Employing the Correspondence of Relations and Connectives to Identify Implicit Discourse Relations via Label Embeddings

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4201–4207 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4201 Employing the Correspondence of Relations and Connectives to Identify Implicit Discourse Relations via Label Embeddings Linh The Nguyen†, Linh Van Ngo†, Khoat Than† and Thien Huu Nguyen‡∗ † Hanoi University of Science and Technology, Hanoi, Vietnam ‡ Department of Computer and Information Science, University of Oregon, USA {linh.nt142585,linhnv,khoattq}@sis.hust.edu.vn,thien@cs.uoregon.edu Abstract It has been shown that implicit connectives can be exploited to improve the performance of the models for implicit discourse relation recogni- tion (IDRR). An important property of the im- plicit connectives is that they can be accurately mapped into the discourse relations conveying their functions. In this work, we explore this property in a multi-task learning framework for IDRR in which the relations and the con- nectives are simultaneously predicted, and the mapping is leveraged to transfer knowledge between the two prediction tasks via the em- beddings of relations and connectives. We pro- pose several techniques to enable such knowl- edge transfer that yield the state-of-the-art per- formance for IDRR on several settings of the benchmark dataset (i.e., the Penn Discourse Treebank dataset). 1 Introduction Discourse parsing reveals the discourse units (i.e., text spans, sentences, clauses) of the documents and how such units are related to each others to improve the coherence. This work focuses on the task of implicit discourse relation recogni- tion (IDRR), aiming to identify the discourse rela- tions (i.e., cause, contrast) between adjacent text spans in documents. IDRR is a fundamental prob- lem in discourse analysis (Knott, 2014; Webber et al., 1999) with important applications on ques- tion answering (Liakata et al., 2013; Jansen et al., 2014) and text summarization (Gerani et al., 2014; Yoshida et al., 2014), to name a few. Due it its importance, IDRR is being studied actively in the literature, leading to the recent advances for this problem based on deep learning (Chen et al., 2016; Qin et al., 2016; Zhang et al., 2016; Lan et al., 2017; Dai and Huang, 2018). ∗Corresponding author. Consider the two following text spans (called arguments) taken from (Qin et al., 2017) as an ex- ample: Argument 1: Never mind. Argument 2: You already know the answer. An IDRR model should be able to recognize that argument 2 is the cause of argument 1 (i.e., the Cause relation) in this case. This is a chal- lenging problem as the models need to rely solely on the text of the arguments to predict accurate discourse relations. The problem would become more manageable if connective/marker cues (i.e., “but”, “so”) are provided to connect the two argu- ments according to their discourse relations (Qin et al., 2017). In the example above, it is bene- ficial for the models to know that “because” can be a connective of the two arguments that is con- sistent with their discourse relation (i.e., Cause). In fact, a human annotator can also benefit from the connectives …

Content truncated to 500 words