Decomposed Local Models for Coordinate Structure Parsing

Proceedings of NAACL-HLT 2019, pages 3394–3403 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3394 Decomposed Local Models for Coordinate Structure Parsing Hiroki Teranishi1 Hiroyuki Shindo1 Yuji Matsumoto1,2 1Nara Institute of Science and Technology 2RIKEN Center for Advanced Intelligence Project (AIP) {teranishi.hiroki.sw5, shindo, matsu}@is.naist.jp Abstract We propose a simple and accurate model for coordination boundary identification. Our model decomposes the task into three sub- tasks during training; finding a coordinator, identifying inside boundaries of a pair of con- juncts, and selecting outside boundaries of it. For inference, we make use of probabilities of coordinators and conjuncts in the CKY parsing to find the optimal combination of coordinate structures. Experimental results demonstrate that our model achieves state-of-the-art results, ensuring that the global structure of coordina- tions is consistent. 1 Introduction Coordination is a frequently occurring structure that consists of conjuncts joined by a coordina- tor word. Since conjunct spans are one of the major ambiguities, identifying them is difficult, even for humans. For instance, in the sentence “Toshiba’s line of portables, for example, features the T-1000, which is in the same weight class but is much slower and has less memory, and the T-1600, which also uses a 286 microprocessor, but which weighs almost twice as much and is three times the size,” we cannot find correct conjuncts for each coordinator at a glance. The presence of coordina- tion makes a sentence more ambiguous and longer, resulting in errors in syntactic parsing. To identify the conjuncts of a given coordina- tor, previous studies have explored two properties of coordinate structures: (1) similarity – conjuncts tend to be similar; (2) replaceability – conjuncts can be replaced. Ficler and Goldberg (2016b) combine the syntactic parser and neural networks to compute the similarity and replaceability fea- tures of conjuncts. Teranishi et al. (2017) also exploit the two properties without deploying any syntactic parser, and achieve state-of-the-art re- sults. Although both approaches outperform the similarity-based approaches (Shimbo and Hara, 2007; Hara et al., 2009), they cannot handle more than two conjuncts in a coordination, and multiple coordinations in a sentence at one time. Hence, their systems may produce coordinations that con- flict with each other. In contrast, Hara et al. (2009) define production rules for coordination in order to output consistent coordinate structures. Here, we propose a new framework for coordi- nation boundary identification. We generalize a scoring function that takes a pair of spans with a coordinator and returns a higher score when the two spans appear to be coordinated. Using this function in the CKY parsing with produc- tion rules for coordination, our system produces globally consistent coordinations in a given sen- tence. To obtain such a function, we decompose the task into three independent subtasks – find- ing a coordinator, identifying the inner boundaries of a pair of conjuncts and delineating its outer boundaries. We use three different neural net- works for the tasks, and the networks are trained on the basis of their local decisions. Our …

Content truncated to 500 words