Description’s role in language and concept learning
Learning from description seems straightforward: if you hear about a “blue dax,” you might think you should learn that daxes are likely to be blue. But, counterintuitively, you should do the opposite. This is because people describe the rare or interesting features of objects, and not their typical features. If you hear about a “blue dax,” it is likely that most daxes are not blue. In this talk, I will draw on corpus analyses of conversation, language modeling, and experiments with adults to suggest that learning from description poses a problem for associative models of language and concept learning. Corpus analyses of parent–child conversations show that when talking to children as young as 14 months old, parents use description to highlight the surprising, atypical features of things (e.g., “brown apple”) rather than their typical, generalizable features (e.g., “red apple”). That is, people remark on features that are surprising given their world knowledge, even when talking to a child whose world knowledge is still nascent. Because language is structured in this way, models of word meaning that associate co-occurring words (here, Word2vec) fail to capture typicality relationships between nouns and adjectives well. Finally, adults take into account that description remarks upon the remarkable when interpreting utterances about novel categories (e.g., “Pass me the blue dax”): they do not simply associate the descriptor with the novel category, but infer that the descriptor points out an atypical feature (e.g., that it is rare for daxes to be blue). Overall, we find that people produce and interpret description based on principles of informative communication—not veridical description of the world—and this raises problems for associative models of language and concept learning as well as natural language processing.