What do you learn from context? Probing for sentence structure in\n contextualized word representations

Contextualized representation models such as ELMo (Peters et al., 2018a) and\nBERT (Devlin et al., 2018) have recently achieved state-of-the-art results on a\ndiverse array of downstream NLP tasks. Building on recent token-level probing\nwork, we introduce a novel edge probing task design and construct a broad suite\nof sub-sentence tasks derived from the traditional structured NLP pipeline. We\nprobe word-level contextual representations from four recent models and\ninvestigate how they encode sentence structure across a range of syntactic,\nsemantic, local, and long-range phenomena. We find that existing models trained\non language modeling and translation produce strong representations for\nsyntactic phenomena, but only offer comparably small improvements on semantic\ntasks over a non-contextual baseline.\n