A Stacking Gated Neural Architecture for Implicit Discourse Relation Classification

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2263–2270, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics A Stacking Gated Neural Architecture for Implicit Discourse Relation Classification Lianhui Qin1,2, Zhisong Zhang1,2, Hai Zhao1,2,∗ 1Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China 2Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China {qinlianhui, zzs2011}@sjtu.edu.cn,zhaohai@cs.sjtu.edu.cn Abstract Discourse parsing is considered as one of the most challenging natural language processing (NLP) tasks. Implicit discourse relation clas- sification is the bottleneck for discourse pars- ing. Without the guide of explicit discourse connectives, the relation of sentence pairs are very hard to be inferred. This paper proposes a stacking neural network model to solve the classification problem in which a convolu- tional neural network (CNN) is utilized for sentence modeling and a collaborative gated neural network (CGNN) is proposed for fea- ture transformation. Our evaluation and com- parisons show that the proposed model outper- forms previous state-of-the-art systems. 1 Introduction As a fundamental task in natural language process- ing (NLP), discourse parsing entails the discovery of the latent relational structure in multi-sentence level analysis. It is also central to many practical tasks such as question answering (Liakata et al., 2013; Jansen et al., 2014), machine translation (Meyer and Popescu-Belis, 2012; Meyer and Webber, 2013) and automatic summarization (Murray et al., 2006; ∗Corresponding author. This paper was partially supported by Cai Yuanpei Program (CSC No. 201304490199 and No. 201304490171), National Natural Science Foundation of China (No. 61170114, No. 61672343 and No. 61272248), National Basic Research Program of China (No. 2013CB329401), Ma- jor Basic Research Program of Shanghai Science and Tech- nology Committee (No. 15JC1400103), Art and Science In- terdisciplinary Funds of Shanghai Jiao Tong University (No. 14JCRZ04), and Key Project of National Society Science Foun- dation of China (No. 15-ZDA041). Yoshida et al., 2014). Discourse parsing is also the shared task of CoNLL 2015 and 2016 (Xue et al., 2015; Xue et al., 2016), and many previous works previous on this task (Qin et al., 2016b; Li et al., 2016; Chen et al., 2015; Wang and Lan, 2016). In a discourse parser, implicit relation recognition has been the bottleneck due to lack of explicit connec- tives (like “because” or “and”) that can be strong indicators for the senses between adjacent clauses (Qin et al., 2016b; Pitler et al., 2009; Lin et al., 2014). This work therefore focuses on implicit re- lation recognition that infers the senses of the dis- course relations within adjacent sentence pairs. Most previous works on PDTB implicit relation recognition only focus on one-versus-others binary classification problems of the top level four classes (Pitler et al., 2009; Zhou et al., 2010; Park and Cardie, 2012; Biran and McKeown, 2013; Ruther- ford and Xue, 2014; Braud and Denis, 2015). Tra- ditional classification methods directly rely on fea- ture engineering, based on bag-of-words, produc- tion rules, and some linguistically-informed fea- tures (Zhou et al., …

Content truncated to 500 words