Taeuk Kim

Hanyang University

Assistant Professor

It has been recently known in the literature that pre-trained language models are to some extent aware of syntactic regularities usually represented by parse trees in linguistics. Inspired by such findings, in this talk, a new paradigm called "Constituency Parse Extraction from Pre-trained Language Models (CPE-PLM)" will be introduced, which enables one to derive constituency parse tree-like structures directly from the fixed parameters of language models. In addition, it will also be demonstrated that the trees extracted by the proposed method can function as a proxy for estimating the extent to which language models understand the notion of syntax.