Language reflects how people organize experience into categories, and cross-linguistic comparison can help to identify general principles that shape categorization. Here we argue that symmetry is one such principle, and present a symmetry-based theory that predicts whether category systems for a given domain tend to include an even or an odd number of categories. We test the theory against cross-linguistic data previously compiled for a range of domains and find that deictic day-naming and tense-marking systems tend to have an odd number of categories, but that systems for domains including seasons, phases of the moon, kinship, and cardinal directions tend to have an even number of categories. Our results therefore provide evidence of the widespread influence of symmetry on categorization across languages and domains.

Basic-level categories such as mouse and apple are thought to be psychologically more natural than superordinates such as animal and fruit. Among other properties, basic-level categories are claimed to be acquired early by children and to emerge early in the historical development of languages, and these claims appear to be supported by an information-theoretic approach developed by Corter and Gluck (1992). We show, however, that under realistic assumptions this information-theoretic approach actually predicts that basic-level categories emerge after superordinate categories. This finding suggests either that additional assumptions are required for an information-theoretic account to capture the basic-level advantage, or that the basic-level advantage itself warrants re-examination.

Recent research has shown that words or morphemes that are closer to each other in linear order tend to have higher statistical inter-predictability, measured as mutual information. We offer an explanation for this in terms of holistic chunking of inter-predictable symbols, which provides an efficiency gain in the retrieval of stored symbols to encode a message. Inter-predictable chunking then interacts with structural priming to produce the schematic linear structures that are characteristic of both syntax and morphology. We thus argue that predictability and efficiency play a key role in the emergence of grammatical structure, going beyond previous information-theoretic analyses of natural language. In this paper we articulate some fundamental principles of chunking and linearisation, and use a simple computational implementation to show that these are sufficient to produce natural-language-like structures, using NP-internal ordering as a case study.

A key function of the lexicon is to express novel concepts as they emerge over time through a process known as lexicalization. The most common lexicalization strategies are the reuse and combination of existing words, but they have typically been studied separately in the areas of word meaning extension and word formation. Here, we offer an information-theoretic account of how both strategies are constrained by a fundamental tradeoff between competing communicative pressures: Word reuse tends to preserve the average length of word forms at the cost of less precision, while word combination tends to produce more informative words at the expense of greater word length. We test our proposal against a large dataset of reuse items and compounds that appeared in English, French, and Finnish over the past century. We find that these historically emerging items achieve higher levels of communicative efficiency than hypothetical ways of constructing the lexicon, and both literal reuse items and compounds tend to be more efficient than their nonliteral counterparts. These results suggest that reuse and combination are both consistent with a unified account of lexicalization grounded in the theory of efficient communication.

People from every culture observe the natural world in detail and organise it into categories, and Western biology builds on this universal impulse towards classification. Here we provide a quantitative analysis of factors that shape folk and scientific classification of birds from areas associated with seven indigenous languages (Anindilyakwa, Innu, Saami, Tlingit, Tobelo, Tzeltal, and Zapotec). We find that traditional Linnaean taxonomies align better with folk categories than do modern phylogenetic classifications, which suggests that human perception is responsible in part for the correspondence between Linnaean and folk taxonomies. Perceptual similarity is difficult to measure at scale, but we use the recently released AVONET database to develop a proxy for the perceptual similarity between pairs of birds and find that traditional Linnaean taxonomies and perceptual similarity both independently predict folk categories. Our results therefore provide quantitative evidence for the view that perceptual similarity influences both scientific and folk classification.

The impressive recent performance of large language models has led many to wonder to what extent they can serve as models of general intelligence or are similar to human cognition. We address this issue by applying GPT-3.5 and GPT-4 to a classic problem in human inductive reasoning known as property induction. Over two experiments, we elicit human judgments on a range of property induction tasks spanning multiple domains. Although GPT-3.5 struggles to capture many aspects of human behaviour, GPT-4 is much more successful: for the most part, its performance qualitatively matches that of humans, and the only notable exception is its failure to capture the phenomenon of premise non-monotonicity. Our work demonstrates that property induction allows for interesting comparisons between human and machine intelligence and provides two large datasets that can serve as benchmarks for future work in this vein.

Language reflects how people organize experience into categories, and cross-linguistic studies have revealed universal tendencies in the categorization of domains such as kinship and color. Here we consider universal tendencies involving the parity of category systems, and develop and test a theory that predicts whether category systems for a given domain tend to include an even or an odd number of categories. Consistent with the theory, we find that deictic day-naming and tense-marking systems tend to have an odd number of categories, but that systems for domains including seasons, phases of the moon, kinship, and cardinal directions tend to have an even number of categories. Our theory is founded on the principle of symmetry, and our results therefore provide evidence of the widespread influence of symmetry on categorization across languages and domains.

Assistive robots have the potential to support independence for older adults with mobility limitations and to alleviate the demands of their care partners. Several design considerations are required to ensure that the users can successfully rely on the robot to carry out their tasks. Therefore, building trustworthy robots is necessary for wider acceptance of these assistive robots. Using a participatory design approach, we assessed various aspects involved in advancing the design of a trustworthy robot in home environments. This is a case study focused on supporting an older adult with mobility limitations and his care partner. Through several iterations of co-active development as a team, most of the tasks were accomplished to meet the needs of the older adult couple interacting with the robot. Our approach highlighted usability challenges, the merits of a multidimensional approach in evaluating trust, and co-design strategies to improve the trustworthiness of the robot.

Categorization is ubiquitous in human cognition and society, and impacts how we perceive and understand the world. In reflecting the needs and perspectives of their creators, no categorization system is entirely objective, and inbuilt biases can have harmful social consequences. Here, we propose methods for quantifying three kinds of category biases in hierarchical category systems. We present a study on two widely used library classification systems (the DDC and LCC) as large-scale examples of human categorization, and use our methods to quantify bias towards content associated with western (vs non-western) concepts in topic areas including history and religion. We find consistent evidence for western bias and show that the DDC tends to exhibit more western bias than the LCC. Our methods are general, and can be used to survey biases across topic areas, bias attributes, and hierarchical category systems.

The human lexicon expresses a wide array of concepts with a limited set of words. Previous work has suggested that semantic categories are structured compactly to enable informative communication. Informativeness is typically quantified with respect to an entire semantic domain and not at the level of individual names. We develop a measure of name informativeness using an information-theoretic framework grounded in visual object representations derived from natural images. Our approach uses computer vision models to characterize informativeness of individual names with respect to large-scale data in a naturalistic setting. We show that our informativeness measure predicts degrees of specificity in lexical categories more precisely than alternative measures based on entropy and frequency. We also show that name informativeness jointly captures within-category similarity and distinctiveness across categories. Our analyses suggest how the variability of names from a broad part of the lexicon may be understood through the lens of information theory.

This dataset contains all data that was used for: Han, S. J., Ransom, K. J., Perfors, A. & Kemp, C. (2023). Inductive reasoning in humans and large language models. Cognitive Systems Research. The code for this project can be found here.

Natural language expresses new concepts by reusing existing words or coining new ones. Previous studies have examined these word formation strategies separately through a functional lens, but it is unclear why one strategy might be preferred over another. In this study, we hypothesize that communicative and cognitive efficiency might predict the choice between lexical reuse and compounding for expressing an emerging concept. We test our hypothesis by developing a computational analysis of English word meanings that emerged over the past century. Our results suggest that strategy choice may be explained partly by a pressure for least effort. Our work contributes a novel connection between strategy choice in word formation and functional theories of language.

The impressive recent performance of large language models such as GPT-3 has led many to wonder to what extent they can serve as models of general intelligence or are similar to human cognition. We address this issue by applying GPT-3 to a classic problem in human inductive reasoning known as property induction. Our results suggest that while GPT-3 can qualitatively mimic human performance for some inductive phenomena (especially those that depend primarily on similarity relationships), it reasons in a qualitatively distinct way on phenomena that require more theoretical understanding. We propose that this emerges due to the reasoning abilities of GPT-3 rather than its underlying representations, and suggest that increasing its scale is unlikely to change this pattern.

Compounding is a common type of word formation exten- sively studied in linguistics and cognitive psychology. A growing line of research suggests that the lexicon supports efficient communication by balancing informativeness and simplicity. We propose that the formation of novel compounds reflects a similar tradeoff between informativeness and word length. We formalize this hypothesis in information-theoretic terms and develop a computational procedure to evaluate our hypothesis on English noun compounds that emerged over the past century. We find that attested compounds achieve more efficient tradeoffs between informativeness and word length than do alternative word forms. Our work demonstrates how word formation and compositionality can be connected with information-theoretic approaches to the design of the lexicon.

construction of larger-scale systems of knowledge: [...] Building these systems takes years, much longer than learning a single new word or concept, but on this scale too the final product of learning far outstrips the data observed

Versatile robotic caregivers could benefit millions of people worldwide, including older adults and people with disabilities. Recent work has explored how robotic caregivers can learn to interact with people through physics simulations, yet transferring what has been learned to real robots remains challenging. Virtual reality (VR) has the potential to help bridge the gap between simulations and the real world. We present Assistive VR Gym (AVR Gym), which enables real people to interact with virtual assistive robots. We also provide evidence that AVR Gym can help researchers improve the performance of simulation-trained assistive robots with real people. Prior to AVR Gym, we trained robot control policies (Original Policies) solely in simulation for four robotic caregiving tasks (robot-assisted feeding, drinking, itch scratching, and bed bathing) with two simulated robots (PR2 from Willow Garage and Jaco from Kinova). With AVR Gym, we developed Revised Policies based on insights gained from testing the Original policies with real people. Through a formal study with eight participants in AVR Gym, we found that the Original policies performed poorly, the Revised policies performed significantly better, and that improvements to the biomechanical models used to train the Revised policies resulted in simulated people that better match real participants. Notably, participants significantly dis-agreed that the Original policies were successful at assistance, but significantly agreed that the Revised policies were successful at assistance. Overall, our results suggest that VR can be used to improve the performance of simulation-trained control policies with real people without putting people at risk, thereby serving as a valuable stepping stone to real robotic assistance.