A Neural Network for Coordination Boundary Prediction

Abstract

We propose a neural-network based model for coordination boundary prediction. The network is designed to incorporate two signals: the similarity between conjuncts and the observation that replacing the whole coordination phrase with a conjunct tends to produce a coherent sentences. The modeling makes use of several LSTM networks. The model is trained solely on conjunction annotations in a Treebank, without using external resources. We show improvements on predicting coordination boundaries on the PTB compared to two state-of-the-art parsers; as well as improvement over previous coordination boundary prediction systems on the Genia corpus.

Hierarchical Coordinate Structure Analysis for Japanese Statutory Sentences Using Neural Language Models

“…Finally, as is the case with our research, Ficler and Goldberg (2016) also proposed a method based on neural networks; however, it takes a different approach from our NLMs. Specifically, this method has three main differences from ours.…”

“…As is well known, a number of methods have been developed for domain-independent coordination analysis. Among them, many methods, especially the older ones, focus on identifying only two conjuncts adjacent to a coordinator (Agarwal and Boggess 1992;Resnik 1999;Goldberg 1999;Chantree et al 2005;Dubey et al 2006;Ficler and Goldberg 2016). Since coordinate structures in Japanese statutory sentences often have three or more conjuncts, 1 we need to identify all of them.…”

“…Our method identifies a conjunct based on two assumptions: (1) the conjuncts in any pair are alike and (2) the fluency of a sentence is maintained even if the conjuncts in a coordinate structure are swapped with each other. The first assumption has also been used in previous works for coordination analysis (Hara et al 2009;Ficler and Goldberg 2016;Huddleston and Pullum 2002). Instead of our second assumption, some previous works (Ficler and Goldberg 2016;Huddleston and Pullum 2002) used another assumption, i.e., that a coordinate structure can be replaced by any of its conjuncts.…”

“…The first assumption has also been used in previous works for coordination analysis (Hara et al 2009;Ficler and Goldberg 2016;Huddleston and Pullum 2002). Instead of our second assumption, some previous works (Ficler and Goldberg 2016;Huddleston and Pullum 2002) used another assumption, i.e., that a coordinate structure can be replaced by any of its conjuncts. However, this assumption on replacement may not work well for an incomplete coordinate structure, which our method generates while identifying coordinate structures with more than two conjuncts.…”

Decomposed Local Models for Coordinate Structure Parsing

“…However, the shortcoming of their work is that it is highly dependent on the external parser. The work of Teranishi et al (2017) developed an end-to-end model, as opposed to the pipeline approach of Ficler and Goldberg (2016b). They also used similarity and replaceability feature representations without information from a syntactic parser.…”