Chart-based Zero-shot Constituency Parsing on Multiple Languages

Zero-shot constituency parsing is a recent methodology in unsupervised parsing that aims to extract parse trees from pre-trained language models (PLMs) with no extra training. This paper improves upon the existing paradigm by introducing the combination of a novel chart-based method and an effective ensemble technique, attaining performance competitive to other unsupervised parsers on English PTB. Furthermore, we broaden the range of zero-shot parsing application by examining languages other than English. Specifically, we first demonstrate that the approach is applicable to the languages that are equipped with their respective monolingual PLMs. Finally, we propose to introduce multilingual PLMs into the zero-shot parsing framework, confirming that it is possible to generate reasonable parses for sentences in nine languages in an integrated and language-agnostic manner.