UECA-Prompt: Universal Prompt for Emotion Cause Analysis

Proceedings of the 29th International Conference on Computational Linguistics, pages 7031–7041 October 12–17, 2022. 7031 UECA-Prompt: Universal Prompt for Emotion Cause Analysis Xiaopeng Zheng1, Zhiyue Liu1,2, Zizhen Zhang1, Zhaoyang Wang1, Jiahai Wang1∗ 1School of Computer Science and Engineering, Sun Yat-sen University Guangzhou, China 2School of Computer, Electronics and Information, Guangxi University Nanning, China {zhengxp26,liuzhy93,wangzhaoy22}@mail2.sysu.edu.cn, {wangjiah,zhangzzh7}@mail.sysu.edu.cn Abstract Emotion cause analysis (ECA) aims to extract emotion clauses and find the corresponding cause of the emotion. Existing methods adopt fine-tuning paradigm to solve certain types of ECA tasks. These task-specific methods have a deficiency of universality. And the relations among multiple objectives in one task are not explicitly modeled. Moreover, the relative po- sition information introduced in most exist- ing methods may make the model suffer from dataset bias. To address the first two problems, this paper proposes a universal prompt tuning method to solve different ECA tasks in the uni- fied framework. As for the third problem, this paper designs a directional constraint module and a sequential learning module to ease the bias. Considering the commonalities among different tasks, this paper proposes a cross-task training method to further explore the capa- bility of the model. The experimental results show that our method achieves competitive per- formance on the ECA datasets. 1 Introduction Recently, emotion cause analysis (ECA) has ob- tained increasing attention. As a classic task of ECA, emotion cause extraction (ECE) was first pro- posed and defined as a clause-level classification problem (Gui et al., 2016a). ECE aims to explore the potential causes behind a certain emotional ex- pression in a clause. However, the applications of ECE are limited in real-world scenarios because the emotion must be annotated before cause ex- traction. Therefore, emotion-cause pair extraction (ECPE) was proposed to identify all emotions and their corresponding causes from an unannotated text at the same time (Xia and Ding, 2019). Chen et al. (2020a) proposed conditional causal relation- ships classification (CCRC) to explore the relation of emotion-cause pairs and contexts. Although there are other emerging tasks proposed (Li et al., ∗*Corresponding author Input c1 c2 c5 Clause-level Encoder h1 h2 h5 Prediction Contextual Encoder Pretrained Embedding || PLMs Attention Attention Position Information Document c1 :Yesterday … c2: a policeman … c3: and told him … c4: The old ... c5: and … h1 h2 h5 Fusion Attention ··· ··· ··· Figure 1: Fine-tuning for emotion cause analysis. 2021b; Bi and Liu, 2020) to expand the applications of ECA, this paper focuses on the most complex and challenging three clause-level sub-tasks: ECE, ECPE, and CCRC. In literature, ECA tasks have been widely stud- ied. On the ECE task, Gui et al. (2016a) released a benchmark Chinese emotion cause dataset and proposed a multi-kernel based method. Several sta- tistical learning (Gui et al., 2016b; Xu et al., 2017) and deep learning methods (Cheng et al., 2017; Hu et al., 2021b) have been applied to ECE, which show competitive performance on emotion cause prediction. On the ECPE task, Xia and Ding (2019) proposed a two-step …

Content truncated to 500 words