Exemplar by feature applicability matrices and other Dutch normative data for semantic concepts

A mixture approach to vagueness and ambiguity

When asked to indicate which items from a set of candidates belong to a particular natural language category inter-individual differences occur: Individuals disagree which items should be considered category members. The premise of this paper is that these inter-individual differences in semantic categorization reflect both ambiguity and vagueness. Categorization differences are said to be due to ambiguity when individuals employ different criteria for categorization. For instance, individuals may disagree whether hiking or darts is the better example of sports because they emphasize respectively whether an activity is strenuous and whether rules apply. Categorization differences are said to be due to vagueness when individuals employ different cut-offs for separating members from non-members. For instance, the decision to include hiking in the sports category or not, may hinge on how strenuous different individuals require sports to be. This claim is supported by the application of a mixture model to categorization data for eight natural language categories. The mixture model can identify latent groups of categorizers who regard different items likely category members (i.e., ambiguity) with categorizers within each of the groups differing in their propensity to provide membership responses (i.e., vagueness). The identified subgroups are shown to emphasize different sets of category attributes when making their categorization decisions.

A mixture approach to vagueness and ambiguity

When asked to indicate which items from a set of candidates belong to a particular natural language category inter-individual differences occur: Individuals disagree which items should be considered category members. The premise of this paper is that these inter-individual differences in semantic categorization reflect both ambiguity and vagueness. Categorization differences are said to be due to ambiguity when individuals employ different criteria for categorization. For instance, individuals may disagree whether hiking or darts is the better example of sports because they emphasize respectively whether an activity is strenuous and whether rules apply. Categorization differences are said to be due to vagueness when individuals employ different cut-offs for separating members from non-members. For instance, the decision to include hiking in the sports category or not, may hinge on how strenuous different individuals require sports to be. This claim is supported by the application of a mixture model to categorization data for eight natural language categories. The mixture model can identify latent groups of categorizers who regard different items likely category members (i.e., ambiguity) with categorizers within each of the groups differing in their propensity to provide membership responses (i.e., vagueness). The identified subgroups are shown to emphasize different sets of category attributes when making their categorization decisions.

Contrast effects in typicality judgements: a hierarchical Bayesian approach

We examine the influence of contrast categories on the internal graded membership structure of everyday concepts using computational models proposed in the artificial category learning tradition. In particular, the generalized context model (Nosofsky, 1986), which assumes that only members of a given category contribute to the typicality of a category member, is contrasted to the similarity-dissimilarity generalized context model (SD-GCM; Stewart & Brown, 2005), which assumes that members of other categories are also influential in determining typicality. The models are compared in a hierarchical Bayesian framework in their account of the typicality gradient of five animal categories and six artefact categories. For each target category, we consider all possible relevant contrast categories. Three separate issue are examined: (a) whether contrast effects can be found, (b) which categories are responsible for these effects, and (c) whether more than one category influences the typicality. Results indicate that the internal category structure is codetermined by dissimilarity towards potential contrast categories. In most cases, only a single contrast category contributed to the typicality. The present findings suggest that contrast effects might be more widespread than has previously been assumed. Further, they stress the importance of characteristics particular of everyday concepts, which require careful consideration when applying computational models of representation of the artificial category learning tradition to everyday concepts.

A formal ideal-based account of typicality

Inspired by Barsalou's (Journal of Experimental Psychology: Learning, Memory, and Cognition, 11, 629-654, 1985) proposal that categories can be represented by ideals, we develop and test a computational model, the ideal dimension model (IDM). The IDM is tested in its account of the typicality gradient for 11 superordinate natural language concepts and, using Bayesian model evaluation, contrasted with a standard exemplar model and a central prototype model. The IDM is found to capture typicality better than do the exemplar model and the central tendency prototype model, in terms of both goodness of fit and generalizability. The present findings challenge the dominant view that exemplar representations are most successful and present compelling evidence that superordinate natural language categories can be represented using an abstract summary, in the form of ideal representations. Supplemental appendices for this article can be downloaded from http://mc.psychonomic-journals.org/content/supplemental.

The generalized polymorphous concept account of graded structure in abstract categories

Abstract categories present with graded structure. The extent to which feature commonality between exemplars and category provides a satisfying account of this graded structure varies from one abstract category to the other (Hampton, 1981). We investigate whether the incorporation of features that exemplars share with external categories yields an improved account of abstract categories' graded structures. In doing so, we follow the suggestion that abstract categories are relational in nature (Goldstone, 1996; Wiemer-Hastings & Xu, 2005). The generalized polymorphous concept model, which incorporates both types of features, is found to improve the account of typicality and category membership in three of seven studied abstract categories. These three categories are found to be the most abstract, suggesting that it is appropriate to think of abstract categories as varying along a continuum of abstractness/interrelatedness rather than as a distinct type of category altogether.

On domain differences in categorization and context variety

Membership in many natural categories is considered all-or-none, while membership in most artifact categories is found to be graded. We introduce an alternative for the prevailing view that this domain difference in categorization results from representational differences. The context variety account posits that an item's gradedness reflects the variety of contexts it appears in. Items that feature in a variety of contexts are assumed to be more likely to elicit a graded categorization response, since the suggested target category only provides one of many solutions to the question of the item's identity. We review earlier work that suggested a domain difference in context variety, with artifactual items appearing in a greater variety of contexts than natural ones. The context variety domain difference is established in two separate experiments but is shown not to explain the domain difference in categorization. A selection of artifactual and natural items, for which the domain difference in context variety is reversed, is presented for categorization in a third experiment. This selection, too, fails to provide evidence for the context variety account of categorization differences. The domain difference in categorization is shown to be robust against this manipulation. Context variety appears to have no bearing on categorization, so the context variety account is not a sustainable alternative to accounts that posit representational differences between natural and artifact categories.

EEG decoding of semantic category reveals distributed representations for single concepts

Achieving a clearer picture of categorial distinctions in the brain is essential for our understanding of the conceptual lexicon, but much more fine-grained investigations are required in order for this evidence to contribute to lexical research. Here we present a collection of advanced data-mining techniques that allows the category of individual concepts to be decoded from single trials of EEG data. Neural activity was recorded while participants silently named images of mammals and tools, and category could be detected in single trials with an accuracy well above chance, both when considering data from single participants, and when group-training across participants. By aggregating across all trials, single concepts could be correctly assigned to their category with an accuracy of 98%. The pattern of classifications made by the algorithm confirmed that the neural patterns identified are due to conceptual category, and not any of a series of processing-related confounds. The time intervals, frequency bands and scalp locations that proved most informative for prediction permit physiological interpretation: the widespread activation shortly after appearance of the stimulus (from 100 ms) is consistent both with accounts of multi-pass processing, and distributed representations of categories. These methods provide an alternative to fMRI for fine-grained, large-scale investigations of the conceptual lexicon.

Similarity, feature discovery, and the size principle

In this paper we consider the "size principle" for featural similarity, which states that rare features should be weighted more heavily than common features in people's evaluations of the similarity between two entities. Specifically, it predicts that if a feature is possessed by n objects, the expected weight scales according to a 1/n law. One justification of the size principle emerges from a Bayesian analysis of simple induction problems (Tenenbaum & Griffiths, 2001), and is closely related to work by Shepard (1987) proposing universal laws for inductive generalization. In this article, we (1) show that the size principle can be more generally derived as an expression of a form of representational optimality, and (2) present analyses suggesting that across 11 different data sets in the domains of animals and artifacts, human judgments are in agreement with this law. A number of implications are discussed.

A crossed random effects diffusion model for speeded semantic categorization decisions

Choice reaction times (RTs) are often used as a proxy measure of typicality in semantic categorization studies. However, other item properties have been linked to choice RTs as well. We apply a tailored process model of choice RT to a speeded semantic categorization task in order to deconfound different sources of variability in RT. Our model is based on a diffusion model of choice RT, extended to include crossed random effects (of items and participants). This model retains the interesting process interpretation of the diffusion model's parameters, but it can be applied to choice RTs even in the case where there are few or no repeated measurements of each participant-item combination. Different aspects of the response process are then linked to different types of item properties. A typicality measure turns out to predict the rate of information uptake, while a lexicographic measure predicts the stimulus encoding time. Accessibility measures cannot reliably predict any component of the decision process.

Category and feature identification

This paper considers a family of inductive problems where reasoners must identify familiar categories or features on the basis of limited information. Problems of this kind are encountered, for example, when word learners acquire novel labels for pre-existing concepts. We develop a probabilistic model of identification and evaluate it in three experiments. Our first two experiments explore problems where a single category or feature must be identified, and our third experiment explores cases where participants must combine several pieces of information in order to simultaneously identify a category and a feature. Humans readily solve all of these problems, and we show that our model accounts for human inferences better than several alternative approaches.

Detecting intra- and inter-categorical structure in semantic concepts using HICLAS

In this paper, we investigate the hypothesis that people use feature correlations to detect inter- and intra-categorical structure. More specifically, we study whether it is plausible that people strategically look for a particular type of feature co-occurrence that can be represented in terms of rectangular patterns of 1s and 0s in a binary feature by exemplar matrix. Analyzing data from the Animal and Artifact domains, we show that the HICLAS model, which looks for such rectangular structure and which therefore models a cognitive capacity of detecting feature co-occurence in large data bases of features characterizing exemplars, succeeds rather well in predicting inter- and intra-categorical structure.

Introduction to the special issue on formal modeling of semantic concepts

We introduce the special issue on formal models of semantic concepts. After outlining the research questions that motivated the issue, we summarize the rich set of data provided by the Leuven Natural Concepts Database, and provide an overview of the seven research articles in the special issue. Each of these articles applies a formal modeling approach to one or more parts of the database, attempting to further our understanding of how people represent and use semantic concepts.

Finding the features that represent stimuli

We develop a model for finding the features that represent a set of stimuli, and apply it to the Leuven Concept Database. The model combines the feature generation and similarity judgment task data, inferring whether each of the generated features is important for explaining the patterns of similarity between stimuli. Across four datasets, we show that features range from being very important to very unimportant, and that a small subset of important features is adequate to describe the similarities. We also show that the features inferred to be more important are intuitively reasonable, and present analyses showing that important features tend to focus on narrow sets of stimuli, providing information about the category structures that organize the stimuli into groups.

Combining feature norms and text data with topic models

Many psychological theories of semantic cognition assume that concepts are represented by features. The empirical procedures used to elicit features from humans rely on explicit human judgments which limit the scope of such representations. An alternative computational framework for semantic cognition that does not rely on explicit human judgment is based on the statistical analysis of large text collections. In the topic modeling approach, documents are represented as a mixture of learned topics where each topic is represented as a probability distribution over words. We propose feature-topic models, where each document is represented by a mixture of learned topics as well as predefined topics that are derived from feature norms. Results indicate that this model leads to systematic improvements in generalization tasks. We show that the learned topics in the model play in an important role in the generalization performance by including words that are not part of current feature norms.

Features of graded category structure: Generalizing the family resemblance and polymorphous concept models

Many real-world categories contain graded structure: certain category members are rated as more typical or representative of the category than others. Research has shown that this graded structure can be well predicted by the degree of commonality across the feature sets of category members. We demonstrate that two prominent feature-based models of graded structure, the family resemblance (Rosch & Mervis, 1975) and polymorphous concept models (Hampton, 1979), can be generalized via the contrast model (Tversky, 1977) to include both common and distinctive feature information, and apply the models to the prediction of typicality in 11 semantic categories. The results indicate that both types of feature information play a role in the prediction of typicality, with common features weighted more heavily for within-category predictions, and distinctive features weighted more heavily for contrast-category predictions. The same pattern of results was found in additional analyses employing rated goodness and exemplar generation frequency. It is suggested that these findings provide insight into the processes underlying category formation and representation.

Similar but not the same: a comparison of the utility of directly rated and feature-based similarity measures for generating spatial models of conceptual data

Spatial models are employed to represent conceptual data in a wide range of fields within psychological research. In order to generate spatial models, it is necessary to first obtain empirical similarity data. A number of methods are available for collecting these data, but little effort has been made to compare their relative utility. In this article, we compare directly rated and five feature-based similarity data types in regard to their ability to be adequately represented by a spatial model (representational goodness of fit), and the ability of the representations to predict three external empirical variables (predictive validity). The results indicate that the representational goodness of fit of the feature-based similarities is noticeably superior to the directly rated similarities, and that the predictions of representations derived from common feature similarity data are substantially more likely than the predictions of all of the alternative representations. It is suggested that these findings are highly relevant to researchers employing spatial models to represent conceptual data, given that direct pairwise ratings have generally been considered the "gold standard" means of obtaining empirical similarities.