Global-Local Modeling with Prompt-Based Knowledge Enhancement for Emotion Inference in Conversation

Findings of the Association for Computational Linguistics: EACL 2023, pages 2120–2127 May 2-6, 2023 ©2023 Association for Computational Linguistics Global-Local Modeling with Prompt-Based Knowledge Enhancement for Emotion Inference in Conversation Renxi Wang and Shi Feng School of Computer Science and Engineering Northeastern University Shenyang, China realreasonwang@gmail.com, fengshi@cse.neu.edu.cn Abstract The ability to recognize emotions in conversa- tions is necessary and important for the online chatbot to do tasks such as empathetic response generation and emotional support. Present re- searches mainly focus on recognizing emotions through a speaker’s utterance, while research on emotion inference predicts emotions of ad- dressees through previous utterances. Because of the lack of the addressee’s utterance, emo- tion inference is more challenging than emo- tion recognition. In this paper, we propose a global-local modeling method based on recur- rent neural networks (RNN) and pre-trained language models (PLM) to do emotion infer- ence, which utilizes the sequence modeling ability of RNNs and abundant knowledge from PLMs. Moreover, we take the whole dialogue history as input of PLM to generate knowledge by in-context learning. Experimental results show that our model with knowledge enhance- ment achieves state-of-the-art performance on all three datasets.1 1 Introduction The task of emotion recognition in conversation (ERC) (Poria et al., 2019b) aims to identify emo- tion labels of an utterance, where the whole dia- logue history along with the current utterance is given. However, in emotion inference in conver- sation (EIC) the current utterance is lacking but the dialogue history and the current addressee are known (Li et al., 2021a). For example, Figure 1 shows a conversation between A and B. In the third turn, ERC detects A’s emotion using all available information while EIC predicts A’s emotion using all the information except the last utterance. ERC is a popular task that has been explored widely and deeply, while EIC is a new task that measures the emotion understanding ability of models from a different perspective. 1The code is available at https://github.com/ Reason-Wang/DialogueGLP. Person A Person B [neutral]: Hi, what are you doing here? [happy]: I am celebrating my child’s birthday. Would you like to join us? [happy]: Thank you! It’s so nice of you to invite me. Figure 1: A conversation between A and B. The text in orange might be helpful for ERC. The text in green might be helpful for EIC. In ERC, some previous works utilize sequence- based neural networks to model the context and speaking parties (Majumder et al., 2019; Hu et al., 2021a; Li et al., 2021a). These approaches first finetune a model on utterances to classify emo- tions. Then this model is used to extract features of utterances. As shown in Figure 1, B makes A happy because he/she is inviting A to join in a party. However, B feels happy because he/she is celebrating his/her child’s birthday. The finetun- ing tends to keep the semantics that is helpful to classify the current emotion and the feature extrac- tion compresses the utterance’s information, which may cause information loss that …

Content truncated to 500 words