Coordination Boundary Identification with Similarity and Replaceability

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 264–272, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Coordination Boundary Identification with Similarity and Replaceability Hiroki Teranishi Hiroyuki Shindo Yuji Matsumoto Nara Institute of Science and Technology {teranishi.hiroki.sw5, shindo, matsu}@is.naist.jp Abstract We propose a neural network model for coordination boundary detection. Our method relies on two common properties — similarity and replaceability in con- juncts — in order to detect both sim- ilar and dissimilar pairs of conjuncts. The model improves the identification of clause-level coordination using bidirec- tional recurrent neural networks incorpo- rating two properties as features. We show that our model outperforms existing state- of-the-art methods for the coordination an- notated Penn Treebank and Genia corpus without any syntactic information from parsers. 1 Introduction Coordination is a common structure and one of major ambiguities in human languages. Although coordination gives a large amount of syntactic or semantic information of coordinated phrases, dis- ambiguating coordination still remains one of the difficult problems that state-of-the-art parsers can- not cope with. Given a coordinator word, how can we find conjuncts? Coordinate structures are character- ized by two properties: (1) similar structures of- ten appear in conjuncts, and (2) one conjunct can be replaced with another conjunct without los- ing sentence consistency in syntax or semantics. However, many previous studies of coordination disambiguation rely only on the similarities be- tween conjuncts, despite the fact that similarities are not always helpful (Shimbo and Hara, 2007; Hara et al., 2009; Hanamoto, 2012). For example, the sentence “[at least two commercial versions have been put on the U.S. market], and [an esti- mated 500 have been sold].” does not have sim- ilar phrases around the coordinating conjunction “and.” Thus, existing methods sometimes fail to capture coordination. In addition to the case where there is a lack of similarities, many similarity- based methods use handcrafted features, heuristic rules, or external resources such as thesauri. To overcome these problems, Ficler and Gold- berg (2016) proposed a neural network model with the replaceability feature as well as the similar- ity feature. Their model produces candidate pairs of conjuncts using probabilities assigned by the Berkeley Parser. All candidate pairs are scored on the basis of the similarity, replaceability and parser-derived features, and then the best scored pair is picked. Their approach outperforms ex- isting constituent parsers for the Penn Treebank and similarity-based coordination disambiguation methods such as those by Shimbo and Hara (2007) and Hara et al. (2009) for the Genia treebank. Al- though Ficler and Goldberg’s (2016) method im- proves performance significantly, it heavily de- pends on the syntactic parser. They use the outputs from the parser not only for candidates generation and the feature for scoring, but also for computa- tion of the similarities. The problems of propa- gated errors from the parser and dependencies on external resources still remain in their work. In this work, we propose a neural network model for coordination disambiguation that does not require any external …

Content truncated to 500 words