Proceedings of the Annual Meeting of the Cognitive Science Society

Modern Darwinist perspective enables to deal with the study of several human phenomena, one of which is deception, that we define as a behaviour unfolded with the deliberate intention of producing or sustaining a state of ignorance or false belief in another person. volutionary Psychology, an emerging area inside Cognitive Science, represents a promising conceptual approach to the study of deception. According to it, knowledge on human mind can be improved by understanding the processes which, during evolution, shaped its rchitecture. This work traces back to the Evolutionary Psychology arguments (for a review see Cosmides & Tooby, 1987; Barkow, Cosmides & Tooby, 1992; Buss, 1995; 1999) and develops the hypothesis that deception is a behaviour underpinned by two psychological echanisms that evolved in response to problems posed by group living: the theory of mind and deontic reasoning.

The question of whether language influences thought or not has been much discussed and disputed in the cognitive science literature. A recent proposal by Lakoff and Johnson (1999) adds an interesting slant to this debate by arguing that although language can influence thought via conceptual metaphors, the overall shape of the human conceptual system is determined by its embodied, perceptual nature. In this way, language is ultimately the slave of thought. We present an experiment aimed at exploring this question empirically. Exploiting evidence that has shown that schema consistent priming can bias the outcome of reasoning tasks, we performed a study in a well mapped conceptual domain in order to examine whether mbodied experience or language is the greater determinant of conceptual inferences. In this study, we found that language, rather than thought, is maybe what counts.

The notion that memory for goals is organized as a stack is central in cognitive theory in that stacks are core constructs leading cognitive architectures. However, the stack over-predicts the strength of goal memory and the precision of goal selection order, while under- predicting the maintenance cost of both. A better way to study memory for goals is to treat them like any other kind of memory element. This approach makes accurate and well-constrained predictions and reveals the nature of goal encoding and retrieval processes. The approach is demonstrated in an ACT-R model of human performance on a canonical goal-based task, the Tower of Hanoi. The model and other considerations suggest that cognitive architectures should enforce a two-element limit on the depth of the stack to deter its use for storing task goals while preserving its use for attention and learning.

Serial attention is the process of focussing mentally on one item at a time. This process has two phases: attention switching and attention maintenance. Attention switching involves rapidly building up the activation of a new item to dominate old items. Attention maintenance involves letting the current item decay while in use to prevent it from intruding on the next item later on. SASM , a model based on this analysis, suggests that this balance of high initial activation followed by gradual decay reflects a strategic adaptation to ask demands on one hand and principles of memory on the other. The model makes novel and accurate predictions about response times and error rates, integrates past use and current context as memory activation sources, and integrates attention switching and attention maintenance into one unified account.

Two experiments explore how pronoun resolution is influenced by a) properties of discourse referents, specifically whether they are underspecified and in need of description, and b) the contribution of the pronoun-containing utterance, specifically whether it provides a description or specifies an event. W e find that these factors interact, such that when an underspecified referent is in focus, reading is facilitated for description continuations, but when a specified referent is in focus, reading is facilitated in event continuations when the specified referent continues as the topic. This study reveals one of the complex interactions that underlies pronoun resolution.

The evidence that the comprehension of abstract and concrete words differ prompts one to consider how the lexical representations for these word types differ. The context-availability model (Schwanenflugel & Shoben, 1983) suggests that abstract words are more difficult to process because associated contextual information stored in memory for these words is more difficult to retrieve than for concrete words. Schwanenflugel (1991) provides two hypotheses regarding how these differences in retrieval of contextual information may come about. Three simulations using context representations from the Hyperspace Analogue to Language (HAL) model of memory (Burgess & Lund, 1997; Lund & Burgess, 1996) are used to evaluate Schwanenflugel's hypotheses, as well as to provide insight into the representational differences between abstract and concrete words.

Many researchers have argued for a description of nature using multiple levels, or modes of description, as we call them. This paper focuses on a confusion that follows from the multiple-mode approach, a confusion due to the notion of causation between modes. Causation between modes is reinterpreted as ordinary causation but with cause and effect described in different modes. In the first part of the paper the framework of modes of description is presented. In the second part it is applied to examples from cognitive science, which are taken from debates on the mind-brain issue and the dynamical systems approach to cognition.

In the linguistic psychological literature, there is a classical distinction between direct and indirect speech acts. In particular, some theories claim that the latter are more difficult to produce and comprehend than the former. W e propose to abandon such a distinction in favour of a novel one between simple and complex speech acts. This distinction applies to any kind of pragmatic phenomena, from standard speech acts to non standard ones, like irony and deceit. Our proposal is based on the types of mental representations and mental operations involved in speech acts production and comprehension.

This study focuses on the "classical" topic of the relation between language and thinking. Empirical studies investigating the interaction of verbalization and problem solving show inconsistent results. Studies differ with respect to the instruction of verbalization and the characteristics of the task. The aim of the study is to compare the performance in a nonverbal problem with and without language. For this purpose we investigate the performance of six groups of subjects working under different conditions: some of them were disturbed in their language behavior, others were encouraged to verbalize. It could be shown that though they had to work on a non-verbal problem, subjects disturbed in their linguistic behavior showed a worse performance than controls. Furthermore it can be shown that ."hinking aloud" in itself does not guarantee an improvement of performance. Moreover there are specific aspects of thinking aloud supporting problem solving. Case studies reveal interesting results with respect to the specific structure of "helpful" verbalization. The differences found cannot be explained by different memory loads or by the degree of distraction.

Functionalists in philosophy of mind traditionally raise two major arguments against the type identity theory: (1) psychological states are multiply realizable so that there are no one-to-one mappings of psychological states onto neural states and (2) the most that evidence could ever establish is the correlation of psychological and neural states, not their identity. We defend a variant on the traditional type identity theory which we call heuristic identity theory (HIT) against both of these objections. Drawing its inspiration from scientific practice, heuristic identity theory construes identity claims as hypotheses that guide subsequent inquiry, not as conclusions of the research.

An important property of analogical reasoning is that resulting inferences can be used to acquire new knowledge in a target domain. However, little is known about what happens to memory for these inferences. In this study, we explore the link between analogical reasoning, inferences, and memory. We gave participants information on a political debate. Some subjects were given a short text and other subjects were given a long text to read. In addition, half the subjects were given an analogy at the end of the text. A week later, subjects were brought back and asked to recall the information. We were particularly interested in whether subjects would (a) remember the analogy, and (b) incorporate analogical inferences into their memory for the text. We found that when they were given more information, subjects did not report the analogy, but falsely included analogical inferences in their recall. Results were different when subjects were given a lesser amount of information they remembered the analogy and did not erroneously recall analogical inferences. Overall, the results indicate that memory for analogical inferences is highly related to the amount of information that people are given.

We ciaim mental models are a framework that allows to shed light on the phenomenon of presuppositions. A plan-based lexical representation for verbs, together with the effect of conversational implicatures that discharge possible mental models, are the key features of this proposal.

Does the language you speak affect how you think about the world English and Mandarin speakers talk about time differently. Is this difference between the two languages reflected in the way their speakers think about time? The findings of two RT experiments show that different ways of talking about time lead to different ways of thinking. In Experiment 1, Mandarin-English bilinguals were compared to native English speakers. The results suggested that Mandarin speakers used a "Mandarin way of thinking" even when they were "thinking for English". In Experiment 2, native English speakers were trained to talk about time in "a Mandarin way". Results showed that even after a short training, native English speakers behaved more like Mandarin speakers than like untrained English speakers. It is concluded that language is a powerful tool in shaping thought.

In this paper, we explore the relationship between metaphor and polysemy. We begin by discussing how novel metaphoric mappings can create new word meanings in the form of domain-general representations. Tuming next to consider the implications of this view for the on-line comprehension of figurative language, we suggest that there is a shift from comparison processing to categorization processing as metaphors are conventionalized. Finally, we describe a series of experimental findings that support the proposed account.

Current theories of language production disagree about the way in which conceptual accessibility influences syntactic processing (e.g. Bock, 1987; De Smedt, 1990). We present theoretical arguments that the assumption of highly incremental processing can only be reconciled with theories in which conceptual accessibility influences word order. We report a sentence recall experiment in Modern Greek that provides empirical support for this position. Our results demonstrate that Greek speakers prefer to place conceptually accessible entities in early word order positions, irrespective of grammatical function, contrary to previous findings for English (Bock & Warren, 1985; McDonald, Bock & Kelly, 1993). We interpret our results as evidence for highly incremental processing.

Philosophy has never settled into a stable position in cognitive science and its role is not well understood. One reason for this is that the methods philosophers use to study cognition look quite peculiar to other cognitive scientists. This paper explores the methods of philosophy, laying out some of the main kinds and looking at some examples, and makes some remarks about their value to cognitive science.

The development of a graphical representation for performing a task can potentially yield a greater understanding of the task domain, but it is itself a demanding task that can distract from the primary one of learning the domain. In this research, we investigated the impact of constructing versus receiving a graphical representation on learning and coaching the analysis of scientific arguments. Subjects studied instructional materials and used the Belvedere graphical interface to to analyze texts drawn from an actual scientific debate. One group of subjects used a box-and-arrow representation, augmented with text, whose primitive elements had preassigned meanings tailored to the domain of instruction. In the other group, subjects used the graphical elements as they wished, thereby creating their own representation. Our results support the following conclusions. From the perspective of learning target concepts, developing one's own representation may not hurt those students who gain a sufficient understanding of the possibilities of abstract representation, although there are costs in time on task and in the quality of the diagrams produced. The risks are much greater for less able students because, if they develop a representation that is inadequate for expressing the concepts targeted by instruction, they will use those concepts less or not at all. From the perspective of coaching students, a predefined representation has a significant advantage. If it is appropriately expressive for the concepts it is designed to represent, it provides a common language and clearer shared meaning between the student and the coach, enabling the coach to understand students' analysis more easily and to evaluate it more effectively against a model of the ideal analysis.

Traditionally, it has been assumed that rules are necessary to explain language acquisition. Recently, Marcus, Vijayan, Rao, & Vishton (1999) have provided behavioral evidence which they claim can only be explained by invoking algebraic rules. In the first part of this paper, we show that contrary to these claims an existing simple recurrent network model of word segmentation can fit the relevant data without invoking any rules. Importantly, the model closely replicates the experimental conditions, and no changes are made to the model to accommodate the data. The second part provides a corpus analysis inspired by this model, demonstrating that lexical stress changes the basic representational landscape over which statistical learning takes place. This change makes the task of word segmentation easier for statistical learning models, and further obviates the need for lexical stress rules to explain the bias towards trochaic stress patterns in English. Together the connectionist simulations and the corpus analysis show that statistical learning devices are sufficiently powerful to eliminate the need for rules in an important part of language acquisition.

In this paper we compare the behaviour of three competing accounts of decision making under uncertainty (a Bayesian account, an associationist account, and a hypothesis testing account) with subject performance in a medical diagnosis task. The task requires that subjects first learn a set of symptom/disease associations. Later, subjects are required to form diagnoses based on limited symptom information. The competing theoretical accounts are embodied in three computational models, each with a single parameter governing the learning rate. Subjects' diagnostic accuracy was used to calibrate the learning rates of the models. The resulting parameter-free models were then used to predict subjects' symptom querying behaviour in a subsequent task. The fit between the Associationist model's predictions and subject behaviour was poor. The fit was slightly better in the case of the Bayesian model, but the hypothesis testing account proved to provide the most adequate account of the data.

Current models of Human Sentence Processing fall into two broad categories: Constraint Satisfaction accounts, which emphasise the immediate access of the comprehension processes to detailed linguistic information as parsing progresses (e.g., MacDonald et al., 1994), and Syntax First accounts, which hold that parsing is essentially a two-stage process, with initial decisions being made on the basis of a subset of available information (see, e.g., Frazier, 1995). In this paper, we examine evidence from Mitchell (1987) which seems strongly to favour a syntax first position, suggesting that basic lexical information about verbs may have little influence on the eariy stages of sentence processing. We provide experimental evidence to show (a) that detailed linguistic information is available early, but (b) that bare NP adverbs (a type of modifier) are read surprisingly fast, a finding which appears difficult to reconcile with many current accounts of sentence processing.

Several studies have shown that similarity judgements involve a process of structural alignment akin to analogical mapping. In particular, it has been shown that people appear to rely more on the relational structure of scenes involving cross-mappings, if they have previously carried out a similarity judgement task on these scenes (e.g., Markman & Centner, 1993b). W e report a study which shows that similarity judgements do not necessarily invoke structural alignment but that other task demands and the materials presented are more critical in selecting the comparison mechanism used in a given situation. The wider implications of these results for models of similarity and comparison are considered.

Recent studies suggest that participants commonly abstract rules when learning concepts, but a remaining question is whether they retain and apply knowledge of individual instances subsequent to rule abstraction. Research in the category learning domain indicates that exemplar information is retained and that exceptions to a category rule have special status in memory (Palmed & Nosofsky, 1995). The present experiment examines whether these findings extend to function learning. Participants learned associations between stimulus and response magnitudes that were related according to a negative linear function. Twelve stimulus-response pairs were given, some consistent with the negative linear rule, others exceptions to the rule. After each of six training sessions, previously studied stimulus magnitudes were presented as tests of learning accuracy. Participants were also given extrapolation trials followed by a final recognition test that included old and new rule-congruent and rule-incongruent items. Extrapolation was extensive. In addition, analyses revealed poorer learning and recognition for exceptions than for rule-congruent items, plus a high rate of false alarms for new rule-congruent items. These findings suggest that although the conceptual knowledge acquired in function learning tasks centers on rules, exceptions to these rules do not have special status in memory.

In hindsight, people often claim to have known more in foresight than they actually did. For example, the confidence for one of several possible outcomes is larger when it is known that this particular outcome occurred. A widespread explanation of hindsight bias assumes that the feedback serves as an anchor. How precisely this anchor takes effect and why it leads to a bias towards the anchor value has not been satisfactorily answered yet. One possible mechanism to explain hindsight bias assumes that the encoding of the feedback leads to a selective activation of the item-specific knowledge base. As a result, specific information units are strengthened and are thus more likely to be recalled when a person tries to reconstruct his or her original judgment. We tested the effect of selective activation in two hindsight experiments. The results showed a clear hindsight bias in that the recalled confidence ratings were distorted towards the feedback. Moreover, the consequences of selective activation were evident in that more information favoring the feedback was recalled.

When comprehending combined concepts (e.g., 'peeled apples'), two kinds of features are potentially accessible. Phrase features are true only of the phrase (e.g., "white"), while noun features are true of both the phrase and the head noun (e.g., "round"). Phrase features are verified more quickly and more accurately than noun features. No satisfactory account of this phrase feature priority has been put forth. We propose that relevance can explain the phrase feature priority. In Experiment 1, the differential accessibility of noun and phrase features was reversed by context paragraphs that made noun features relevant. Experiment 2 more subtly replicated this effect using a single-word context. We conclude that the phrase feature priority is attributable to the discourse strategy of assigning relevance to modifiers of combined concepts.

When assessing the likelihood of an event, human judgment is often inconsistent with the rules inherent in standard probability theory. For example, the judged probability of an event can be heavily influenced by the alternatives that are explicitly presented. Tversky and Koehier (1994) attempted to account for this phenomenon by arguing that probability judgments are made by comparing the amount of cognitive support one holds in favour of the event in question relative to all other possibilities. They suggested that different descriptions of the same event elicit different amounts of support resulting in different probability ratings. In addition to the role played by explicitly considered alternatives, the present paper suggests that people are also sensitive to the influence of alternatives that are not considered explicitly. We present the term "implied numerosity' in an attempt to indicate that probability ratings are influenced by a general impression of the number of potential alternatives that exist. Systematic differences in probability estimations may result from systematic changes in the perceived size of the category being evaluated.

We provide a computational account of the integration of various constraints proposed to be involved in the resolution of the direct object/sentence complement ambiguity. In the first part, competition-integration simulations show that a constraint-based model accounts for the results of Garnsey, Pearlmutter, Myers, and Lotocky (1997) at least as well as the garden-path model. In the second part, we compare the efficacy of norming techniques for capturing plausibility effects. Simulations show that norms designed to tap people's conceptual knowledge of events better capture plausibility effects than do norms that are biased toward tapping linguistic knowledge. We conclude that local information concerning event plausibility is an important constraint for understanding ambiguity resolution.

Researchers (see Siegler, 1987; Newell, 1973) have demonstrated the dangers of aggregating data over strategies. In this paper, we provide a current demonstration of this point using our recent work in the study of cognitive skill acquisition as a case study. Moreover, we call particular attention to the relation between cognitive modeling and microanalysis as driving forces toward a more thorough understanding of the role of strategies in cognitive skill acquisition.

For all intents and purposes, catastrophic interference, the sudden and complete forgetting of previously stored information upon learning new information, does not exist in healthy adult humans. But does it exist other animals? In light of recent research done by McClelland, McNaughton, & O'Reilly (1995) and McClelland & Goddard (1996) on the role of the hippocampal-neocortical interaction in alleviating catastrophic interference, it is of particular interest to ascertain whether catastrophic interference occurs in nonhuman higher animals, especially in those animals with a hippocampus and a neocortex, such as the rat. In this paper, we describe experimental evidence to support our claim that this type of radical forgetting does, in fact, exist for certain types of learning in some higher animals, specifically, in the rat's learning of time-durations. We develop a connectionist model that could provide an insight into how the rat might be encoding time-duration information.

Traditionally, thematic relatedness (chicken and egg) and similarity (chicken and turkey) have been thought of as distinct phenomena, the former the result of associative processes, and the latter reflecting comparison processes. However, recent studies (Bassok & Medin, 1996; Wisniewski & Bassok, 1996) suggest that similarity is a result of both association and comparison. This could call for a radical redefinition of similarity as inherently fused with association. We term this view the integration account. We consider an alternative, the confusability account, under which thematic influences intrude upon assessments of similarity but are not an essential part of the similarity process. W e present two experiments supporting the confusability account. The first indicates that comparison and association are independent processes. The second shows that thematic influences rise with increased cognitive load. We believe that while a redefinition of similarity is not warranted, similarity is more vulnerable to error and intrusion than is generally thought.

We propose a model of the representation and processing of uncertainty and use it to account for data from an experimental study of the use of probability words. Given two sentences, one using a probability word and the other phrased in terms of reasons-to-believe, subjects were asked to judge if the second was an acceptable paraphrase for the first. For certain word/paraphrase pairs there was a high degree of consensus about acceptability, for others the subjects were divided. We model the decision process as involving two stages. First, a concrete "mental" model is constructed which is consistent with the first phrase. The second phrase is then tested for compatibility with this model. In simulations two different representations for the meanings of phrases were tested, one based on probability intervals, and one based on qualitative argument structures. Both versions of the model give a good account for the data, both in terms of which paraphrases are judged to be acceptable and the relative proportions of subjects agreeing or disagreeing.

The mental model theory postulates that reasoners build models of the situations described in premises, and that each model represents a possibility. The present paper proposes that causal relations, such as A causes B and A allows B. have meanings that concern only possibilities and a temporal constraint that B cannot precede A. This theory predicts that causes and enabling conditions differ in meanings, contrary to a long tradition in philosophy and psychology that they are logically indistinguishable. It also predicts that individuals should reason about causation on the basis of mental models rather than on fully explicit models. Three experiments corroborated these predictions.

This paper examines certain claims of "cognitive significance" which (wisely or not) have been based upon the theoretical powers of two distinct classes of connectionist networks, namely, the "universal function approximators", and recurrent finite-state simulation networks. Each class will be considered with respect to its potential in the realm of cognitive modeling. Regarding the first class, I argue that, contrary to the claims of some influential connectionists, feed-forward networks do not possess the theoretical capacity to approximate all functions of interest to cognitive scientists. By contrast, I argue that a certain class of recurrent networks (i.e., those which closely approximate deterministic finite automata, DFA) shows considerably greater promise in some domains. However, serious difficulties arise when we consider how the relevant recurrent networks (RNNs) could acquire the weight vectors needed to support DFA simulations.

We introduce a computational model of language comprehension that combines locality of syntactic and semantic analysis with incrementality of processing. As the model incorporates inheritance-based abstraction mechanisms we are able to specify a parsimonious inventory of abstract, simple and domain-independent semantic interpretation schemata.

In a series of experiments. Bock and colleagues have demonstrated that subjects show a reliable increase in the use of particular syntactic constructions after having heard and repeated that construction in an unrelated sentence. Aspects of the data seem to indicate that it is syntactic constituent structure, independent of meaning, that underlies the facilitation in these situations. In this study we investigate whether more semantic factors might also lead to priming, and specifically whether the assignment of a semantic role to a particular participant in a prime sentence can increase the probability of a target sentence whose structure allows a similar assignment. To test this we replicate Bock's study and include a further set of primes (provide-with primes) which have the syntactic constituent structure of the dative, but share semantic role assignment with the ditransitive. If syntactic priming were triggered by constituent structure alone, primes like this would lead to more dative responses, relative to a ditransitive prime. If semantic involvement were crucial, on the other hand, this prime should elicit more ditransitive responses. In this study we find significantly more ditransitive responses following the provide-with sentence than following a dative prime, and no difference between the provide-with and ditransitive primes, suggesting that semantic factors indeed play a role.

One of the hallmarks of inductive reasoning by adults is the diversity effect, namely that subjects draw stronger inferences from a diverse set of premise statements than from a homogenous set of premises (Osherson et al., 1990). However, past developmental work (Lopez et al., 1992; Gutheil & Gelman, 1997) has not found diversity effects with children age 9 and younger. In our own experiments, we found robust and appropriate use of diversity information in children as young as 5 years. For stimuli we used pictures of people and their possessions, rather than the stimuli concerning animals and their biological properties in past studies. We discuss implications of these results for models of inductive reasoning.

We present a category learning experiment in which subjects faced the knowledge selection problem, i.e., they needed to use their observations to determine which prior knowledge would be useful for learning. The issue of putting prior knowledge into neural network models is reviewed, and we present a new model which addresses the knowledge selection problem. This model gives a good account of the experimental results.

Some theories of analogical mapping predict that finding mappings based on relations between objects requires greater working-memory capacity than finding mappings based on attributes of individual objects. It follows that the ability to make relational mappings will be impaired by any manipulation that constricts available working memory capacity. This prediction was tested in two experiments using a mapping task that required finding correspondences between pairs of pictures in which a critical object was "cross-mapped" (attribute similarity supporting one mapping, relational similarity another). Working memory was constricted in Experiment 1 by requiring participants to maintain a digit load while performing the mapping, and in Experiment 2 by inducing anxiety using a speeded subtraction task administered prior to the analogy task. Both manipulations caused participants to produce fewer relational responses and more attribute responses. The findings support the postulated links among working memoiy, anxiety, and the ability to perform complex analogical mapping.

Despite centuries of using mathematical notation, surprisingly little is known about how mathematicians perceive equations. The present experiment provides an initial step in understanding what sort of internal representation is used by experienced mathematicians. In particular, we examined if mathematical syntax plays a role in how mathematicians encode algebraic equations, or if just a simple memory strategy is used. Participants in the experiment performed a memory recognition task that required them to identify both well-formed (syntactically correct) and non-well-formed sub-expressions of equations. As hypothesised, performance was significantly better for well-formed sub-expressions, a result which suggests that mathematicians do indeed use an internal representation based on mathematical syntax to encode equations.

This paper has three goals: (i) to present a partial description of the intricate semantic selectional restrictions on the noun phrases in what we call here the Causal Have Construction (CHC), (ii) to show that four and five-year old children are sensitive to these selectional restrictions without much exposure to CHCs, and (iii) to discuss some implications of these findings for theories of language and language acquisition. Our interest in this topic derives from the possibilities it opens up for a deeper understanding of the organization of the mental structures that give rise to these semantic selectional facts, an understanding which we believe implicates an intricate and nontrivial interaction between grammatical and conceptual knowledge.

Two-process theories of human cognition, that state that learning can occur by both associative and rule-based processes, are currently popular. We report two experiments which support such a view. Both employed a set of six stimuli which varied along a luminance dimension, and followed the same general design. That is, participants were trained to discriminate between the two stimuli in the middle of this set, before being tested on the whole set. In Experiment I, the length of training was varied. Following short training, participants' performance on test exhibited a peak-shift, and therefore may be explained in associative terms. After longer training, however, their behavior was consistent with rulebased learning. In Experiment II, the contingency during the training phase was varied. Participants in the 'Full Contingency' group performed in a manner consistent with rule-learning, while the 'Reduced Contingency' condition produced a peak-shift. These results are discussed in terms of McLaren, Green & Mackintosh's (1994) version of the associative/rule-based distinction.

In this paper we show that adaptivity is reduced when people become fatigued. Fatigued people adapt worse to changing probability distributions as compared to non-fatigued individuals. In an ACT-R model of the task we show that this decreased adaptivity is due to a decrease in the use of one specific strategy. We argue that the use of this strategy is decreased, because it places high demands on working memory. In previous research we also found indications that mental fatigue is related to changes in working memory functioning. We argue that modeling individual differences in performance will provide better insight in the processes involved in mental fatigue.

The elimination model is proposed as an account of the inverse base-rate effect (D. L. Medin & S. M. Edelson, 1988). A key-assumption is that participants sometimes rely on eliminative inference to decide among candidate categories. A new prediction is that there will be an inverse base-rate effect also for an entirely novel symptom presented in the transfer phase—a prediction that contrasts with that by ADIT (J. K. Kruschke, 1996). This was tested and confirmed in 2 experiments.

Verbal elaboration of a worked example has been shown to be helpful to learners before attempting to solve similar problems. This has been termed as the self-explanation effect. (Chi, Bassok, Lewis, Reimann & Glaser, 1989). This study examined how self-explanation changes before and after sequential problem solving rounds. We found that changes in self-explanation within an individual may affect individual performance across a series of problem solving episodes. Also, some participants appear to use the worked-out example as a self-generated feedback (SGF) mechanism to help with their problem solving rounds, while other participants do not. Locations or points in a worked-out example where self-explanation (elaboration) is most likely to occur for students with higher performance scores versus those with lower performance scores, is discussed. The implications of these differences for the design of a computational cognitive model are also addressed.

We present the beginnings of a model of the human capacity to learn abstract invariants, such as square. The model is founded on four primary assumptions, which we believe to be neurally plausible and generic: Metric space, Topology, Comparison operations (subtraction, greater-than/less-than), and Extraction of vertices. The model successfully learns to discriminate simple planar quadrilaterals, and generalizes that learning across variations in viewpoint and modest variations in shape.

A cascading neural loop model is proposed to address the question of how to represent continuous experience. A prediction of the model is that short-term memory decay should exhibit a set of bumps or dips superimposed on a smooth exponential base. The prediction was tested using a Brown- Peterson distractor task, with distractor intervals from 1 to 24 seconds spaced every second apart. In one study with 22 participants, fits of nested regression models indicated that peaking functions with periods near harmonics of 1.6 seconds provided a better description of the data than an exponential function alone. In a replication study with 29 participants, peaking functions with a period of 3.2 seconds provided the best fit. In both studies, 5 % rises above an exponential base were evident near 7, 10 to 11, 13 to 14, and 16 seconds. This short-term memory effect has not been reported before and needs further replication.

Insight problems cause impasses because they deceive the problem solver into constructing an inappropriate initial representation. The main theoretical problem of explaining insight is to identify the cognitive processes by which impasses are resolved. In past work, we have hypothesized two such processes: constraint relaxation and chunk decomposition. In the study reported here, we derive detailed predictions about the structure of eye movements from these hypotheses. Eye movement data from a study of match stick algebra problems were consistent with the predictions. The results support the view that a key component of creative thinking is to overcome the processing imperatives of past experience.

Evans, Barston and Pollard, (1983) found that on the syllogistic evaluation task participants tended to endorse believable conclusions as being valid but reject unbelievable conclusions as invalid. A phenomenon known as "Belief Bias". Additionally, they collected verbal protocols from participants and established that this influence of belief was primarily associated with initial reference to the conclusions of these syllogistic arguments. In contrast, better logical reasoning was associated with initial reference to the premises. This experiment was designed to try to direct participants' anention to either the conclusion or the premises of a syllogistic argument with the intention of manipulating participants' logical reasoning ability and susceptibility to belief. The results reflected an inability to alter the influence of beliefs, but in one condition where the conclusion was presented prior to the premises, there was a successful reduction in participants' reasoning ability. The results are discussed with respect to the current theories of belief bias.

In this paper, we compare the rhetoric that sometimes appears in the literature on computational models of category learning with the growing evidence that different theoretical paradigms typically produce similar results. In response, we suggest that concrete computational models, which currently dominate the field, may be less useful than simulations that operate at a more abstrcict level. We illustrate this point with an abstract simulation that explains a challenging phenomenon in the area of category learning - the effect of consistent contrasts - and we conclude with some general observations about such abstract models.

To date, the vast amount of research done on the isochrony of English speech rhythm has not accounted for the emerging organization of rhythmicity. Our observation that speech rhythmicity is naturally occurring and even preferred as a strategy for optimizing the production and perception of a language-related task has been left untested. A set of experiments were devised to simulate list reading, i.e., a finite set of word tokens that a speaker must convey to hearers. Three lists were used that differed in prosodic structure to investigate the effect of stress pattern on isochrony. The results are analyzed as a low-dimensional dynamical system in which stress determines the cycle of an oscillator. The subjects show consistency in their speech rhythm across all list conditions. There is evidence of attractor dynamics in list reading.

A model of humans playing the simple game of Paper Rock Scissors based on the ACT-R architecture (Anderson, 1993; Anderson & Lebiere, 1998) is presented. This model stores in long-term memory sequences of moves and attempts to anticipate the opponent's moves by retrieving from memory the most active sequence. This results in a tightly linked dynamical system in which each player drives the play of its opponent. The performance of this model as a function of the length of the sequences stored and the amount of noise in the system is investigated, and is compared to the performance of human subjects.

This research investigates the impact of time pressure and individual differences on learning in a Real-Time Dynamic Decision Making (RTDDM) task. Our empirical results indicate that high time pressure generates high cognitive loads inhibiting learning. The results also show that high time pressure have a differential impact on the learning of individuals with high or low Working Memory (WM) capacity. W e present a cognitive model based on ACT-R intended to explain learning in tiiis task. Our cognitive model simulates learning by recognizing regularities in the decision task, and building "chunks" that guide decision making (instance-based learning). We describe how the model will be used to explain the impact of time pressure and WM capacity by varying the number of chunks acquired by the system given alternative time pressure conditions and individual differences.

This study explores the self-organizing neural network as a model of lexical and morphological acquisition. We examined issues of generalization, representation, and recovery in a multiple feature-map model. Our results indicate that self-organization and Hebbian learning are two important computational principles that can account for the psycholinguistic processes of semantic representation, morphological generalization, and recovery from generalizations in the acquisition of reversive prefixes such as un- and dis-. These results attest to the utility of self-organizing neural networks in the study of language acquisition.

A set of analyses are presented that replicate the mediated priming effect (e.g., lion-stripes) using a naming latency task, and demonstrate that the mediated priming effect is influenced by individual differences in sensitivity to this priming effect. Previous research (Livesay & Burgess, 1998) has shown that stimulus differences are a major factor in whether or not mediated priming is obtained. The present research explores the influence of verbal ability on this effect. The primary finding is that individuals with low verbal ability are not sensitive to mediated word relationships, whereas, individuals with high verbal ability manifest a robust mediated priming effect

Previous studies of children's inductive reasoning have attempted to demonstrate that label information is preferred to perceptual similarity as the basis for inductive inference (Gelman and Markman, 1986; Gelman and Markman, 1987; Gelman, 1988). A connectionist model of the development of inductive reasoning predicts that this will only be true when the percepnial variability of category exemplars is high (Loose and Mareschal, 1997). W e report three studies investigating the model's predictions. Study 1 demonstrates that patterns of categorization can depend on percepnial variability. In study 2 we develop a set of stimuli with differing variability but equal discriminability. Study 3 demonstrates that young children's patterns of reasoning are more affected by the presence of category labels when the inference is from an exemplar of a more perceptually variable category. This study also demonstrates that the basis of inference is not explicable in terms of the ease of the ability to categorize of the stimuli. Implications for the original model are discussed.

An experiment was performed to test the hypothesis that people sometimes take physical actions to make themselves more effective problem solvers. The task was to generate all possible words that could be formed from seven Scrabble letters. In one condition, participants could use their hands to manipulate the letters, and in another condition, they could not. Results show that more words were generated with physical manipulation than without. However, an interaction was obtained between the physical manipulation conditions and the specific letter sets chosen, indicating that physical manipulation helps more for generating words in some circumstances than in others. Overall, our findings can be explained in terms of an interactive search process in which external, physical activity effectively complements internal, cognitive activity. Within this framework, the interaction can be explained in terms of the relative difficulty of generating words from the letters given in the different sets.

When subjects are asked to move items in a visual display in response to spoken instructions, their eye movements are closely time-locked to the unfolding speech signal. A recently developed eye-tracking method, the "visual world paradigm", exploits this phenomenon to provide a sensitive, continuous measure of ambiguity resolution in language processing phenomena, including competition effects in spoken word recognition (Tanenhaus, Spivey-Knowlton, Eberhard, & Sedivy, 1995). With this method, competition is typically measured between names of objects which are simultaneously displayed in front of the subject. This means that fixation probabilities may not reflect competition within the entire lexicon, but only that among items which become active because they are displayed simultaneously. To test this, we created a small, artificial lexicon with specific lexical similarity characteristics. Subjects learned novel names for 16 novel geometric objects. Objects were presented with high, medium or low frequency during training. Each lexical item had two potential competitors. The crucial comparison was between high-frequency items which had either high- or low-frequency competitors. In spoken word recognition, performance is correlated with the number of frequencyweighted neighbors (phonologically similar words) a word has, suggesting that neighbors compete for recognition as a function of frequency and similarity (e.g., Luce & Pisoni, 1998). W e found that in the visual world paradigm, fixation probabilities for items with high-frequency neighbors were delayed compared to those for items with low-frequency neighbors, even when the items were presented with unrelated items. This indicates that fixation probabilities reflect the internal structure of the lexicon, and not just the characteristics of displayed items.

This paper presents a connectionist model of correlation-based categorization by 10-month-old infants (Younger, 1985). Simple autoencoder networks were exposed to the same stimuli used to test 10-month-olds. Both infants and networks used co-variation information (when available) to segregate items into separate categories. The model provides a mechanistic account of category learning within a test session. It shows how distinct categories are developed and demonstrates how categorization arises as the product of an inextricable interaction between the subject (the infant) and the environment (the stimuli).

Neonates seem to perceive two ends of a partly occluded rod as two separate objects. However, by 4 months of age infants often appear to perceive a similar stimulus as comprised of a single unified object. Little is known about the mechanisms of development underlying this change. We constructed four connectionist models of how perception of object unity might develop in human infants, based on experience with a variety of visual cues known to be important to infants' performance. After exposure to a simulated visual environment, ail the models were able to perceive a partly occluded object as unified. A rich perceptual environment and the presence of units for internal representations were found to improve generalization of acquired unity knowledge. These results lend plausibility to mechanistic accounts of human percepnial development, based on learning the statistical regularities inherent in the normal visual environment.

The traditional language switch hypothesis, according to which bihnguals can selectively activate and deactivate either language, has been repeatedly challenged in recent studies. In particular, an eyetracking experiment investigating spoken language processing suggests that bilinguals maintain both languages active in parallel even during monolingual input. The present study extends this finding to circumstances exhibiting between-language competition, within-language competition, or both. In this experiment, we find evidence for lexical items in the first language interfering with processing of the second language. We find that, in addition to competing activation between languages, bilinguals (like monolinguals) encounter competition within languages. Moreover, the results suggest that when simultaneous competition is encountered from items in both languages, within-language competition may be stronger than between-language competition. It appears that a bilingual's irrelevant language continues to be processed even when not actively used. However, this phenomenon is considerably influenced by language mode, even when such variables as word frequency, phonetic overlap, and language preference are taken into account.

Research with bilinguals may provide insights into the complex relationship between autobiographical memory and language. The present paper suggests existence of language-dependent memory, where linguistic factors at the time of recall influence memory retrieval. In two experiments, Russian-English bilingual immigrants were interviewed using the word-prompt technique. In the first experiment, bilinguals retrieved more autobiographical memories when there was a match between language of recall and language of encoding than when there was a mismatch. More memories from the period before immigration were recalled in Russian than in English and more memories from the United States were recalled in English than in Russian. To examine the mechanisms underlying these results, the ambiance language and the word-prompt language were considered separately in the second experiment. Both the linguistic ambiance and the word prompt were found to influence recall of autobiographical memories. These results, and particularly the effect of linguistic ambiance on recall, suggest language-dependent memory.

We propose a formal criterion for delineating literal from figurative speech (metonymies, metaphors, etc.). It is centered around the notion of categorization conflicts that follow from the context of the utterance. In addition, we consider the problem of granularity, which is posed by the dependence of our approach on the underlying ontology, and compare our distinction with alternative reference-based explanations.

Simple Recurrent Networks (SRNs) have been widely used in natural language processing tasks. However, their ability to handle long-term dependencies between sentence constituents is somewhat limited. NARX networks have recently been shown to outperform SRNs by preserving past information in explicit delays from the network's prior output. However, it is unclear how the number of delays should be determined. In this study on a shift-reduce parsing task, we demonstrate that comparable performance can be derived more elegantly by using a SARDNET self-organizing map. The resulting architecture can represent arbitrarily long sequences and is cognitively more plausible.

We discuss two different kinds of thinking about what might have been: Counterfactual "if only" thinking about how things might have been different and semifactual "even if" thinking about how things might have tumed out the same. We report the results of an experiment that showed that the two kinds of thinking have different effects on cause and emotion judgements. The experiment provides the first demonstration that semifactual "even if" thoughts reduce peoples judgements of causality and their emotional reactions compared to no thoughts about what might have been, and it replicates recent findings that counterfactual "if only" thoughts increase peoples judgements of causality and their emotional reactions.

In order to produce coherent text, natural language generation systems must have the ability to generate pronouns in the appropriate places. In the past, pronoun usage was primarily investigated with respect to the accessibility of referents. That is, it was assumed that a pronoun should be generated whenever the referent was sufficiently accessible so as to make its resolution easy. W e found that such an explanation does not seem to account well for the patterns of pronoun usage found in naturally occurring texts. We present an algorithm for generating appropriate anaphoric expressions which takes into account the temporal structure of texts (as a discourse structuring device) and knowledge about ambiguous contexts. Other important factors in our algorithm are sentence boundaries and the distance from the last mention of the anaphor. We back up our hypotheses with some empirical results indicating that our algorithm chooses the right referring expression in 85% of the cases.

Participants made a same-different judgment of the internal features of two faces presented simultaneously on screen. Whereas responding to upright faces on "same" trials relied upon holistic processing strategies, responding to upright faces on "different" trials, as well as responding to inverted faces, relied upon part-based processing strategies. Our results are also cfrntrary to earlier reports in that we found that when attention is focused upon the internal features, presentation of these features alrnie is sufficient to form a discrimination judgment.

Readers who self-explain texts aloud understand more from a text and construct better mental models of the content. This study examined the effects of providing self-explanation training on text comprehension, as well as course grades. Effects of prior knowledge and reading skill were also examined in relation to the benefits of self-explaining and self-explanation training. In general, low-knowledge readers gained more from training than did high-knowledge readers.

Recently, Bayesian principles have been successfully applied to connectionist networks with an eye towards studying the formation of internal representations. Our current work grows out of an unsupervised, generative framework being applied to understand the representations used in visual cortex (Olshausen & Field, 1996) and to discover the underlying structure in hierarchical visual domains (Lewicki & Sejnowski, 1997). We modified Lewicki and Sejnowski's approach to study how incorporating two specific constraints—context and sparse coding—affect the development of internal representations in networks learning a feature based alphabet. Analyses of the trained networks show that (1) the standard framework works well for limited data sets, but tends to poorer performance with larger data sets; (2) context alone improves performance while developing minimalistic internal representations; (3) sparse coding alone improves performance and actually develops internal representations that are somewhat redundant; (4) the combination of context and sparse coding constraints increases network accuracy and forms more robust internal representations, especially for larger data sets. Furthermore, by manipulating the form of the sparse coding constraint, networks can be encouraged to adopt either distributed or local encodings of surface features. Feedback connections in the brain may provide context information to relatively low-level visual areas, thereby informing their abiUty to discover structure in their inputs.

This paper proposes that the set of frequencies that the human language processor keeps track of are those that are useful to it in learning. In a computational experimental setting, we investigate four liguistically motivated features which distinguish subclasses of intransitive verbs, and suggest that those features that are the most useful to automatically classify verbs into lexical semantic classes are related to mechanisms used in adult processing to resolve structural ambiguity.

Psychometric measures of ability are unsuited to computational descriptions of tasks, primarily because they cannot take process into account. Studies of aptitude-treatment interactions have often failed to replicate from task to task precisely because of this difficulty. The current study aligns psychometric measures with process accounts in the domain of multimodal reasoning. Learning from multimodal logic courses transfers to other reasoning tasks, and this transfer has been found to relate to differences in strategic use of graphical representations in proof construction. The current study is a replication and an extension of these findings. Different goal types are distinguished in terms of: their modality; whether they involve proofs of consequence or non-consequence; and whether they can be solved by constructing single or multiple cases. We report on the interaction of a range of psychometric measures, and the ways in which they relate to the development and deployment of strategies. In particular, students who develop coping strategies to overcome difficulties with certain problems find that these strategies arise at the expense of appropriate use of a variety of strategies. Our approach, which characterises goals in terms of their logical as well as phenomenal properties, supports a computational perspective on psychometric measures in reasoning tasks.

This study examined the effects of emotion on the long-term acquisition of a procedural skill over a five-day period. Two tasks were employed: a word association task (WAT) and a visual discrimination task (VDT). Over the initial four days of the study participants went through a mood induction procedure (MIP) then subsequently completed both tasks. Both tasks showed a reduction in reaction time consistent with the power law of learning. No significant change in reaction time between day four and day five (one week later) was noted suggesting the change in reaction time was robust. These data further suggest that emotion modifies the rate at which the VDT is acquired.

Information acquired early in life is processed faster than information acquired late in life. Moore and Valentine (1998) report celebrities' faces follows the same pattern of results. This is problematic for the account of age of acquisition (AoA) based on language development because knowledge of celebrities is acquired after early representations are formed in the phonological lexicon. Also, the effects of AoA in lexical decision tasks (LDT) are assumed to be the result of automatic activation of phonology from the printed word. Such an account would predict null effects of AoA on face processing tasks not requiring name production (i.e. names are not automatically accessed, Valendne, Hollis & Moore, 1998). Significant effects of AoA were established in three Experiments: reading aloud printed names, making familiarity decisions to celebrities' names and faces. It is argued that temporal order of acquisition rather than age of acquisition may be the chief determinant of processing speed.

Traditional theories of text memory and comprehension posit that text is represented and reconstructed based upon its semantic content. In contrast, Rubin (1995) found that poetic materials are remembered based not only on semantic content, but based also on the schematic constraints, such as rhythm and rhyme, present in the surface structure of the verse. Rubin's research has done much to record the phenomenon of memory for poetic, structured materials. The present study is an investigation of the effects of multiple schematic constraints on participants' recall for words in limericks. This study provides support for Rubin's claims that surface structure and schematic constraints facilitate recall for schema-consistent poetic materials. In addition, the present study extends the analysis of the effects of schematic constraints, illustrating that the schematic constraints present in structured verse serve to guide recall for schema-inconsistent material, making the inconsistent material schema-consistent upon recall.

The aim of this paper is to expand and replicate the findings of Wheeldon and Levelt (1995). They employed an internal speech monitoring task which required Dutch speakers to monitor silently generated words for target syllable or phoneme sequences. On the basis of the obtained data several claims were made concerning the locus, time-course and nature of the internal speech code. The series of experiments reported here examined these predictions using English stimuli. In contrast to the Dutch study, no evidence of any reaction time advantage to syllable over nonsyllable strings was found. A phoneme monitoring experiment replicated the left-to-right pattern of results observed by Wheeldon and Levelt. In addition, a perception version of the task failed to replicate these effects suggesting that they were independent of the position of the target in the speech stream. Implications of the results in terms of the time course of phonological encoding are discussed.

We examined the development of the ability to differentiate logically determinate from logically indeterminate problems. The results indicated that a) young children tend to reduce the number of empirical possibilities via "cutting" the second half of less informative propositions, b) these errors do not stem from encoding or recall errors, c) from elementary to middle school, children tend to increase their understanding of logical form, and d) this increase corresponds to a decrease in the rate of cuts.

We propose an account for the acquisition of grammatical relations using the concepts of connectionist learning and a construction-based theory of grammar. The proposal is based on the observation that early production of childhood speech is formulaic and the assumption that the purpose of language is communication. If one assumes that children's comprehension of multiword speech is not globally systematic, but based initially on semi-rote knowledge (socalled "pivot grammars"), a pathway through small-scale systematicity to grammatical relations appropriate to the child's target language can be seen. W e propose such a system and demonstrate a portion of the emergence of grammatical relations using a connectionist network.

An important determinant of picture and word naming speed is the age at which the words were learned, that is, their age of acquisition (AoA). Two possible interpretations of these effects are that they reflect differences between words in their cumulative frequency of use, or that they reflect differences in the amount of time early- and late-acquired words have spent in lexical memory. Both theories predict that differences between early- and late-acquired words will be smaller in older than younger adults. We report three experiments in which younger and older adults read words varying in AoA or frequency, or named objects varying in AoA. There was no effect of word frequency when AoA was controlled. In contrast, strong AoA effects which did not diminish with age were found. The implications of these results for theories of how AoA affects lexical processing are discussed.

Young children's performance on certain mapping tasks can be improved by introducing relational language (Gentner, 1998). We show that children's performance on a spatial mapping task can be modeled using the Structure-Mapping Engine (SME) to simulate the comparisons involved. To model the effects of relational language in our simulations, we vary the quantity and nature of the spatial relations and object descriptions represented. The results reproduce the trends observed in the developmental studies of Loewenstein & Gentner (1998; in preparation). The results of these simulations are consistent with the claim that gains in relational representation are a major contributor to the development of spatial mapping ability. We further suggest that relational language can promote relational representation.

A key question motivating research in perception and attention is how the brain represents visual information. One aspect of this representation is the coordinate or reference frame with respect to which visual features are encoded. To determine the frames of reference involved in human vision and attention, neurological patients with unilateral neglect have been extensively studied. Neglect patients often fail to orient toward, explore, and respond to stimuli on the left. The interesting question is: with respect to what frame of reference is neglect of the left manifested? W h e n a neglect patient shows a deficit in attentional allocation that depends not merely on the location of an object with respect to the viewer but on the extent, shape, or movement of the object itself, the inference is often made that attentional allocation must be operating in an object-based frame of reference. Via simulations of an existing connectionist model of spatial attention (Mozer, 1991; Mozer & Sitton, 1998), w e argue that this inference is not logically necessary: object-based attentional effects in neglect can be obtained without object-based frames of reference.

This study investigated the model selection problem in cognitive psychology: How should one decide between two computational models of cognition? The focus was on model "faithfulness, " which refers to the degree to which a model's behavior originates from the theoretical principles that it embodies. The guiding principle is that among a set of models that simulate human performance equally well, the model whose behavior is most stable or robust with variation in parameter values should be favored. This is because such a model is likely to have captured the underlying mental process in the least complex way while at the same time being faithful to the theoretical principles that guided the model's development. Sensitivity analysis is introduced as a tool for assessing model faithfulness. Its application is demonstrated in the context of two localist connectionist models of speech perception, TRACE and MERGE.

The suppression of logically valid inferences by the content or context of premises can be seen as an instance of knowledge having a detrimental influence on reasoning. Although Henle (1962) has claimed that invalid deductions are due to additional premises drawn from background knowledge, current research on content effects ignores the methodological implications of this claim. Elaborating on the suppression effect in conditional reasoning (Byrne, 1989), we present a knowledge-based approach that makes relevant features of background knowledge an integral part of the analysis. After identifying the sufficiency and necessity of conditions as the type of knowledge mediating the effect, we construct and validate task materials independently from any assessment of reasoning (Experiment 1). We then replicate and extend suppression effects in syllogism tasks (Experiment 2) and show that participants are able to couch their background knowledge in formally correct wordings (Experiment 3).

The question whether connectionism offers a new way of looking at the cognitive architecture, or if its main contribution is as an implementational account of the classical (symbol) view, has been extensively debated for the last decade. Of special interest in this debate has been to achieve tasks which easily can be explained within the symbolic framework, i.e., tasks which seemingly require the possession of a systematicity of representation and process, in a novel way in connectionist systems. In this paper we argue that connectionism can offer a new explanational framework for aspects of cognition. Specifically, we argue that connectionism can offer new notions of compositionality, content and context-dependence based on connectionist primitives, i.e., architectures, learning, weights and internal activations, which open up for new variations of systematicity.

Recurrent attractor networks have many virtues which have prompted their use in a wide variety of connectionist cognitive models. One of these virtues is the ability of these networks to leam articulated attractors — meaningful basins of attraction arising from the systematic interaction of explicitly trained patterns. Such attractors can improve generalization by enforcing "well formedness" constraints on representations, massaging noisy and ill formed patterns of activity into clean and useful patterns. This paper investigates methods for learning articulated attractors at the hidden layers of recurrent backpropagation networks. It has previously been shown that standard connectionist learning techniques fail to form such structured attractors over internal representations. To address this problem, this paper presents two unsupervised learning rules that give rise to componential attractor structures over hidden units. The performance of these learning methods on a simple structured memory task is analyzed.

In this research, we investigated the factors that mediate the use of prior knowledge in learning new procedures. Participants learned to operate two different versions of four tasks on a hypothetical device interface. At a conceptual level, all devices were operated in the same way. However, in some conditions, the appearance of the two versions was manipulated by changing the graphical appearance of the interface. A second manipulation concerned the physical layout: The position of the device controls, graphics, and gauges was either the same or different from one version to the next. Providing the same appearance and providing the same physical layout both increased the amount of transfer. These effects were additive, suggesting that the factors contribute independently to learning. Our interpretation is that appearance affects the use of semantic constraint, while layout affects the use of structural analogy.

The schema repertoire model claims that an explanation is constructed by selecting and articulating a schema. Novice evolutionary explanations are analyzed to identify the relevant schemas and to demonstrate competition among schemas. An intervention study shows that a newly acquired schema does not necessarily win the competition against previously acquired schemas. The difference between schemas and beliefs is emphasized.

The sensory sampling model (SESAM, P. Juslin & H. Olsson, 1997) accounts for the underconfidence observed in sensory discriminations with pair-comparisons. In the present study the model is applied to a single-stimulus task and a comparison is made with pair- comparisons. The model predicts that in the single-stimulus condition training with feedback should lead to poorer calibration with more underconfidence. In pair-comparison the feedback should have little or no effect on calibration. The results confirm these predictions.

An experimental study of task design expertise is reported wherein a set of 12 mathematics tasks were sorted by specialist designers of mathematics tasks and by experienced mathematics teachers without specialist design experience. Contrary to the frequent finding of increasing conceptual depth with increasing expertise, conceptual depth did not differ between groups. Teachers sorted on the basis of mathematical content earlier than designers, and were more specific in their content-based categories. Designers produced more sorts than teachers and were more individualistic in their sorting. These findings suggest that domain expertise does not necessarily impair creative problem solving, as has been suggested in other studies. Instead, expertise includes the ability to shift perspectives with respect to the domain.

This study investigates how speakers of typologically different languages, Turkish (verb-framed) and English (satellite-framed) express motion events in their speech and accompanying gestures. 14 English and 16 Turkish speakers narrated an animated cartoon and one motion event scene was selected for analysis. English speakers depicted this scene with one verb with a satellite "the cat rolls down ", combining manner and path of the motion in one clause. Whereas Turkish speakers used two verbal clauses (e.g., yuvarlanarak iniyor (rolling descends)), separating manner from path. Gestures showed a similar pattern. Turkish speakers compared to English were more likely to use a) pure rotation gestures (representing manner only) and b) pure trajectory gestures (representing path only). These findings support the claim that speakers of typologically different languages conceptualize motion events in different ways during on-line speaking. While more Turkish speakers represent two components of a motion event as separate, English speakers represent them as one unit.

Are perceptual categorization and recognition memory subserved by a single memory system or by separate memory systems? A critical piece of evidence for multiple memory systems is that amnesics can categorize stimuli as well as normals but recognize those same stimuli significantly worse than normals (Knowlton & Squire, 1993). An extreme case is E.P., a profound amnesic who can categorize as well as normals but cannot recognize better than chance. This paper demonstrates that the paradigm used to test E.P. and other amnesics may be fundamentally flawed in that memory may not even be necessary to categorize the test stimuli in their paradigm. We "induced" profound amnesia in normals by telling them they had viewed subliminally presented stimuli that were never actually presented. Without any prior exposure to training stimuli, subjects' recognition performance was completely at chance, as expected, yet their categorization performance was quite good.

We examined the time-course of the utilization of background knowledge in perceptual categorization by manipulating the meaningfulness of labels associated with categories and by manipulating the amount of time given to subjects to make a categorization decision. Extending a paradigm originally reported by Wisniewski and Medin (1994), subjects learned two categories of children's drawings that either were given standard labels (drawing by children from group 1 or group 2) or were given theory-based labels (drawings by creative or noncreative children); meaningfulness of the label had a profound effect on how new drawings were categorized. Half of the subjects were given unlimited time to respond, the other half of the subjects were given a quick response deadline; speeded response conditions had a relatively large effect on categorization decisions by subjects given the standard labels but had a relatively small effect on categorization decisions by subjects given the theory-based labels. These results suggest that background knowledge may have its influence at relatively early stages in the timecourse of a categorization decision.

As electronic data becomes widely available, the need for tools that help people gain insight from data has arisen. A variety of techniques from statistics, machine learning, and neural networks have been applied to databases in the hopes of mining knowledge from data. Multiple regression is one such method for modeling the relationship between a set of explanatory variables and a dependent variable by fitting a linear equation to observed data. Here, we investigate and discuss some factors that influence whether the resulting regression equation is a credible model of the data.

Current models of graph understanding typically address the encoding and interpretive processes involved during the course of comprehension and largely focus on the visual properties of the graph. An experiment comparing reasoning with two types of graph is presented. On the basis and scope of existing models, performance with the two graphs would not be predicted to differ substantially. There are substantial computational differences between the graphs, however. It is suggested, therefore, that an adequate model of graph use must incorporate different combinations of visual properties of the graphs, levels of graph complexity, interpretive schemas and task requirements.

PDP networks that use nonmonotonic activation functions often produce hidden unit regularities that permit the internal structure of these networks to be interpreted (Berkeley et al, 1995; Dawson, 1998; McCaughan, 1997). In some cases, these regularities are associated with local interpretations (Dawson, Medler & Berkeley, 1997). Berkeley has used this observation to suggest that there are fewer differences between symbols and subsymbols than one might expect (Berkeley, 1997). We suggest below that this kind of conclusion is premature, because it ignores the fact that regardless of their content, the local features of these networks are not combined symbolically. W e illustrate this point with the interpretation of a network trained on a variant of Hinton's (1986) kinship problem, and show how the network's behavior depends on the coarse coding of information represented by hidden unit bands, even when these bands have local interpretations. We conclude that nonmonotonic PDP networks actually provide an excellent example of the differences between symbolic and subsymbolic processing.

In the experiments of comparative visual search reported here, each half of a display contains simple geometrical objects of three different colors and forms. The two hemifields are identical except for one mismatch either in color or form. The subject's task is to find this difference. Eye-movement recording yields insight into the interaction of mental processes involved in the completion of this demanding task. We present a hierarchical model of comparative visual search and its implementation as a computer simulation. The evaluation of simulation data shows that this Three-Level Model is able to explain about 9 8 % of the empirical data collected in six different experiments.

We propose a model to characterize the type of knowledge acquired in Artificial Grammar Learning (AGL). In particular, we suggest a way to compute the complexity of different test items in an AGL task, relative to the training items, based on the notion of Shannon entropy: The more predictable a test item is from training items, the higher the likelihood that it will be selected as compatible to the training items. Our model is an attempt to formalize some aspects of inductive inference by providing a quantitative measure of the knowledge abstracted by experience. We motivate our particular approach from research in reasoning and categorization, where reduction of entropy has also been seen as a plausible cognitive objective. This may suggest that reducing (Shannon) uncertainty may provide a single explanatory framework for modeling as diverse aspects of cognition, as learning, reasoning, and categorization.

Current theories of human deductive reasoning make different claims about the representation of logical statements in memory. Syntactically-based theories claim that abstract logical forms are represented veridically in memory, separate from content, whereas semantic theories propose that naive reasoners represent combinations of possibilities that are based on the content of statements. We tested these predictions in two experiments in which participants had to recall and recognize statements of different logical forms. Results indicate that memory for logical form is not veridical, thus failing to support the syntactic view. In particular, results suggest that naive participants tend, whenever possible, to represent only a single possibility for a statement of any logical form. These findings are consistent with semantic theories of human deductive reasoning and have significant implications for all theories of reasoning.

In this paper we criticize existing computational models of lexicon for assuming that for every word there is a fixed number of word sense that must be searched for the proper meaning of that word in a context. We reject this sense enumerative view and argue for a different model of lexicon in which the effects of context are not limited to selecting a word sense, and selected senses can be contextually modulated. W e also explain how patterns of contextual effects could evolve in an exemplar-based fashion. A prototype implementation of this model is also discussed.

The paper presents theoretical propositions for modeling the expert radiologist. The propositions are twofold. First, a basic model is given to complement a recent connectionist symbolic framework (Raufaste, Eyrolle, & Marine, 1998). Empirical data have showed dissociation between two kinds of experts ("basic" and "super") with regard to cognitive flexibility. The difference is conceived as a kind of perseveration in basic experts. Hence, the basic model was combined with a Supervisory Attentional System (Norman & Shallice, 1986) into an "extended model". An analysis of cognitive activity is then presented within this framework, along with a new theoretical explanation of cognitive flexibility.

We present a feedback or recurrent, auto-associative model that captures several important aspects of causal learning and causal reasoning that cannot be handled by feed-forward models. First, our model learns asymmetric relations between cause and effect, and can reason in both directions between cause and effect. As a result it can represent an important distinction in causal reasoning, that between necessary and sufficient causes. Second, it predicts cue competition among effects and provides a mechanism for them, something which can only be done with feed-forward models by assuming that two separate networks are learned, a highly non parsimonious assumption. Finally, we show that contrary to previous claims, a feed-forward model cannot handle Discounting and Augmenting in causal reasoning, although a feedback model can. The success of our feedback model argues for a greater focus on such models of causal learning and reasoning.

Implicit and explicit learning were originally distinguished in terms of accessibility to verbal report. We identify evidence for the proposal that the implicit/explicit contrast corresponds to a divide between connectionist and symbolic representations. We show that explicit learning shows marked improvement between 4 and 8 years of age. This finding contrasts against very early implicit learning abilities, and concurs with other evidence on the progressive development of symbolic reasoning abilities.

The domain-specificity principle implies that domain-specific knowledge is the main determinant of scientific discovery. An alternative view is that scientists make discoveries by assembling and articulating abstract schemas. If so, prior activation of the relevant abstractions should facilitate discovery and comprehension. Two in vitro studies showed that abstract information can have as much or larger impact on scientific thinking as domain-specific information.

In this article I propose that categorization decisions are often made relative to causal models of categories that people possess. According to this causal-model theory of categorization, evidence of an exemplar's membership in a category consists of the likelihood that such an exemplar can be generated by the category's causal model. Bayesian networks are proposed as a representation of these causal models. Causal-model theory was fit to categorization data from a recent study, and yielded better fits than either the prototype model or the exemplar-based context model, by accounting, for example, for the confirmation and violation of causal relationships and the asymmetries inherent in such relationships.

A method is proposed for argumentation on the basis of information that characterizes the structure of arguments. The proposed method cein be used both to detect arguments emd to generate candidate arguments for rebuttal. No assumption of a priori knowledge about attack and support relations between propositions, advanced by the agents participating in a dialog, is made. More importantly, by using the method, the relations are dynamically established while the dialog is taking place. This aJlows incremental processing since the agent need only consider the current utterance advanced by the dialog participant, not necessarily the entire argument, to be able to continue processing.

In many recent models of word recognition, words compete to activate distributed semantic representations. Reports of faster visual lexical decisions for ambiguous words compared with unambiguous words are problematic for such models; why does increased semantic competition between different meanings not slow the recognition of ambiguous words? This study challenges these findings by showing that visual lexical decisions to ambiguous words whose meanings were judged to be unrelated were slower than either unambiguous words or ambiguous words whose meanings were judged to be related. We suggest that previous reports of an ambiguity advantage are due to the use of ambiguous words with highly related meanings.

A neural network instantiation of Decision Field Theory (Busemeyer & Townsend, 1993) for multiple choice decision tasks is presented. First it is shown how under certain situations this dynamic model reduces to two well-known static models of choice. Next, model simulations of two well-known findings in multiple choice decision literature are presented. The first is the effect of similarity (Tversky, 1972). Several choice models also predict this effect. However, a more challenging effect, which is not predicted by numerous static choice models is the decoy effect (Huber, Payne, & Puto, 1982). Simulations show that the current model predicts this finding by using the concept of lateral inhibition. Finally, predictions of the model are made about the dynamic nature of the deliberation process in the decoy effect. If empirical results are found to be in agreement with this prediction, it would be a strong test of the model.

To explain the origins of new scientific ideas, historians and philosophers of science point to examples where scientists appear to have drawn analogies between their scientific domain and some very different domain. By contrast, research from the psychology lab suggests that those kinds of analogies are very difficult to obtain in even the simplest situations. To resolve this potential conflict, we examine the analogies that occur in psychology lab group and formal colloquium settings. This approach can be viewed as a cross-sectional approximation of an historical analysis. We find that as the setting moves further away from the original discovery, the way different types of analogies appear to be used changes. In particular, analogies between very different domains are never used in reasoning in the lab group, whereas they are frequently used in reasoning in formal colloquium presentations. Yet, we find that analogy between very similar domains remains an important source of new ideas and a method for solving problems in scientific settings.

There is a growing debate among developmental theorists concerning the perception of causality in young infants. Some theorists advocate a top-down view, e.g., that infants reason about causal events on the basis of intuitive physical principles. Others argue instead for a bottom-up view of infant causal knowledge, in which causal perception emerges from a simple set of associative learning rules. In order to test the limits of the bottom-up view, we propose an optimal control model (OCM) of infant causal perception. OCM is trained to find an optimal pattern of eye movements for maintaining sight of a target object. We first present a series of simulations which illustrate OCM's ability to anticipate the outcome of novel, occluded causal events, and then compare OCM's performance with that of 9-month-old infants. The implications for developmental theory and research are discussed.

In analogical problem solving, non-isomorphic source/target relations are typically only investigated in contrast to the ideal case of isomorphism. We propose to give a closer look to different types of non-isomorphic source/target relations and varying degrees of structural overlap. We introduce a measure of graph distance which captures the "size" of partial isomorphism between two structures and we present two experiments investigating the influence of different non-isomorphic relations on analogical transfer In the first experiment we contrast transfer performance for isomorphic vs. source inclusive problems with high vs. low superficial similarity. In the second experiment we explore different types of partial isomorphisms: source inclusiveness, target exhaustiveness, and different degrees of source/target overlap. The results indicate that (1) transfer of isomorphs is not significantly influenced by superficial similarity but transfer of partial isomorphs is, and (2) partial isomorphs can be transferred successfully if the amount of structural overlap is at least as high as structurally differences. The experiments were inspired by some open design questions for the analogy module of IPAL (a computational model integrating problem solving and learning).

Two experiments investigated the grammatical encoding processes during the production of noun phrases consisting of an article, an adjective, and a noun. Experiment 1 shows that for noun phrases in German, with the adjective in prenominal position, the lemmas of the noun and the adjective, and the noun's grammatical gender are selected before utterance onset. Experiment 2 shows that for noun phrases in French, with the adjective in postnominal position, only the noun lemma and its grammatical gender are selected. This suggests that grammatical advance planning at the level of grammatical encoding can operate with the smallest full phrase which can be expanded rightwards during articulation. Furthermore, the data show that gender is selected irrespective of whether it surfaces in the eventual phonological form of the noun phrase or not. This result is in line with the assumption that the grammatical encoder operates independently of the phonological encoder.

How much information about meaning is contained in the statistical structure of the environment? LSA is a theoretical and practical tool that is challenging previous notions about what is contained in the statistical structure of the environment. This paper examines what kind of category knowledge can be obtained from the environment using LSA. In particular, two experiments are conducted with LSA to test what kind of category structure it embodies. LSA ratings about the relatedness of categories to their properties are compared with human judgments regarding the centrality of properties to the categories. LSA is found to capture aspects of property centrality for some object and event categories. However, it is found to only capture those aspects related to typicality: how often do members of the category have that property? LSA fails to capture other aspects of centrality that can be found in human category judgments. Thus, it appears that humans do bring other constraints to bear in shaping their categories.

Three experiments were designed to examine the influence of shape feature and stimulus familiarity on saccadic selectivity during visual search. Robust shape feature based guidance was found in Experiment 1. In contrast, familiarity-based guidance was much smaller in magnitude and was observed with an unfamiliar target (Experiments 2 & 3) but not with a familiar target (Experiments 1, 2 & 3). Results from the current study suggest that there are qualitative and quantitative differences between the saccadic selectivity produced by stimulus familiarity and that produced by low-level features.

We expand upon the case for believing that the initial precise splitting of the foveal projection to the visual cortex fundamentally conditions the whole process of visual word recognition. We explore the optimal behaviour of a split architecture that attempts to divide its processing load equally between its two halves. We successfully model three aspects of fixation behaviour in human readers: (a) the positioning of the optimal viewing position to the left of the midpoint of the word, (b) a displaced Gaussian curve of letter-report accuracy resembling an RVF advantage, (c) the tendency for shorter words not to be directly fixated.

The consonance constraint satisfaction model, recently used to simulate the major paradigms of cognitive dissonance theory, is extended to deal with emotional arousal phenomena in dissonance. The impact of arousing drugs is implemented in the simulations by a scalar that modulates the intensity of unit activations representing the relevant cognitions and the connection weights representing their implications. The simulations show that even exotic dissonance phenomena can be explained in terms of the relatively commn process of constraint satisfaction.

Contrary to a recent claim that neural network models are unable to account for data on infant habituation to artificial language sentences, the present simulations show successful coverage with cascade-correlation networks using analog encoding. The results demonstrate that a symbolic rule-based account is not required by the infant data.

This study examined the cognitive processes of decision making in an urban high school classroom in which tenth graders analyzed scientific evidence about current issues of technology and society. A computer program, called Convince Me (Schank, Ranney & Hoadley, 1996), provided scaffolding for making evidence-based decisions for the experimental group. During the course of instruction, both the control and experimental classes completed open-ended assessments. Student progress, in using evidence to support claims and in weighing benefits and drawbacks, was mixed. Reasons for the changes in decision making are offered.

Associative memoiy for familiar faces was investigated in two experiments. Pairs of familiar faces were presented for deep or shallow encoding; memory for these pairs was tested by presenting old-intact pairs, old-recombined pairs, and pairs consisting of one or two new faces. In Experiment 1, pairs consisted of two different individuals whereas in Experiment 2, pairs consisted of different views of the same individual, ia both experiments, explicit recognition was best for old-intact pairs under deep encoding conditions. No associative priming effects were obtained in either experiment despite using a simultaneous familiarity-judgment task, similar to one that has produced associative priming effects with words (e.g., Goshen-Gottstein & Moscovitch, 1995a). It is proposed that the different associative priming effects obtained with the two types of stimuli may arise from differences in the modular perceptual representation systems for faces and words.

Important component of expertise is the rapid pickup complex. task-relevant pattern structure, yet such skills seldom trained explicitly. We report initial results ying principles of perceptual learning to the essing of structure in mathematics, specifically the ection between graphed functions and their symbolic essions. Subjects in two experiments viewed graphs inctions and made a speeded, forced choice match from ral equations. Training consisted of many short trials lis active classification task involving examples of a tion (e.g., sine) subjected to various transformations. scaling, shifting, reflection). Experiment 1 used rastive feedback — the graph for a trial was shown erimposed on the canonical function to accentuatejformations. Subjects showed substantial performance s from 45 minutes of training and transferred to new inces, new function families and a new task. In eriment 2, with contrastive feedback removed, subjects ved no transfer to new functions. The results indicate value of perceptual training in producing lematical expertise and the value of contrastive back in particular.

Fedforward neural network models of cognitive development are reviewed within the framework of a functional distinction between learning and development. This analysis suggests that static architecture networks implement a learning theory, whereas generative architecture networks combine learning and development. Both types of networks are then evaluated m terms of genetic costs. Within a levels-of-innateness framework, generative architectures are viewed as more plausible than static ones. Static architecture networks appear to implement a form of nativistic elicitation.

ive reasoning and problem solving is error-prone. One such pattern is manifested in that people err more often when problems are indeterminate than when problems are determinate W e suggest that an incomplete problem representation could account for the observed pattern of errors. W e further contend that in verbal reasoning such incomplete representation stems from a lack of systematic representations of connectives (e.g., and, or, if, etc.), and, therefore, extemalization of relations denoted by sentential connectives should improve people's representations of multiple possibilities. These predictions were tested in three reported experiments. Results indicate that determinate problems were easier to represent and recall than indeterminate problems. Furthermore, there was a tendency to represent and recall indeterminate problems as if they were determinate ones by fruncating the number of possibilities compatible with the problem. Finally, external aids dramatically improved representation and recall of indeterminate problems. These results are discussed in relation to theories of representation and reasoning.

The mental model theory of reasoning postulates that reasoners build models of the situations described in premises, and that these models normally make explicit only what is true. The theory has an unexpected consequence: it predicts the occurrence of inferences that are systematically invalid. These inferences should arise from reasoners failing to take into account what is false. We report an experiment that corroborated the occurrence of these illusory inferences, and that eliminated a number of altemative explanations for them. Results illuminate the controversy among various current theories of reasoning.

Research on connectionist mapping from written to spoken forms in natural language is presented. For this task, the more plausible Simple Recurrent Networks were used instead of static Neural Networks. The model was trained on a Dutch monosyllabic corpus. The effects of frequency, length and consistency were examined and were found similar to reported data in psycholinguistic experiments.

Multiple agents, equipped with a feature-based phonetic model and a connectionist cognitive model, interact via the naming game, with lexicon formation and change as emergent properties of this complex adaptive system. We present a new description of the naming game, situating it as a general, implementation-independent paradigm. Our addition of richerphonetic and cognitive models provides the agents with a greater degree of cognitive validity than does earlier work, while enhancing the flexibility of the system and reproducing empirical results. Feature-based phonetics, piecewise reinforcement learning, and a connectionist architecture with local representation allows language discrimination based on schemata instead of entire utterances.

Recent work has emphasised the importance of thematic domains in sentence processing. T w o questionnaire studies examined whether thematic domains influence attachment of relative clauses to complex NPs in Japanese. The results suggest that definitions of thematic domains should be revised to cover left-branching stmctures, but do not support a distinction between domains associated with clauses and adpositional phrases.

Analyses of an idealized corpus of English speech to infants revealed that simple conditional decision rules can separate frequent bisyllabic wordsftxjm bisyllables not corresponding to words. If infants accurately represent speech in terms of syllables, and compute conditional statistics over these syllables, such computations have the potential to inform infants of likely English words.

In this paper I will present a detailed ACT-R model of how the task-specific knowledge for a new. complex task is learned. The model is capable of acquiring its knowledge through experience, using a declarative representation that is gradually compiled into a procedural representation. The model exhibits several characteristics that concur with Fitts' and Anderson's theories of skill learning, and can be used to show that individual differences in working-memory capacity initially have a large impact on performance, but that this impact diminished after sufficient experience, which is consistent with Ackermans's theory of skill learning. Some preliminary experimental data support these findings.

Speech is produced incrementally. The Incremental Parallel Formulator (De Smedt, 1996) is a computational model of grammatical encoding that takes this notion of incrementality into account. It predicts that the order and time-scale with which conceptual fragments activate lexical segments affect the syntactic shape of an utterance. We derived predictions firom this model and tested these in two online experiments. In these experiments, participants described computer animations in which two objects moved in upward or downward directions. We manipulated the availability of pieces of the conceptual input by withholding either the information about the movement direction, or about the identity of one of the objects for various amounts of time. The experiments showed that both the type and the temporal availability of conceptual information strongly affect the syntactic shape of an utterance.

We explore the effects of taking notes on problem-solving and learning in a scientific discovery domain. Participants solved a series of five scientific reasoning problems in a computer environment in which they had access to an online, unstructured notepad. The results show that participants who used the notepad performed better than those who did not use it. This improvement held even when these participants no longer used the notepad on subsequent tasks. However, not all uses of the notepad were equally effective; only those that involved deeper levels of processing were related to improved performance.

Bredart, Valentine, Calder and Gassi (1995) described an interactive activation and competition (lAC) model in which the lexical representations of people's names have inhibitory connections between each other, but do not receive inhibition from the representation of biographical properties. The model predicts that people would be slower to name a celebrity for whom two names are equally available than they would be to name an equally familiar celebrity for whom only one name is available. However, naming should only be slowed by competition from a competing name; a highly available biographical property should not increase face naming latency. These predictions were confirmed in a simulation of the model. The effect is referred to as the nominal competitor effect. Experiment 1 showed that participants who had practiced naming actors using both the actor's name (e.g. John Cleese) and the character's name (e.g. Basil Fawlty) were slower to produce the actor's name at test than were participants who had practiced producing only the actor's name. However, practice in naming the relevant television series (e.g. Fawlty Towers) did not inhibit subsequent production of the actor's name. In contrast to the semantic competitor effect in picture naming, the effect reported here was found to be long-lasting (Experiment 2).

In this paper, I present experimental data bearing on the controversial issue of the possible relationship between the frequency of language types and how easily they can be learnt. Using simple, artificial languages which only differ with respect to the properties we are interested in, I show that there does appear to be a relationship of some kind, although not as strong as one might have hoped. In particular, if a language type can be learnt relatively easily, then the models fail to predict its actual frequency in the real world. On the other hand, the connectionist models provide evidence that the language types which are unattested or highly infrequent are also impossible or hard to learn.

The standard approach guiding research on the relationship between categories and causality views categories as reflecting causal relations in the world. We provide evidence that the opposite direction also holds: Categories that have been acquired in previous leaming contexts may influence subsequent causal leaming. In three experiments we show that identical causal leaming experiences yield different attributions of causal capacity depending on the pre-existing categories that the leaming exemplars are assigned to. There is a strong tendency to continue to use old conceptual schemes rather than switch to new ones even when the old categories are not optimal for predicting the new effect. This tendency is particularly strong when there is a plausible semantic link between the categories and the new causal hypothesis under investigation.

Complex reasoning, such as that elicited by GRE 'logical' reasoning problems, is demanding for human reasoners and beyond the competence of any existing computer program. We report four experiments carried out to investigate the question of what makes these problems difficult. The experiments established three causes of difficulty: the nature of the logical task (Experiment 1), the nature of the incorrect foils (Experiment 2), and the nature of the correct conclusions (Experiments 3 and 4).

Number comparison is a task that has been widely used to investigate the mental representation of number magnitudes. It is frequently assumed that the mapping from numerals to a "mental number line" is compressive (i.e., logarithmic) or that magnitude representations have the property of scalar variability. In this study, w e simulate the process of selecting the larger of two numbers in a neural network model. We show that it is possible to account for the main experimental effects (e.g., the distance effect and the number size effect) with a simple architecture using a linear representation of numerical magnitudes. The compressive effects that are found in the reaction times emerge from the non-linear interactions that are intrinsic to the decision process.

One influential view about the attentional demands of the reading processes maintains that phonological assembly is less automatic and more attention-demanding than phonological retrieval. The strongest evidence is this respect is the release-from-competition (RFC) effect (Paap & Noel, 1991), in which the pronunciation of low frequency exception words is speeded when participants have to perform a concurrent memory task. However, the results of follow-up investigations have led to a sharp controversy regarding whether the phenomenon is real and whether it can be replicated or not. The debate has reached stalemate, partly because the discussion about architectural and processing assumptions has been carried out only in verbal terms. This paper investigates the RFC phenomenon through simulations with two computational models of reading, the Connectionist Dual-Process model (Zorzi et al., 1998) and the DRC model (Coltheart et al., 1993). Both models failed to reproduce the RFC effect, even when the specific assumptions made by Paap and Noel were accurately implemented in the simulations. This finding casts further doubts about the reality of the phenomenon.

The current debate over what conditions a scheme of mental representation needs to satisfy in order to explain the systematicity of thought is characterized in such a way that (contrary to Fodor, Pylyshyn, and McLaughhn) any complete representational scheme (whether classical or non-classical) can explain the systematicity of thought. Though FPM might reply that non-classical schemes only satisfy these conditions in an unprincipled fashion, this shifts the discussion to less empirical considerations. Recasting the debate, we show that FPM can maintain their objection of unprincipledness only at the price of representational pluralism. Our thesis is that one can maintain representational monism if one uses what we call structured encodings. This will be accomplished by spelling out a representational taxonomy that makes evident what properties need obtain for a given representational scheme to exhibit systematicity effects.