Joint Alignment of Multi-Task Feature and Label Spaces for Emotion Cause Pair Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 6955–6965 October 12–17, 2022. 6955 Joint Alignment of Multi-Task Feature and Label Spaces for Emotion Cause Pair Extraction Shunjie Chen1, Xiaochuan Shi1, Jingye Li1, Shengqiong Wu2, Hao Fei2, Fei Li1†, Donghong Ji1 1Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University, Wuhan, China 2School of Computing, National University of Singapore, Singapore {shunjiechen, shixiaochuan, theodorelee, lifei_csnlp, dhji}@whu.edu.cn swu@u.nus.edu, haofei37@nus.edu.sg Abstract Emotion cause pair extraction (ECPE), as one of the derived subtasks of emotion cause anal- ysis (ECA), shares rich inter-related features with emotion extraction (EE) and cause ex- traction (CE). Therefore EE and CE are fre- quently utilized as auxiliary tasks for better fea- ture learning, modeled via multi-task learning (MTL) framework by prior works to achieve state-of-the-art (SoTA) ECPE results. How- ever, existing MTL-based methods either fail to simultaneously model the specific features and the interactive feature in between, or suffer from the inconsistency of label prediction. In this work, we consider addressing the above challenges for improving ECPE by performing two alignment mechanisms with a novel A2Net model. We first propose a feature-task align- ment to explicitly model the specific emotion- &cause-specific features and the shared inter- active feature. Besides, an inter-task alignment is implemented, in which the label distance between the ECPE and the combinations of EE&CE are learned to be narrowed for better label consistency. Evaluations of benchmarks show that our methods outperform current best- performing systems on all ECA subtasks. Fur- ther analysis proves the importance of our pro- posed alignment mechanisms for the task.1 1 Introduction Emotion cause analysis (ECA), detecting potential causes for certain emotion expressions in a doc- ument, has been a hot research topic in natural language processing (NLP) community (Lee et al., 2010; Gui et al., 2016; Fan et al., 2019). ECA has derived three associated tasks: EE, CE and ECPE. As illustrated in Figure 1(a), EE and CE detects the emotion and cause clauses respectively, while ECPE identifies both the emotion and cause clauses †Corresponding author 1Our code is available at https://github.com/ csj199813/A2Net_ECPE (b) Shared Method (c) Parallel Method (d) Feature-Task Alignment (e) Inter-Task Alignment (a) Example Interaction Feature Cause-Specific Feature ECPE EE CE ECPE EE CE ECPE EE CE ECPE EE CE Prediction ≈ [c1] If one person is rescued, [c2] everyone will be very happy, [c3] and they will eat better at night. [c4] If no one is rescued, [c5] the heart will be very sad. EE: [c2, c5] CE: [c1, c4] ECPE: [(c2, c1), (c5, c4)] Emotion-Specific Feature Figure 1: (a) illustrates three subtasks of ECA. (b) and (c) depicts the shared and parallel features encoding method, respectively. In (d) and (e) we show our pro- posed feature-task alignment mechanism and inter-task alignment mechanism, respectively. as well as their semantic relation. By jointly model- ing the clauses detection and the relational pairing, ECPE effectively relieves the noise introduction in the pipeline process, and thus receives most …

Content truncated to 500 words