ECPE-2D: Emotion-Cause Pair Extraction based on Joint Two-Dimensional Representation, Interaction and Prediction

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3161–3170 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3161 ECPE-2D: Emotion-Cause Pair Extraction based on Joint Two-Dimensional Representation, Interaction and Prediction Zixiang Ding, Rui Xia∗, Jianfei Yu School of Computer Science and Engineering, Nanjing University of Science and Technology, China {dingzixiang,rxia,jfyu}@njust.edu.cn Abstract In recent years, a new interesting task, called emotion-cause pair extraction (ECPE), has e- merged in the area of text emotion analysis. It aims at extracting the potential pairs of e- motions and their corresponding causes in a document. To solve this task, the existing re- search employed a two-step framework, which first extracts individual emotion set and cause set, and then pair the corresponding emotion- s and causes. However, such a pipeline of t- wo steps contains some inherent flaws: 1) the modeling does not aim at extracting the final emotion-cause pair directly; 2) the errors from the first step will affect the performance of the second step. To address these shortcom- ings, in this paper we propose a new end-to- end approach, called ECPE-Two-Dimensional (ECPE-2D), to represent the emotion-cause pairs by a 2D representation scheme. A 2D transformer module and two variants, window- constrained and cross-road 2D transformers, are further proposed to model the interaction- s of different emotion-cause pairs. The 2D representation, interaction, and prediction are integrated into a joint framework. In addi- tion to the advantages of joint modeling, the experimental results on the benchmark emo- tion cause corpus show that our approach im- proves the F1 score of the state-of-the-art from 61.28% to 68.89%. 1 Introduction Emotion cause extraction (ECE), as a sub-task of emotion analysis, aims at extracting the potential causes of certain emotion expressions in text. The ECE task was first proposed by Lee et al. (2010) and defined as a word-level sequence labeling prob- lem. Gui et al. (2016a) released a new corpus and re-formalized the ECE task as a clause-level ex- traction problem. Given an emotion annotation, ∗Corresponding author the goal of ECE is to predict for each clause in a document if the clause is an emotion cause. This framework has received much attention in the fol- lowing studies in this direction. Although the ECE task was well defined, it has two problems: Firstly, the emotion must be annotated manually before cause extraction, which greatly limits its practical application; Secondly, the way to first annotate the emotion and then extract the causes ignores the fact that emotions and causes are mutually indicative. To address this problem, we have proposed a new task named emotion-cause pair extraction (ECPE), aiming to extract the potential pairs of emotion- s and their corresponding causes together in our previous work (Xia and Ding, 2019). Specifically, ECPE is defined as a fine-grained emotion analysis task, where the goal is to extract a set of valid emotion-cause pairs, given a document consisting of multiple clauses as the input. Figure 1 (a) shows an example of the ECPE task. The …

Content truncated to 500 words