How to Plant Trees in Language Models: Data and Architectural Effects on the Emergence of Syntactic Inductive Biases

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1118–1131 July 9-14, 2023 ©2023 Association for Computational Linguistics Joint Constrained Learning with Boundary-adjusting for Emotion-Cause Pair Extraction Huawen Feng, Junlong Liu, Junhao Zheng, Haibin Chen, Xichen Shang, Qianli Ma∗ School of Computer Science and Engineering, South China University of Technology, Guangzhou, China 541119578@qq.com, qianlima@scut.edu.cn Abstract Emotion-Cause Pair Extraction (ECPE) aims to identify the document’s emotion clauses and corresponding cause clauses. Like other rela- tion extraction tasks, ECPE is closely associ- ated with the relationship between sentences. Recent methods based on Graph Convolutional Networks focus on how to model the multiplex relations between clauses by constructing dif- ferent edges. However, the data of emotions, causes, and pairs are extremely unbalanced, but current methods get their representation using the same graph structure. In this paper, we pro- pose a Joint Constrained Learning framework with Boundary-adjusting for Emotion-Cause Pair Extraction (JCB). Specifically, through constrained learning, we summarize the prior rules existing in the data and force the model to take them into consideration in optimiza- tion, which helps the model learn a better repre- sentation from unbalanced data. Furthermore, we adjust the decision boundary of classifiers according to the relations between subtasks, which have always been ignored. No longer working independently as in the previous frame- work, the classifiers corresponding to three sub- tasks cooperate under the relation constraints. Experimental results show that JCB obtains competitive results compared with state-of-the- art methods and prove its robustness on unbal- anced data. 1 Introduction Emotion cause analysis aims to capture causal rela- tionships between human emotions and their corre- sponding causes, which has drawn extensive schol- arly attention in recent years (Russo et al., 2011; Neviarouskaya and Aono, 2013; Ghazi et al., 2015; Gui et al., 2018). Emotion cause extraction (ECE), first proposed by Lee et al. (2010), is a branch of emotion analysis tasks. ECE aims at extracting potential causes for given emotions. However, it *Corresponding author requires emotions to be marked first, which limits the applications in real-world scenarios. Hence, Emotion-Cause Pair Extraction (ECPE) (Xia and Ding, 2019) aims to extract all potential pairs of emotions and corresponding causes simultane- ously. Early methods for ECPE are two-stage mod- els (Xia and Ding, 2019), which predict emotions and causes first and then filter out wrong pairs from all possible pairs. Unfortunately, error propagation happens frequently because the predictions in the first stage directly affect the set of possible pairs in the second stage. To this end, the previous work adopts end-to-end frameworks (Ding et al., 2020b; Cheng et al., 2020; Singh et al., 2021) instead of two-stage models. These methods get the represen- tation of emotions and causes separately and then model the pair with them. The distance between the pair of causes is also taken into account because two distant clauses being an emotion-cause pair is usually impossible. With the rapid development of Graph Convo- lutional Networks (Kipf and Welling, 2016; Def- ferrard et al., …

Content truncated to 500 words