Large language models are better than theoretical linguists at theoretical linguistics

Abstract

Large language models are better than theoretical linguists at theoretical linguistics, at least in the domain of verb argument structure; explaining why (for example), we can say both The ball rolled and Someone rolled the ball, but not both The man laughed and *Someone laughed the man. Verbal accounts of this phenomenon either do not make precise quantitative predictions at all, or do so only with the help of ancillary assumptions and by-hand data processing. Large language models, on the other hand (taking text-davinci-002 as an example), predict human acceptability ratings for these types of sentences with correlations of around r = 0.9, and themselves constitute theories of language acquisition and representation; theories that instantiate exemplar-, input- and construction-based approaches, though only very loosely. Indeed, large language models succeed where these verbal (i.e., non-computational) linguistic theories fail, precisely because the latter insist – in the service of intuitive interpretability – on simple yet empirically inadequate (over)generalizations.