Topic Tensor Network for Implicit Discourse Relation Recognition in Chinese

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 608–618 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 608 Topic Tensor Network for Implicit Discourse Relation Recognition in Chinese Sheng Xu, Peifeng Li, Fang Kong, Qiaoming Zhu and Guodong Zhou Institute of Artificial Intelligence, School of Computer Science and Technology, Soochow University, China sxu@stu.suda.edu.cn;{pfli,kongfang,qmzhu,gdzhou}@suda.edu.cn Abstract In the literature, most of the previous studies on English implicit discourse relation recog- nition only use sentence-level representations, which cannot provide enough semantic infor- mation in Chinese due to its unique paratac- tic characteristics. In this paper, we propose a topic tensor network to recognize Chinese im- plicit discourse relations with both sentence- level and topic-level representations. In par- ticular, besides encoding arguments (discourse units) using a gated convolutional network to obtain sentence-level representations, we train a simplified topic model to infer the latent topic-level representations. Moreover, we feed the two pairs of representations to two fac- tored tensor networks, respectively, to capture both the sentence-level interactions and topic- level relevance using multi-slice tensors. Ex- perimentation on CDTB, a Chinese discourse corpus, shows that our proposed model sig- nificantly outperforms several state-of-the-art baselines in both micro and macro F1-scores. 1 Introduction As a critical component of discourse parsing, dis- course relation recognition focuses on determining how two adjacent discourse units (e.g., clauses, sentences, and sentence groups), called argu- ments, semantically connect to one another. Ob- viously, identifying discourse relations can help many downstream NLP applications, such as auto- matic summarization, information extraction and question answering. In principle, the discourse connectives between two arguments are important for recognizing the relationship between them. For explicit dis- course relation recognition where the discourse connectives explicitly exist in the text, a simple frequency-based mapping table can achieve high performance due to the critical role of a connective in determining the discourse relations (Xue et al., 2016). For implicit discourse relation recognition, it is much more challenging due to missing an ex- act connective and it normally depends on the un- derstanding of the whole text (Pitler et al., 2009). This paper focuses on recognizing implicit dis- course relations in Chinese. In contrast to English, which is a hypotactic language (formal cohesion), Chinese is a paratactic language (semantic cohe- sion) that tends to pro-drop clause connectives. Our statistics indicate that the implicit relations in the Chinese CDTB corpus account for 75.2%, while the proportion in the English PDTB corpus declines to only 40%. Hence, recognizing implicit discourse relations in Chinese becomes more cru- cial than in English. In the literature, most of previous studies fo- cused on English, with only a few on Chinese. Compared with traditional feature-based methods (Pitler et al., 2009; Lin et al., 2009; Wang et al., 2017; Kong and Zhou, 2017) that directly rely on feature engineering, recent neural network models (Liu et al., 2017; Qin et al., 2017; Guo et al., 2018; Bai and Zhao, 2018) can capture deeper semantic cues and learn better representations …

Content truncated to 500 words