Semantic categories of artifacts and animals reflect efficient coding

Introduction

• The IB account had explained cross-language variation in color naming and why soft categories and inconsistent naming can be efficient, but its generality beyond color remained unclear.Zaslavsky, Kemp, Regier, and Tishby (2018) also proposed an evolutionary process that roughly recapitulates Berlin and Kay’s (1969) hierarchy while capturing continuous aspects of color-category evolution.Source: Introduction · p. 1
• The paper tests whether the same communication model applies to artifacts and animals, domains qualitatively different from color and not previously comprehensively addressed in terms of efficient communication.Container overlap and inconsistent naming are established characteristics of the domain (Ameel, Storms, Malt, & Sloman, 2005; Ameel, Malt, Storms, & Assche, 2009).Source: Introduction · p. 1
• Container naming in Dutch and French, including overlapping and inconsistent patterns across monolinguals and bilinguals, is near-optimally efficient in the IB sense.The analysis considers the full naming distributions across a large set of stimuli.Source: Introduction · p. 1
• Aspects of Brown’s proposed animal-taxonomy hierarchy are captured by a sequence of efficient animal-naming systems along the IB theoretical limit.The results also support a role for both perceptual and functional features in shaping animal categories across languages.Source: Introduction · p. 1

• The Information Bottleneck (IB) balances lexicon complexity against reconstruction accuracy, with complexity measured by mutual information between meanings and words and accuracy inversely related to expected KL divergence.The framework assumes a Bayesian listener reconstructs the speaker’s intended meaning from the word.Source: Theoretical framework and predictions · p. 2
• An IB system minimizes the tradeoff objective, while β controls the balance and the optimal systems trace the theoretical efficiency limit as β shifts from 0 to infinity.Along this trajectory, systems become more fine-grained and complex while reaching maximal accuracy for their complexity level.Source: Theoretical framework and predictions · p. 2, Figure 2 · p. 2
• The framework predicts that language naming systems should deviate only slightly from the IB optimum, with their distributions also close to corresponding IB systems.The supplied passage describes these as predictions for testing the efficiency account.Source: Theoretical framework and predictions · p. 2
• Applying the framework to new domains requires specifying a meaning space and a prior p(m), which determines how frequently each meaning needs to be communicated.Meanings are represented as distributions over a set of relevant features.Source: Theoretical framework and predictions · p. 2

Study I: Container names

• Container naming tests whether IB predictions extend to overlapping categories, environmental adaptation, and bilingual lexicon convergence, including whether bilingual compromise carries an efficiency cost.Earlier work found English, Spanish, and Chinese container naming efficient against hypothetical systems, but did not consider full speaker naming distributions or explicitly contrast monolingual and bilingual systems.Source: Study I: Container names · p. 3

Data.

• The data include 192 household-container images, naming by Dutch and French monolinguals and bilinguals, and Dutch sorting judgments used to estimate container similarity.Naming involved 32 Dutch and 30 French monolingual speakers and 30 bilingual speakers; bilinguals completed the task in both languages, while 65 Dutch speakers performed sorting.Source: Data · p. 3

Model.

• The model represents each container through similarity-based distributions over the stimulus set and defines the need distribution by averaging language-specific least informative priors.Because container similarity data came from Dutch speakers, the study assumes they reflect a shared perceptual representation; the authors note that this representation requires further validation.Source: Model · pp. 3–4

Results

• Bilingual complexity-accuracy tradeoffs are closer to each other than monolingual tradeoffs, and a joint bilingual system has a 0.16% complexity reduction.The authors suggest this small reduction may accumulate across domains; the calculation excludes similar word forms.Source: Results · p. 4
• Dutch and French container naming lies near the IB theoretical limit in monolingual and bilingual conditions, with bilinguals achieving similar efficiency.The corresponding IB solution is at β_l ≈ 1.2 in all four conditions.Source: Table 1 · p. 4, Results · p. 4
• Both languages outperform all 10,000 hypothetical variants on lower-is-better fit scores; reruns with a uniform need distribution produce similar results.The uniform-prior fit to actual systems is not as good as with the least-informative prior.Source: Table 1 · p. 4, Results · p. 4
• Dutch and French IB systems reproduce soft category structure and similar inconsistent naming patterns, including categories around fles/bouteille and doos/boîte.The IB systems also capture the category tube in both languages, though some apparent discrepancies remain.Source: Results · pp. 4–5
• The findings support an efficiency account of container naming in Dutch and French, but two closely related languages do not establish generality across languages.The authors identify testing other, preferably unrelated, languages as future work.Source: Results · p. 5

Study II: Folk biology

• Brown’s six-stage hierarchy is treated as a rough cross-linguistic approximation because the underlying animal-taxonomy data are not fine-grained and the proposal has been criticized for data accuracy.The study therefore makes only coarse comparisons between Brown’s reported patterns and IB-derived systems.Source: Study II: Folk biology · p. 5
• The animal model uses 113 classes and 757 human-generated features, representing each class by its feature distribution and weighting classes by familiarity.The underlying database’s feature data and familiarity ratings were collected from Dutch speakers and translated to English.Source: Study II: Folk biology · pp. 5–6
• The IB hierarchy progresses from a non-informative single category to systems that roughly correspond to Brown stages 2 and 6.The k=2 system distinguishes fish from other animals; the k=4 system refines the bird-mammal category.Source: Study II: Folk biology · p. 6
• The k=3 system includes fish and wug, with a bird-mammal category of probability mass 0.8 versus 0.14 for wug, departing from Brown’s sequence.It lacks a snake category; the authors suggest Dutch-participant feature data may be biased toward Western societies.Source: Study II: Folk biology · p. 6
• The hierarchy captures some cross-linguistic taxonomic tendencies despite Western-biased feature data, while fine-grained cross-linguistic naming data are still needed for a fuller test.The authors interpret the result as supporting, to some extent, a shared underlying representation of animals across cultures.Source: Study II: Folk biology · p. 6
• Salient features in the IB systems are both perceptual, such as being big, and functional, such as being edible.The authors suggest that both feature types may shape animal categories across languages, consistent with pressure for efficiency.Source: Study II: Folk biology · p. 6
• The authors also propose a possible within-language reading of the hierarchy, but describe the link between within-language structure and cross-language evolution as speculative.They note that English speakers recapitulated a color hierarchy in a sequential pile-sorting task, and suggest a similar phenomenon may hold for animal categories.Source: Study II: Folk biology · p. 6

• The findings suggest that efficient coding under the Information Bottleneck principle governs semantic representations across colors, containers, and animals.Source: General discussion · p. 6
• The account explains how pressure for efficiency may relate to soft categories, inconsistent naming in monolinguals and bilinguals, and language evolution.Source: General discussion · p. 6
• Further research should test whether the findings extend to other semantic domains and, ideally, the lexicon as a whole.Source: General discussion · p. 6