Learning Syntax from Naturally-Occurring Bracketings

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2941–2949 June 6–11, 2021. ©2021 Association for Computational Linguistics 2941 Learning Syntax from Naturally-Occurring Bracketings Tianze Shi Cornell University tianze@cs.cornell.edu Ozan ˙Irsoy Bloomberg L.P. oirsoy@bloomberg.net Igor Malioutov Bloomberg L.P. imalioutov@bloomberg.net Lillian Lee Cornell University llee@cs.cornell.edu Abstract Naturally-occurring bracketings, such as an- swer fragments to natural language ques- tions and hyperlinks on webpages, can reflect human syntactic intuition regarding phrasal boundaries. Their availability and approx- imate correspondence to syntax make them appealing as distant information sources to incorporate into unsupervised constituency parsing. But they are noisy and incom- plete; to address this challenge, we develop a partial-brackets-aware structured ramp loss in learning. Experiments demonstrate that our distantly-supervised models trained on naturally-occurring bracketing data are more accurate in inducing syntactic structures than competing unsupervised systems. On the En- glish WSJ corpus, our models achieve an un- labeled F1 score of 68.9 for constituency pars- ing.1 1 Introduction Constituency is a foundational building block for phrase-structure grammars. It captures the notion of what tokens can group together and act as a single unit. The motivating insight behind this paper is that constituency may be reflected in mark- ups of bracketings that people provide in doing natural tasks. We term these segments naturally- occurring bracketings for their lack of intended syn- tactic annotation. These include, for example, the segments people pick out from sentences to refer to other Wikipedia pages or to answer semantically- oriented questions; see Figure 1 for an illustration. Gathering such data requires low annotation ex- pertise and effort. On the other hand, these data are not necessarily suitable for training parsers, as they often contain incomplete, incorrect and some- times conflicting bracketing information. It is thus an empirical question whether and how much we 1Our code is publicly available at https://github. com/tzshi/nob-naacl21. [ Republicans ] have been imploring the White House [ to compromise on [ the wage issue ] ]. Q: Who have been imploring something? Q: What have someone been imploring? Q: What will someone compromise on? A: Republicans A: To compromise on the wage issue A: The wage issue QA-SRL Wikipedia Science fiction (sometimes shortened to sci-fi or SF) is a genre of speculative fiction that typically deals with imaginative and futuristic concepts such as advanced science and technology, space exploration, time travel, parallel universes, and extraterrestrial life. Republicans have been imploring the White House to compromise on the wage issue. Context: Q&As: Corresponding bracketings: Figure 1: Two example types of naturally-occurring bracketings. Blue underlined texts in the Wikipedia sentence are hyperlinks. We bracket the QA-SRL sen- tence in matching colors according to the answers. could learn syntax from these naturally-occurring bracketing data. To overcome the challenge of learning from this kind of noisy data, we propose to train discrimina- tive constituency parsers with structured ramp loss (Do et al., 2008), a technique previously adopted in machine translation (Gimpel and Smith, 2012). Specifically, we propose two loss …

Content truncated to 500 words