Wouter Voorspoels

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

Features are at the core of many empirical and modeling endeavors in the study of semantic concepts. This article is concerned with the delineation of features that are important in natural language concepts and the use of these features in the study of semantic concept representation. The results of a feature generation task in which the exemplars and labels of 15 semantic categories served as cues are described. The importance of the generated features was assessed by tallying the frequency with which they were generated and by obtaining judgments of their relevance. The generated attributes also featured in extensive exemplar by feature applicability matrices covering the 15 different categories, as well as two large semantic domains (that of animals and artifacts). For all exemplars of the 15 semantic categories, typicality ratings, goodness ratings, goodness rank order, generation frequency, exemplar associative strength, category associative strength, estimated age of acquisition, word frequency, familiarity ratings, imageability ratings, and pairwise similarity ratings are described as well. By making these data easily available to other researchers in the field, we hope to provide ample opportunities for continued investigations into the nature of semantic concept representation. These data may be downloaded from the Psychonomic Society’s Archive of Norms, Stimuli, and Data, www.psychonomic.org/archive.

Norms of age of acquisition and concreteness for 30,000 Dutch words

Word processing studies increasingly make use of regression analyses based on large numbers of stimuli (the so-called megastudy approach) rather than experimental designs based on small factorial designs. This requires the availability of word features for many words. Following similar studies in English, we present and validate ratings of age of acquisition and concreteness for 30,000 Dutch words. These include nearly all lemmas language researchers are likely to be interested in. The ratings are freely available for research purposes.

Exemplars and prototypes in natural language concepts: A typicality-based evaluation

Are natural language categories represented by instances of the category or by a summary representation? We used an exemplar model and a prototype model, both derived within the framework of the generalized context model (Nosofsky, 1984, 1986), to predict typicality ratings for 12 superordinate natural language concepts. The models were fitted to typicality ratings averaged across participants and to the typicality judgments of individual participants. Both analyses yielded results in favor of the exemplar model. These results suggest that higher-level natural language concepts are represented by their subordinate members, rather than by a summary representation.

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.

How do people learn from negative evidence? Non-monotonic generalizations and sampling assumptions in inductive reasoning.

A robust finding in category-based induction tasks is for positive observations to raise the willingness to generalize to other categories while negative observations lower the willingness to generalize. This pattern is referred to as monotonic generalization. Across three experiments we find systematic non-monotonicity effects, in which negative observations raise the willingness to generalize. Experiments 1 and 2 show that this effect emerges in hierarchically structured domains when a negative observation from a different category is added to a positive observation. They also demonstrate that this is related to a specific kind of shift in the reasoners hypothesis space. Experiment 3 shows that the effect depends on the assumptions that the reasoner makes about how inductive arguments are constructed. Non-monotonic reasoning occurs when people believe the facts were put together by a helpful communicator, but monotonicity is restored when they believe the observations were sampled randomly from the environment.

Raising Argument Strength Using Negative Evidence: A Constraint on Models of Induction

Both intuitively, and according to similarity-based theories of induction, relevant evidence raises argument strength when it is positive and lowers it when it is negative. In three experiments, we tested the hypothesis that argument strength can actually increase when negative evidence is introduced. Two kinds of argument were compared through forced choice or sequential evaluation: single positive arguments (e.g., “Shostakovich’s music causes alpha waves in the brain; therefore, Bach’s music causes alpha waves in the brain”) and double mixed arguments (e.g., “Shostakovich’s music causes alpha waves in the brain, X’s music DOES NOT; therefore, Bach’s music causes alpha waves in the brain”). Negative evidence in the second premise lowered credence when it applied to an item X from the same subcategory (e.g., Haydn) and raised it when it applied to a different subcategory (e.g., AC/DC). The results constitute a new constraint on models of induction.

Representation at different levels in a conceptual hierarchy

The present study examines the influence of hierarchical level on category representation. Three computational models of representation - an exemplar model, a prototype model and an ideal representation model - were evaluated in their ability to account for the typicality gradient of categories at two hierarchical levels in the conceptual domain of clothes. The domain contains 20 subordinate categories (e.g., trousers, stockings and underwear) and an encompassing superordinate category (CLOTHES). The models were evaluated both in terms of their ability to fit the empirical data and their generalizability through marginal likelihood. The hierarchical level was found to clearly influence the type of representation: For concepts at the subordinate level, exemplar representations were supported. At the superordinate level, however, an ideal representation was overwhelmingly preferred over exemplar and prototype representations. This finding contributes to the increasingly dominant view that the human conceptual apparatus adopts both exemplar representations and more abstract representations, contradicting unitary approaches to categorization.

Can race really be erased? A pre-registered replication study

When encountering an unknown individual, social categorization of the individual has been shown to automatically proceed on the basis of three fundamental dimensions: People seem to mandatorily encode race, sex and age. In contradiction to this general finding, Kurzban et al. (2001) showed that race encoding is not automatic and inevitable, but rather a byproduct of categorization in terms of coalitions. In particular, they argue and empirically support that when other coalitional information is present, the encoding of race is spectacularly reduced. In the present contribution, we present a replication of the race-erased effect reported by Kurzban et al. First, we give a detailed overview of the hypotheses, the experimental methodology, the derivation of the sample size required to achieve a power of 95%, and the criteria that need to be met for a successful replication. Then we present the findings of an empirical test that met the requirements of our power analyses. Our results indicate that the encoding of race is indeed reduced when another coalitional cue is available, yet this reduction is less marked than in the original study. This experiment was preregistered before data collection at Open Science Framework, osf.io/vnhrm/.

Idealness and similarity in goal-derived categories: A computational examination

The finding that the typicality gradient in goal-derived categories is mainly driven by ideals rather than by exemplar similarity has stood uncontested for nearly three decades. Due to the rather rigid earlier implementations of similarity, a key question has remained—that is, whether a more flexible approach to similarity would alter the conclusions. In the present study, we evaluated whether a similarity-based approach that allows for dimensional weighting could account for findings in goal-derived categories. To this end, we compared a computational model of exemplar similarity (the generalized context model; Nosofsky, Journal of Experimental Psychology. General 115:39–57, 1986) and a computational model of ideal representation (the ideal-dimension model; Voorspoels, Vanpaemel, & Storms, Psychonomic Bulletin & Review 18:1006-114, 2011) in their accounts of exemplar typicality in ten goal-derived categories. In terms of both goodness-of-fit and generalizability, we found strong evidence for an ideal approach in nearly all categories. We conclude that focusing on a limited set of features is necessary but not sufficient to account for the observed typicality gradient. A second aspect of ideal representations—that is, that extreme rather than common, central-tendency values drive typicality—seems to be crucial.

Caveats for the spatial arrangement method: Comment on Hout, Goldinger, and Ferguson (2013).

The gold standard among proximity data collection methods for multidimensional scaling is the (dis)similarity rating of pairwise presented stimuli. A drawback of the pairwise method is its lengthy duration, which may cause participants to change their strategy over time, become fatigued, or disengage altogether. Hout, Goldinger, and Ferguson (2013) recently made a case for the Spatial Arrangement Method (SpAM) as an alternative to the pairwise method, arguing that it is faster and more engaging. SpAM invites participants to directly arrange stimuli on a computer screen such that the interstimuli distances are proportional to psychological proximity. Based on a reanalysis of the Hout et al. (2013), data we identify three caveats for SpAM. An investigation of the distributional characteristics of the SpAM proximity data reveals that the spatial nature of SpAM imposes structure on the data, invoking a bias against featural representations. Individual-differences scaling of the SpAM proximity data reveals that the two-dimensional nature of SpAM allows individuals to only communicate two dimensions of variation among stimuli properly, invoking a bias against high-dimensional scaling representations. Monte Carlo simulations indicate that in order to obtain reliable estimates of the group average, SpAM requires more individuals to be tested. We conclude with an overview of considerations that can inform the choice between SpAM and the pairwise method and offer suggestions on how to overcome their respective limitations.