Tag-Enhanced Tree-Structured Neural Networks for Implicit Discourse Relation Classification

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 496–505, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Tag-Enhanced Tree-Structured Neural Networks for Implicit Discourse Relation Classification Yizhong Wang1 Sujian Li1,2 Jingfeng Yang1 Xu Sun1 Houfeng Wang1,2 1Key Laboratory of Computational Linguistics, Peking University, MOE, China 2Collaborative Innovation Center for Language Ability, Xuzhou, Jiangsu, China {yizhong, lisujian, yjfllpyym, xusun, wanghf}@pku.edu.cn Abstract Identifying implicit discourse relations be- tween text spans is a challenging task be- cause it requires understanding the mean- ing of the text. To tackle this task, re- cent studies have tried several deep learn- ing methods but few of them exploited the syntactic information. In this work, we explore the idea of incorporating syntac- tic parse tree into neural networks. Specif- ically, we employ the Tree-LSTM model and Tree-GRU model, which are based on the tree structure, to encode the argu- ments in a relation. Moreover, we fur- ther leverage the constituent tags to con- trol the semantic composition process in these tree-structured neural networks. Ex- perimental results show that our method achieves state-of-the-art performance on PDTB corpus. 1 Introduction It is widely agreed that text units such as clauses or sentences are usually not isolated. Instead, they correlate with each other to form coherent and meaningful discourse together. To analyze how text is organized, discourse parsing has gained much attention from both the linguistic (Weiss and Wodak, 2007; Tannen, 2012) and computational (Marcu, 1997; Soricut and Marcu, 2003) commu- nities, but the current performance is far from sat- isfactory. The most challenging part is to identify the discourse relations between text spans, espe- cially when the discourse connectives (e.g., “be- cause” and “but”) are not explicitly shown in the text. Due to the absence of such evident linguistic clues, trying to model and understand the mean- ing of the text becomes the key point in identifying such implicit relations. economic performance future NP JJ NN JJ NP measure VB VP (a) (b) to expand likely VP TO VB JJ ADJP is VBZ S the economy DET NN NN VP (Expansion.Restatement.Specification, wsj_0233) Arg2: A figure above 50 indicates the economy is likely to expand Arg1: The index is intended to measure future economic performance Figure 1: An example of two sentences with their discourse relation as Expan- sion.Restatement.Specification. Subfigure (a) and (b) are partial parse trees of the two important phrases with yellow background. Previous studies in this field treat the task of recognizing implicit discourse relations as a clas- sification problem and various techniques in se- mantic modeling have been adopted to encode the arguments in each relation, ranging from tradi- tional feature-based models (Lin et al., 2009; Pitler et al., 2009) to the currently prevailing deep learn- ing methods (Ji and Eisenstein, 2014; Liu and Li, 2016; Qin et al., 2017). Despite of the superior ability of the deep learning models, the syntactic 496 information, which proves to be helpful for iden- tifying discourse relations in many early studies (Subba …

Content truncated to 500 words