Multi-level Emotion Cause Analysis by Multi-head Attention Based Multi-task Learning

Multi-level Emotion Cause Analysis by Multi-head Attention Based Multi-task Learning Xiangju Li, Shi Feng, Yifei Zhang, Daling Wang∗ School of Computer Science and Engineering, Northeastern University No.195, Chuangxin Road, Hunnan District, Shenyang, China, 110207 lixiangju100@163.com, fengshi@cse.neu.edu.cn, zhangyifei@cse.neu.edu.cn, wangdaling@cse.neu.edu.cn Abstract Emotion cause analysis (ECA) aims to identify the potential causes behind certain emotions in text. Lots of ECA models have been designed to extract the emotion cause at the clause level. However, in many scenarios, only extracting the cause clause is ambiguous. To ease the problem, in this paper, we introduce multi-level emotion cause analysis, which focuses on identifying emotion cause clause (ECC) and emotion cause keywords (ECK) simultaneously. ECK is a more challenging task since it not only requires capturing the specific understanding of the role of each word in the clause but also the relation between each word and emotion expression. We observe that ECK task can incorporate the contextual information from the ECC task, while ECC task can be improved by learning the correlation between emotion cause keywords and emotion from the ECK task. To fulfill the goal of joint learning, we propose a multi-head attention based multi-task learning method which utilizes a series of mechanisms including shared and private feature extractor, multi-head attention, emotion attention and label embedding to capture features and correlations between the two tasks. Experimental results show that the proposed method consistently outperforms the state-of-the-art methods on a benchmark emotion cause dataset. 1 Introduction Emotion cause analysis (ECA), a new field in emotion analysis, attempts to comprehend a given text, and then extracts potential causes that lead to emotion expressions in the text. There has been an increasing interest in the research community on ECA more recently since it is widely used in many scenarios. For example, restaurants are eager to find out why people like or dislike their food or services from users’ comments or reviews. Similarly, instead of gauging public opinions towards policies or political issues just using frequency counts, governments would like to further know the triggering factors of certain attitudes expressed online. ECA is a challenging emotion analysis task since it requires a comprehensive understanding of natural languages and the ability to do further inference. Restricted by the lack of annotated corpora, early studies used rule-based methods (Chen et al., 2010; Lee et al., 2010) and crowd-sourcing methods (Russo et al., 2011) to tackle this task. Until recently, Gui et al. (Gui et al., 2016) released a reasonable ECA corpus, based on which they developed the first deep learning model for the task (Gui et al., 2017), and by following that, various other ECA approaches were proposed and achieved superior results (Li et al., 2018; Li et al., 2019; Ding et al., 2019; Xia et al., 2019; Fan et al., 2019). Example 1 [x1] : Entertainment reporter Jucy interviewed that LAM Raymond and Xinyue Zhang to get married in 2020. [x2] : Also, [x3] : she interviewed a piece of explosive news that the wedding ceremony of Tina Tang will …

Content truncated to 500 words