Description:

Adina Williams
Postdoctoral researcher in linguistics at FAIR NYC
https://wp.nyu.edu/adinawilliams/

Title: Sentence Understanding with Natural Language Inference

Abstract: Many recent advancements in Natural Language Understanding have relied on the semantically deep task of Natural Language Inference (NLI). NLI is a difficult and valuable task, because it requires a model to know enough of the meanings of two sentences to determine which entailment relation they stand in. We present a new, crowd-sourced NLI corpus constructed using insights from experimental linguistics for data collection. With our new corpus, we evaluate the performance of two popular sentence encoders. Models that learn to parse with only distant supervision from a downstream semantic task have been showing large performance gains, suggesting they might learn to construct linguistically sophisticated sentence representations. Thus, we ask: What grammars do such models discover? Are they similar to grammars of the kind devised by formal linguists? We find that, while their parsing strategies differ, neither of the two sentence encoders employs a (human-like) grammar. In sum, this talk provides two examples of how a robust computational semantics research program can potentially contribute both to advancing our scientific understanding of linguistics, and to engineering useful, artificially intelligent systems.