is based on a well-documented example: the invention of the telephone by Alexander Graham Bell. We propose mechanisms to explain: (1) how Bell's early thematic goals gave rise to the new goals to invent the multiple telegraph and the telephone, and (2) how the new goals interacted opportunistically. Finally we describe our implemented computational model, ALEC (Analogical Learning by Explaining Creatively), that accounts for the role of enterprise goals in invention.

The importance of discrimination shifts to learning and developmental psychology is highlighted. Basic tasks used in continuous and total change paradigms are presented, and major theoretical accounts are briefly reviewed. The lack of a general and comprehensive interpretation of human shift learning is identified, and a recent model based on neural network research is described. This model suggests that human adult performance in discrimination shifts differs from preschool performance because of a process called spontaneous overtraining. This hypothesis has been previously used in neural network simulations to successfully capture developmental regularities in continuous discrimination shifts (e.g., reversal and nonreversal shifts). In the present paper, new simulations using this model are applied to total change discrimination shifts (e.g., intradimensional and extradimensional shifts). Several developmental regularities are successfully captured by the networks. The contribution of the spontaneous overtraining hypothesis is discussed.

MMD is an aigorithm that learns without supervision intensional definitions of classes of knowledge errors using data (similarity) and theory (causality). Causality can be especially useful when similarity fails to discover certain errors due to their entanglement in complex behaviors, while similarity can be especially useful when no causal relationships for robust co-occurring discrepancies are present in the background knowledge. This paper examines the individual and combined effectiveness of MMD's similarity and causality components in discovering error classes and classifying behaviors in which these errors occur. Experimental results show how similarity and causality can serve to complement each other in the discovery of novice PROLOG programmer errors.

Among data implying pessimistic conclusions about human rationality, one might include evidence from the notorious Newcomb's Problem (Nozick 1969), which has hitherto, however, been largely confined to the philosophical literature. After nearly thirty years of inconclusive discussion, Newcomb's Problem is still widely seen as exposing inadequacies of the current standard theory of rational decision since the most plausible principles of choice give conflicting recommendations. Thus, Jeffrey (1983) says that Newcomb's Problem may be seen "as a rock on which...Bayesianism...must founder". Despite a staggeringly vast literature of great technical subtlety and complexity, no solution has emerged. I offer a novel analysis which goes beyond merely giving the right answer to the choice problem by also revealing the source of its persistent intractability. If my solution suggests good news about human capacity for rational choice, it entails bad news about other important problem-solving abilities central to cognitive science.

How do people deal with ambiguity and indeterminacy of incoming information? The results of the two reported experiments indicate that both young children and adults tend to reduce ambiguity, systematically 'converting' noninformative propositions into more informative ones. Although young children and older participants use different strategies, high rates of conversions were found in both groups. These conversions seem to represent an adaptive cognitive constraint — a tendency to reduce ambiguity and to increase the informativeness of incoming information.

A passive spreading activation theory of incubation effects states that hints, encountered by chance after an unsolved problem has been put aside, direct spreading activation to solutions in memory. Results from three experiments reject this explanation. Pretested hints that were seen seconds before unsolved problems were retested did not aid resolution unless hints were intentionally used to help problem solving.

The problem of multiple instantiation is the ability to handle different instances of a unique object at the same time. For connectionist models that do not use a working area containing copies of items from a long-term knowledge base, the problem of multiple instantiation is a particularly difficult one. While people are able to deal with multiple instances, their performance when doing so is nonetheless poorer, which is not the case for symbolic models. A cognitive model should reflect competence, as well as its limits. Some connectionist solutions to the problem of multiple instantiation are mentioned in this paper. An new solution which makes use of semi-distributed representations is presented. This model does not separate the long term knowledge base from a working area and has no recourse to copies. This solution limits the process of multiple instantiation in a way that should better reflect human data.

Responding appropriately to multiple, fast evolving, high risk situations is difficult. We investigated the design of a decision support tool called a Response Manager within the context of naval air warfare. Our question was whether support, in the fonn of presenting response options for consideration, should be generic to all aircraft or specifically tailored to different types of aircraft. Specifically tailored options limit clutter on a display, but perhaps at the price of constraining options and decision making to an inappropriately small set. In our experiment, air warfare-trained naval ofiicers saw snapshots of air warfare situations consisting of a map of an airspace, detailed data about one aircraft, plus a set of response options. We varied the contents of the response sets. The results indicate that participants tended to give orders to execute responses that were congruent with the presented response set. Highly experienced ofiicers showed as large an influence of response set presentation as did less experienced officers. Separate threat assessment ratings, however, indicated that the specifically tailored response sets were not influencing threat assessments; they were influencing response selection only. We concluded that the response manager works more like a memory aid, by orienting attention toward the presented response options, rather than by biasing situation interpretations.

We specify a model for the conceptual interpretation of relative adjectives (like "big") , which covers a crucial aspect of the underlying comprehension process - the comparison to a norm that is associated with a comparison class. Building on an elaborate domain ontology and knowledge about intercorrelations, comparison classes are dynamically created depending on the context in which adjectival utterances occur.

We describe an experiment that examines observational learning of either rules or instances. Subjects were asked to learn a dynamic computer control task and were given either a specific goal, to make the computer produce a specific response, or a nonspecific goal, to find the pattern underlying the computer's behavour. Subjects either interacted directly with the computer (the 'models') or observed a model's learning trials (the 'observers'). Both the goal of the models and the goal of the observers were varied so that specific goal and non-specific goal models were crossed with specific goal and non-specific goal observers. We predicted that the goal of the observer and not the goal of the model would determine whether observers learned rates or instances and that learning through observation would hinder instance learning. These predictions were confirmed. Non-specific goal models learned rules whereas specific goal models learned instances. Non-specific goal observers also learned rules, irrespective of the goal of the model, but specific goal observers failed to learn at all. A subsequent test confirmed that the failure of the specific goal observers to learn was due to the lack of feedback about correct responses. When such feedback was provided, specific goal observers learned instances. However, the presence of feedback was detrimental to rule learning. When non-specific goal observers received feedback, they learned only instances. These results support the view that both goal specificity and the presence or absence of feedback guide learning by directing attention to either instance space or both instance space and rule space.

This paper examines the way in which high level semantic information influences the production and comprehension of pronouns. It reports a new type of verb semantic processing bias. We examine the effects of this bias on language comprehension.

We present a skill learning model Clarion. Different from existing models of high-level skill learning that use a topdown approach (that is, turning declarative knowledge into procedural knowledge), we adopt a bottom-up approach toward low-level skill learning, where procedural knowledge develops first and declarative knowledge develops later. Clarion is formed by integrating connectionist, reinforcement, and symbolic learning methods to perform on-line learning. We compare the model with human data in a minefield navigation task. A match between the model and human data is found in several respects.

Design sketches are believed to play essential roles in early conceptual design processes. Exploration of how sketches are essential for the formation of new design ideas is expected to bring important implications for design education and design support systems. Little research has been done, however, to empirically examine the ways in which designers cognitively interact with their own sketches. Using a protocol analysis technique, we examined the design thoughts of an architect from the following point of view; how he drew depictions, inspected depicted elements, perceived visuo-spatial features, and thought of non-visual functional or conceptual information. The findings suggest that design sketches serve not only as external memory or as a provider of visual cues for association of non-visual information, but also as a physical setting in which design thoughts are constructed on the fly.

This study investigated the dimensions of the psychological similarity space associated with decisions people commonly encounter in their life. Four distinct clusters were found within a three-dimensional space, which suggested that people classifiy decisions based on content per se (professional-personal), as well as general characteristics of the decisions, namely, importance and complexity. These findings have implications for decision strategy selection.

In the first experiment, participants learned an easy rule that allowed perfect categorization that they had to automate. During learning, a new ancillary dimension was systematically associated with the defining feature of each category. In the test phase, items in which the association created during learning was broken were categorized more slowly than those in which the association was present, even for participants who did not notice the association. However, when the category-defining and the ancillary features were reversed in a second experiment, we did not get the anticipated results: there was no effect of the implicit association created during the learning phase. Results are explained in terms of dependencies between properties during processing. It is argued that similarity to previous exemplars does not explain the results obtained here.

The debate between single and dual route accounts of cognitive processes has been generated predominantly by the application of connectionist modeling techniques to two areas of psycholinguistics. This paper draws an analogy between this debate and bilingual language processing. A prominent question within bilingual word recognition is whether the bilingual has functionally separate lexicons for each language, or a single system able to recognize the words in both languages. Empirical evidence has been taken to support a model which includes two separate lexicons working in parallel (Smith, 1991; Gerard and Scarborough, 1989). However, a range of interference effects has been found between the bilingual's two sets of lexical knowledge (Thomas, 1997a). Connectionist models have been put forward which suggest that a single representational resource may deal with these data, so long as words are coded according to language membership (Thomas, 1997a, 1997b, Dijkstra and van Heuven, 1998). This paper discusses the criteria which might be used to differentiate single route and dual route models. An empirical study is introduced to address one of these criteria, parallel access, with regard to bilingual word recognition. The study fails to find support for the dual route model.

Much of the research on scientific reasoning has investigated the use of explicit, hypothesis-testing strategies. However, there is evidence that scientific reasoning problems can be solved by exploration of the experiment space. This study investigates the strategies by which people explore the experiment space. We examine the relationship between the systematicity of this search and successful performance and find that improved problem-solving may be a function of systematic data collection strategies.

This paper proposes an experimental system called IST for the facilitation of users' analogical transfer by providing image-schemas. First, by taking the radiation problem as an example, we hypothesized that a proper image-schema can promote human analogical problem solving owing to its plasticity and ascertained the hypothesis based on a cognitive experiment. We then constructed the IST system which can provide image-schemas with plasticity, by using the extended techniques of analogical mapping and of constraint-based graphics.

In Wason's (1960) inductive learning task, subjects must discover a rule that governs the production of sequences of three numbers, such as '2-4-6', by generating new triples that receive feedback. Data obtained with Wason's original procedure suggest that people test few hypotheses before announcing their guess and mostly proceed on the basis of a positive-test strategy. These features are commonly regarded as lamentable aspects of reasoning agents who fail to appreciate normative models of hypothesis-testing. Such interpretations, however, are relative to the inferential context in which the behavior is observed. In the present study, Wason's original procedure was modified such that in one condition desirable consequences were associated with the production of positive exemplars and undesirable consequences with negative exemplars. In a second condition, the consequences were reversed. Subjects in the latter condition produced more exemplars, of a greater variety, and were more likely to discover the rule than subjects in the first condition. It seems then that in this second condition the hypothesis-testing strategy emerging from the subjects' appreciation of the cost and benefit of generating certain kinds of triples coincided with the normative strategy. However, since subjects in both conditions aimed to achieve different goals their hypothesis-testing strategies can, in that respect, be characterized as rational.

Although tutoring by expert human tutors is usually effective, it is not always. By contrasting cases where tutoring does and does not result in learning, we can find out what causes learning during tutoring. Approximately 125 hours of physics tutorial dialog were analyzed to see what features of the dialog were associated with learning. Successful learning appears to require that the student make an error or reach an impasse; too much help can prevent learning. Features of successful tutorial explanations appear to be different for different pieces of knowledge. For instance, some pieces of knowledge are learned only if the tutor emphasizes generalization, whereas other learning requires that the tutor first explain why the student's error is wrong.

Goal-specificity has been found to affect performance: In difficult tasks, specific goals may be detrimental for learning. Locke and Latham (1990) claimed that goal-specificity has an impact on performance via motivation. Vollmeyer and Rheinberg's (1998) cognitive-motivational process model proposed that cognitive and motivational processes interact. Therefore, we investigated if goal-specificity may change the nature of this interaction, by trying to fit different structural equations models for groups given a specific goal (SG) or a nonspecific goal (NSG). Before beginning a complex dynamic task, the SG group was given a specific goal to reach, but the NSG group only received a goal when they had to transfer their knowledge. We found that the SG group learnt less and had lower motivation during learning. Contrary to earlier claims, there was no direct effect of goal-specificity on initial motivation, but it did alter the interaction between strategies and motivation during learning. The empirical model for the SG group showed a strong effect of initial motivation on the learning process and goal-directed strategies were effective. For the NSG group motivation during the task and systematic strategies were important.

The purpose of this study is to propose a process model to explain elementary school students' solutions of three simple problems in elementary mechanics. Forty eight 5th grade students were given three drawings depicting objects of various sizes in different kinetic states or being pushed by a human agent. They were asked to say whether a force was being exerted on the objects and to explain why. A process model has been proposed to explain students' answers to the three questions. The innovation of the process model is that it attempts to link two levels of representation: A semantic level, where a concept is analysed in terms of the presuppositions, beliefs, and mental models that underlie it and a syntactic level that specifies how concepts are related to other concepts in hierarchical categories. The work has been validated by a computer model designed by the AI team (Vosniadou, Kayser, Champesme, loannides & Dimitrakopoulou, in press).

Associationist theories of causal induction model learning as the acquisition of associative weights between cues and outcomes. An important deficit of this class of models is its insensitivity to the causal role of cues. A number of recent experimental findings have shown that human learners differentiate between cues that represent causes and cues that represent effects. Our Bayesian network model overcomes this restriction. The model starts learning with initial stractural assumptions about the causal model underlying the learning domain. This causal model guides the estimation of causal strength, and suggests integration schemas for multiple cues. In this way, causal models effectively reduce the potential computational complexity inherent in even relatively simple learning tasks. The Bayesian model is applied to a number of experimental findings, including studies on estimation of causal strength, cue competition, base rate use, and learning linearly and nonlinearly separable categories.

Ideas, empirical data and methodologies from a broad range of disciplines are deployed in exploring the functional borderline between subsymbolic and symbolic processing in human cognition. Initial clarification of functional relationships between the two forms of representation involves a brain monitoring study based on the concept of 'semantic transparency.' The search for further clarification focusses on two major issues, the ontogenetic and phylogenetic origins of local neural areas and processes underlying formation of distal associations between them. Pursuing these objectives has proved to be a challenging, interdisciplinary enterprise. A model of development of local neural areas is presented which assigns a critical role to astrocytes and their interaction with adjacent neurons. An extension to include the phylogenetic dimension, draws on the concept of 'cortical inheritance', a largely ignored aspect of genetic theory. An account of distal association formation involves co-option of hippocampal place fields far a new use.

This paper explores the uncertainty aspects of human abductive reasoning. Echo, a model of abduction based on the Theory of Explanatory Coherence (Thagard, 1992a), captures many aspects of human abductive reasoning, but fails to sufficiently manage the uncertainty in abduction. In particular, Echo does not handle belief acquisition and dynamic belief revision, two essential components of human abductive reasoning. We propose a modified Echo model (UEcho), in which we add a learning mechanism for belief acquisition and a dynamic processing mechanism for belief revision. To evaluate the model, we report an empirical study in which base rate learning serves as a testbed for belief acquisition and the order effect serves as a testbed for belief revision.

Practicing artists, art students, and non artists were asked to respond to six different two-dimensional infinite patterns which they viewed via a new methodology that presents the stimuli as iterating dots on a computer screen. As evidenced in the drawings they made, most viewers searched for shapes with clearly defined edges in the "negative" background space, rather than for shapes as defined by clusters of dots. The process of shape definition using the figure/ground distinction and the issue that past experience influences our perceptions are discussed.

The idea that GOMS can be used to model HCI tasks within the organizational environment in which they occur is discussed and reviewed. An example in terms of satellite operations is provided.

A constructivist neural network model is presented that learns the past tense of English verbs. The model builds its architecture in response to the learning task in a way consistent with neurobiological and psychological evidence. The model outperforms existing connectionist and symbolic past tense models in terms of learning and generalization behavior, and it displays a U-shaped learning curve for many irregular verbs. The trained model develops a modular architecture with dissociations between regular and irregular verbs, and lesioning the different pathways leads to results comparable with neurological disorders. It is argued that the successof the model is due to its constructivist nature, and that the distinction between fixed-architecture and constructivist models is fundamental. Given this distinction, constructivist systems provide more realistic models of cognitive development.

This paper explores the use of an artificial neural network to investigate the mental representation of abstract noun meanings. Unlike concrete nouns, abstract nouns refer to entities that cannot be pointed to. Cues to their meaning must therefore be in their context of use. It has frequently been shown that the meaning of a word varies with its contexts of use. It is more difficult, however, to identify which elements of context are relevant to a word's meaning. The present study demonstrates that a connectionist network can be used to examine this problem. A feedforward network learned to distinguish among seven abstract nouns based on characteristics of their verbal contexts in a corpus of randomly selected sentences. The results suggest that, for our sample, the contextual representation of abstract nouns is in principle sufficient to identify and distinguish abstract nouns and thus meets the functional requirements of concept representation.

This paper describes a cross-disciplinary extension of previous work on infeiring the meanings of unknown verbs from context. In earlier work, a computational model was developed to incrementally infer meanings while processing texts in an information extraction task setting. In order to explore the space of possible predictors that the system could use to infer verb meanings, we performed a statistical analysis of the corpus that had been used to test the computational system. There were various syntactic and semantic features of the verbs that were significantly diagnostic in detemiining verb meaning. We also evaluated human performance at inferring the verb in the same set of sentences. The overall number of correct predictions for humans was quite similar to that of the computational system, but humans had higher precision scores. The paper concludes with a discussion of the implications of these statistical and experimental findings for future computational work.

Advocates of distributed cognition argue that cognitive accomplishments rely in part on structures outside the individual mind - structures located in other minds or in artifacts that we think with. This paper argues that, in some cases, interactional structure can also make essential contributions to cognition. The data are transcribed classroom discussions, in which teachers and students use language to establish both referential and interactional patterns. The analyses use techniques from linguistic pragmatics, to uncover emergent interactional structure in the conversations and to show how this structure might make essential contributions to the cognitive value of those conversations.

We investigated category-based inference tasics, contrasting the predictions of structural alignment theory as applied to categorization with those of feature-overlap models of similarity. We provide evidence for the differential level of importance of causal information in category-based inference tasks, as predicted by the systematicity principle (Centner, 1983). Our basic paradigm consists of a task in which participants decide between inferences based on shared causal antecedents or shared attributes. Experiment I demonstrated a preference for the causal inference when the target animal shares one attribute with one of the base animals and one causal antecedent with the other base. In Experiment 2, we found that this preference holds even when the target animal shares greater attribute similarity with the noncausal base (i.e., the target shares two attributes with one base and one causal antecedent with the other). Experiment 2 also served to demonstrate that this result can indeed be attributed to the influence of causal structure, and not to surface stimulus properties, such as sentence length. Overall, the results agreed with the predictions of structural alignment theory and were inconsistent with a feature-overlap account.

This study examined the acquisition of Japanese numeral classifiers in Japanese preschool children, ages 3 to 6, with a primary emphasis on developing comprehension ability. Numeral classifiers, which exist in a large number of Asian languages, are a group of morphemes that usually occur adjacent to quantity expressions. The selection of numeral classifiers is determined by the inherent semantic properties of the noun whose quantity is being specified, suggesting that developing patterns of comprehension should be linked to underlying patterns of semantic and conceptual development. Previous research claims that children acquire certain distributional patterns very early but that the acquisition of the semantic system is a very slow process. We argue instead that, different techniques and stimulus contrast sets reveal a much greater sensitivity to semantic relations in young children than was previously considered possible. Reasons for the apparent slowness in classifier acquisition are also discussed as are the broader implications for relations between grammatical and conceptual development.

The mental model theory postulates that reasoners build models of the situations described in premises, and that these models normally make explicit only what is true. A computer program revealed an unexpected consequence of the theory: it predicts that certain inferences should have compelling but erroneous conclusions. Two experiments corroborated the existence of such illusions in inferences about what is possible given quantified assertions, such as 'At least some of the plastic beads are not red.' Experiment 1 showed that, as predicted, participants erroneously inferred that impossible assertions were possible, and that possible situations were impossible, but they performed well with control problems. Experiment 2 demonstrated the existence of similar illusions in inferences from dyadic assertions, e.g. 'All the boys played with the girls'.

This paper describes a semantically based computational theory of natural language comprehension. The theory argues for a semantically rich lexicon whose entries can be described as monosemic, generative and image-like. The comprehension process uses the basic definition of a word to decide how new information is to be combined with what has been interpreted so far. Next, and more importantly, the background information is used to generate the meaning of the combined words. Other semantically based approaches are also reviewed, one each from the disciplines of AI, Cognitive Science, and Linguistics.

The field of Judgement and Decision Making has for some time been dominated by normative theories which attempt to explain behaviour in mathematical terms. We argue that such approaches provide httle insight into the cognitive processes which govern human decision making. The dominance of normative theories cannot be accounted for by the intractability of processing models. In support of this view, we present a processing account of performance on a simulated medical diagnosis task. The performance of the model, which includes learning, is compared with that of a normative (Bayesian) model, and with subject performance on the task. Although there are some caveats, the processing model is found to provide a more adequate account of subject performance than the Bayesian model.

Two main ideas are proposed in this article. Richard and Tijus (in press) shown that problem solving can be explained by object properties that subjects take into account during the solving process. Stable properties are those which can not be modified by an action (for instance, an object's size, shape, etc.) and unstable properties are those which can be modified by an action (for instance, an object's location). Our purpose is that the problem space (Newell & Simon, 1972) can be described by state properties and that this description permits explaining the subjective distance (in the subject's mind) between two states. We suggest that similarity between state properties guides a subject's search through the problem space and can lead subjects through irrelevant paths. We think that in this condition, the well known beneficial effect of impasse situations consists in the fact that they permit subjects to discover the relevant properties of objects, problem constraints, and goal properties. Two experiments are proposed here. Results obtained in the first experiment show that working on impasse situations before solving the problem improves performance. Results of the second experiment show that working on impasse situations allow subjects to discover the relevant properties of a problem space, and that the benefit can be extended to all problems sharing the same problem space (which naturally contain the same impasses), even if their initial and final states are different. These results shed some light on the beneficial effects of impasses in problem solving.

This study examines the effects of representational forms on the acquisition and transfer of problem solving strategies. Three isomorphic representations of the Tic-Tac-Toe are used as the experimental tasks. The experiment shows that different representations of a common structure lead to the discovery of different forms of a common strategy with varying degrees of generality. With a better representation, subjects not only learn faster but also acquire more general forms of the strategy. The transfer across different representations can be either positive or negative, and it is based on strategies, not on problem structures.

Do Remote Associates Test (RAT) problems measure the process of remote association? In the present study a new set of RAT problems was generated, and association norms were determined for each test word, providing an index of the remoteness of die association needed to solve each problem. The observed remoteness of each problem correlated with the difficulty of the problems.

This article addresses the effectiveness of search reformulation using query refinement mechanisms on the Internet. Cognitive load was measured using a secondary digit-monitoring task. The load was found to be lower when using the refinement mechanisms than when perusing document summaries - suggesting that the development of refinement mechanisms can make Internet searching easier. Two refinement mechanisms, one based on statistical term co-occurrence and the other on a shallow natural language parsing technique were tested. No difference in load was found, possibly because of the limited time that subjects spent in the refinement process.

Tonality, contour, interval, and hemisphere are important predictors of melody recognition. Using forced-choice comparisons, listeners attempted to recognize the contour and interval information for diatonic and nondiatonic melodies presented to the left or right ear. For diatonic melodies, scale was more salient than contour whereas listeners relied on contour in nondiatonic melodies.

Tests of the influence of categorization and comparison on the representations of relational categories show: 1) a category-based similarity effect, and 2) an indirect role of comparison via facilitated learning with pairwise presentation.

A hippocampal phenomenon known as the sharp wave is correlated with a cell firing pattern that recapitulates an earlier cell firing pattern. The earlier cell firing pattern is driven by external stimuli while the recapitulative cell firing arises spontaneously from within the hippocampus. We postulate that the sharp wave associated cell firing that occurs in the awake state provides the basis for several well-known phenomena that involve self remindings. The hypothesis explains the resolution of cognitive impasses by hypothesizing an explicit, localized, internal mechanism that reminds one of an initially unsuccessful memory retrieval. Combining this hypothesis with ideas expressed by others provides a two-fold view of sharp wave associated cell firing: Recapitulative cell firing (1) mediates the consolidation of intermediate hippocampal memory into long-term neocortical memory during slow wave sleep, and (2) drives implicit (unconscious) neocortical reprocessing of unresolved issues.

Cohen and Eichenbaum (C&E, 1993) proposed that the hippocampus supports compositionality and inherently, flexible relational transfer of learning. Based on this proposal, rats were tested for symmetrical transfer of learning after training on relations between locations. Since a rat's hippocampus supports its spatial abilities, and since a relational test was being conducted, it was predicted that a strong degree of transfer would be obtained. The finding, however, was a general lack of relational transfer of learning. These results appear to limit the generality of C&E's theory and also seem to constrain the theory that the hippocampus is the biological seat of compositionality.

We are currently developing a Route Guidance system, based on a mobile phone that gives Location Information. One of our research goals is to discover what features a 'Route Guidance Service' needs in order to be useful for pedestrians.

This paper describes the theoretical background of an Act-R model of discovery learning in a simulated, conceptual domain: optics. It is assumed that learning in a simulation context consists of both implicit and explicit learning. The Act-R model under development tries to capture both learning types.

By simulating game playing with neural networks, and by using human subjects, it is demonstrated that the interaction between two players in a game of Paper, Rocks and Scissors can give rise to emergent properties that are not inherent in the individual players.