Enhancing Emotion Recognition in Textual Conversation by Leveraging Common-sense

Abstract

A core aspect of human-like artificial intelligence is the awareness of emotions. Recognizing emotions in conversations (ERC) is especially difficult to solve because of several challenges like contextual modeling, interlocutor profiling, recognizing emotion shifts, and multiparty conversations. Analyzing the results from the state-of-the-art techniques reveal that ERC models could be improved by using common-sense knowledge. Once incorporated correctly, the advancements in general common-sense knowledge models could be leveraged directly in ERC models. Furthermore, insights from these experiments will be applicable in other language tasks as well. In this thesis, I propose two approaches for incorporating common-sense knowledge in ERC models: first, implicit incorporation by fine-tuning language models using the common-sense inferences for the given data, and second, explicit incorporation by applying cross-attention on common-sense knowledge for each utterance. I demonstrate that the proposed methods perform well on IEMOCAP, a widely used dyadic conversation dataset with human-annotated emotions.