Implicit Discourse Relation Recognition with Context-aware Character-enhanced Embeddings

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1914–1924, Osaka, Japan, December 11-17 2016. Implicit Discourse Relation Recognition with Context-aware Character-enhanced Embeddings 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 For the task of implicit discourse relation recognition, traditional models utilizing manual fea- tures can suffer from data sparsity problem. Neural models provide a solution with distributed representations, which could encode the latent semantic information, and are suitable for recog- nizing semantic relations between argument pairs. However, conventional vector representations usually adopt embeddings at the word level and cannot well handle the rare word problem without carefully considering morphological information at character level. Moreover, embeddings are assigned to individual words independently, which lacks of the crucial contextual information. This paper proposes a neural model utilizing context-aware character-enhanced embeddings to alleviate the drawbacks of the current word level representation. Our experiments show that the enhanced embeddings work well and the proposed model obtains state-of-the-art results. 1 Introduction It is widely agreed that in a formal text, units including clauses and sentences are not isolated but instead connected logically, semantically, and syntactically. Discourse parsing is a fundamental task in natural language processing (NLP) that analyzes the latent relation structure and discovers those connections across text units. It could benefit various downstream NLP applications such as question answering (Chai and Jin, 2004; Verberne et al., 2007), machine translation (Hardmeier, 2012; Guzm´an et al., 2014), sentiment analysis (Bhatia et al., 2015; Hu et al., 2016b), and automatic summarization (Maskey and Hirschberg, 2005; Murray et al., 2006). For discourse parsing, Penn Discourse Treebank (PDTB) (Prasad et al., 2008) provides the lexically- grounded annotations of discourse relations. Each discourse relation consists of two abstract object arguments and the corresponding sense annotations, which can be roughly characterized according to whether explicit connectives could be drawn from the texts. In Explicit relations, explicit connectives can be found in the texts; when such indicators are not given directly, an inferred connective expression could be inserted, forming Implicit relations. The following two examples describes these two kinds of discourse relations: the former has an explicit connective “so” which reveals the Explicit relation, while in the latter case, an inferred Implicit connective “that is” has to be inserted to express the relation. (1) Arg1: We’re standing in gasoline. Arg2: So don’t smoke. (Contingency.Cause.Result - wsj 0596) (2) Arg1: The ploy worked. Arg2: Implicit=that is The defense won. (Contingency.Cause - wsj 1267) It has been shown that discourse connective is crucial for high-accuracy relation recognition (Pitler et al., 2009; Lin et al., 2014). Compared to explicit discourse relations in which senses between adjacent ∗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), …

Content truncated to 500 words