Transition-based Bubble Parsing: Improvements on Coordination Structure Prediction

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 7167–7182 August 1–6, 2021. ©2021 Association for Computational Linguistics 7167 Transition-based Bubble Parsing: Improvements on Coordination Structure Prediction Tianze Shi Cornell University tianze@cs.cornell.edu Lillian Lee Cornell University llee@cs.cornell.edu Abstract We propose a transition-based bubble parser to perform coordination structure identifica- tion and dependency-based syntactic analysis simultaneously. Bubble representations were proposed in the formal linguistics literature decades ago; they enhance dependency trees by encoding coordination boundaries and in- ternal relationships within coordination struc- tures explicitly. In this paper, we introduce a transition system and neural models for pars- ing these bubble-enhanced structures. Ex- perimental results on the English Penn Tree- bank and the English GENIA corpus show that our parsers beat previous state-of-the-art ap- proaches on the task of coordination structure prediction, especially for the subset of sen- tences with complex coordination structures.1 1 Introduction Coordination structures are prevalent in treebank data (Ficler and Goldberg, 2016a), especially in long sentences (Kurohashi and Nagao, 1994), and they are among the most challenging constructions for NLP models. Difficulties in correctly identify- ing coordination structures have consistently con- tributed to a significant portion of errors in state- of-the-art parsers (Collins, 2003; Goldberg and El- hadad, 2010; Ficler and Goldberg, 2017). These errors can further propagate to downstream NLP modules and applications, and limit their perfor- mance and utility. For example, Saha et al. (2017) report that missing conjuncts account for two-thirds of the errors in recall made by their open informa- tion extraction system. Coordination constructions are particularly chal- lenging for the widely-adopted dependency-based paradigm of syntactic analysis, since the asymmet- ric definition of head-modifier dependency rela- tions is not directly compatible with the symmetric 1Code at github.com/tzshi/bubble-parser-acl21. Bubble Tree: I prefer hot coffee or tea and a bun conj cc conj amod conj cc conj det nsubj obj UD Tree: I prefer hot coffee or tea and a bun nsubj obj amod conj cc conj cc det Figure 1: Bubble tree and (basic) UD tree for the same example sentence. (For clarity, we omit punctuation and single-word bubble boundaries.) Bubbles explic- itly encode the scope of the shared modifier hot with respect to the nested coordination, whereas the UD tree gives both tea and bun identical relationships to hot. nature of the relations among the participating con- juncts and coordinators.2 Existing treebanks usu- ally resort to introducing special relations to rep- resent coordination structures. But, there remain theoretical and empirical challenges regarding how to most effectively encode information like modi- fier sharing relations while still permitting accurate statistical syntactic analysis. In this paper, we explore Kahane’s (1997) alter- native solution: extend the dependency-tree repre- sentation by introducing bubble structures to ex- plicitly encode coordination boundaries. The co- heads within a bubble enjoy a symmetric relation- ship, as befits a model of conjunction. Further, bub- ble trees support representation of nested coordina- tion, with the scope of shared modifiers identifiable by …

Content truncated to 500 words