TSAM: A Two-Stream Attention Model for Causal Emotion Entailment

Proceedings of the 29th International Conference on Computational Linguistics, pages 6762–6772 October 12–17, 2022. 6762 TSAM: A Two-Stream Attention Model for Causal Emotion Entailment Duzhen Zhang∗, Zhen Yang, Fandong Meng, Xiuyi Chen†, Jie Zhou Pattern Recognition Center, WeChat AI, Tencent Inc, Beijing, China {bladedancer957,hugheren.chan}@gmail.com {zieenyang,fandongmeng,withtomzhou}@tencent.com Abstract Causal Emotion Entailment (CEE) aims to discover the potential causes behind an emo- tion in a conversational utterance. Previous works formalize CEE as independent utter- ance pair classification problems, with emo- tion and speaker information neglected. From a new perspective, this paper considers CEE in a joint framework. We classify multiple utterances synchronously to capture the cor- relations between utterances in a global view and propose a Two-Stream Attention Model (TSAM) to effectively model the speaker’s emotional influences in the conversational his- tory. Specifically, the TSAM comprises three modules: Emotion Attention Network (EAN), Speaker Attention Network (SAN), and inter- action module. The EAN and SAN incor- porate emotion and speaker information in parallel, and the subsequent interaction mod- ule effectively interchanges relevant informa- tion between the EAN and SAN via a mu- tual BiAffine transformation. Extensive ex- perimental results demonstrate that our model achieves new State-Of-The-Art (SOTA) perfor- mance and outperforms baselines remarkably. 1 Introduction With the recent proliferation of open conversational data on social media platforms, such as Twitter and Facebook, Emotion Analysis in Conversations (EAC) has become a popular research topic in the field of Natural Language Processing (NLP). Most of the existing works on EAC mainly focus on Emo- tion Recognition in Conversations (ERC), i.e., rec- ognizing emotion labels of utterances (e.g., happy, sad, etc.) (Poria et al., 2017, 2019b; Wang et al., 2020; Zhang et al., 2020). However, Poria et al. (2021) point out that these studies lack further rea- soning about emotions, such as understanding the ∗This work was done when Duzhen Zhang was interning at Pattern Recognition Center, WeChat AI, Tencent Inc, China. †Corresponding author. 1, SA 2, SB You know I am getting married! Wow! That’s great news. Who is the lucky person? excited happy 1, SA 2, SB I like winter. Me too. happy neutral 3, SA It’s snowing heavily. What about taking a walk? happy 4, SB That’s a good idea. Let’s go! happy 5, SA happy What a heavy snow! Look! The water is frozen. a) b) Figure 1: Example conversations sampled from the benchmark dataset (Poria et al., 2021) stimuli and the cause of the emotion. Since Rec- ognizing Emotion Cause in Conversations (REC- CON) holds the potential to improve the inter- pretability and performance of affect-based mod- els, Poria et al. (2021) put forward a new promising task, named RECCON, which includes two differ- ent sub-tasks: Causal Span Extraction (CSE) at word/phrase level and Causal Emotion Entailment (CEE) at utterance level. Due to the simplicity and sufficiency describing emotion causes at the utter- ance level, we focus on the CEE task in this paper, whose goal is to predict which particular utterances in the conversational history contain the cause of non-neutral emotion …

Content truncated to 500 words