Learning to communicate about conceptual hierarchies

Robert D. Hawkins, Kenny Smith, Michael Franke, Noah D. GoodmanProceedings of the 12th International Conference on the Evolution of Language (Evolang12) 2018 · DOI

Abstract

Natural language provides remarkable flexibility in how we label the world. A single object can be a "Dalmatian," a "dog," or an "animal." While we often favor a "basic-level" term, our lexicons maintain a rich hierarchy of overlapping meanings. This research investigates why these multiple levels of reference coexist, proposing that hierarchical vocabularies are driven by the need for informativity across varying communicative contexts.

The Challenge of Hierarchical Lexicons

In cognitive science, the "basic-level" preference explains why we usually say "dog" instead of "mammal." However, the existence of more specific and more general terms remains a puzzle. While probabilistic models explain how we choose between existing terms, they don't explain why a language would evolve multiple levels of specificity in the first place. The authors hypothesize that the structure of a vocabulary is a direct response to environmental communicative pressures.

Experimental Design: The Reference Game

To test this, researchers employed a repeated reference game where pairs of participants developed a new communication system from scratch.

The experimental setup: (A) example stimuli, (B) the alien vocabulary, and (C) the underlying conceptual hierarchy.

Participants were assigned roles as "speaker" and "matcher." The speaker had to guide the matcher to a target object using a single word from a 16-word "alien" vocabulary. The objects were organized into a three-level hierarchy based on shape, pattern, and color.

The key manipulation involved the statistics of the context:

  • Uniform Condition: Trials required distinctions at fine-grained, intermediate, and broad levels with equal frequency.
  • Control Conditions: Trials were biased toward a single level of distinction (e.g., always needing to distinguish very similar objects).

Key Insights: Diversity Breeds Complexity

The study reveals that hierarchical lexicons are a product of contextual diversity.

  • Functional Relevance: A shared language only reflects the mind's hierarchical organization when that structure becomes communicatively relevant. If a broad term like "thing" always works, specific terms like "Dalmatian" never emerge.
  • Pragmatic Balancing: Language evolution balances effort and clarity. When the required level of specificity shifts frequently, the vocabulary evolves multiple tiers to maintain efficiency.

Results and Implications

The "Uniform" condition, which mirrored the diversity of real-world communication, was significantly more likely to produce hierarchical lexica. In these groups, participants developed "abstract terms" for intermediate levels, whereas control groups focused on labels for individual targets. This suggests that human coordination can rapidly converge on complex, functional vocabularies when faced with diverse communicative needs.

Limitations and Future Work

The study used a fixed vocabulary and a relatively shallow hierarchy. Future research could explore how these systems evolve when participants can create their own words or when navigating much deeper conceptual trees. Additionally, scaling these findings from pairs to larger communities would provide insight into the stability of hierarchical language systems.

Summary updated

Abstract

Natural languages provide speakers with remarkable flexibility in the labels they may use to refer to things In addition to the combinatorial explosion of modifiers afforded by compositionality (Partee, 1995), we have a number of lexicalized nominal terms at our disposal. Dalmatian, dog, and animal can all truthfully be used to talk about the same Dalmatian at different levels of specificity, with one level of the conceptual hierarchy -the basic-level -generally privileged over the others How these overlapping meanings are learned, and why speakers choose different levels of specificity in different contexts, is increasingly accounted for by probabilistic models of pragmatic language use (e.g.
BibTeX
@inproceedings{hawkins2018learning,
  author = {Robert D. Hawkins and Kenny Smith and Michael Franke and Noah D. Goodman},
  title = {{Learning to communicate about conceptual hierarchies}},
  booktitle = {Proceedings of the 12th International Conference on the Evolution of Language (Evolang12)},
  year = {2018},
  doi = {10.12775/3991-1.038},
}
Figure