CTF-ERC: Coarse-to-Fine Reasoning for Emotion Recognition in Conversations

Emotion recognition in conversation (ERC) has always been a research highlight in human-computer interaction (HCI). Existing methods mainly rely on hard-to-access structured data and partial session information, ignoring the global session context and the speaker’s historical experience. To address these issues, we propose a novel multilevel reasoning framework ranging from coarse-grained to fine-grained. Specifically, we employ both coarse-grained and fine-grained reasoning to capture the global information of the session and the speaker’s historical experience. At the stage of the coarse-grained inference, an information enhancement module is designed to alleviate the category imbalance in some datasets. In the fine-grained reasoning, we use information distillation to capture the speaker’s background knowledge and historical experience. In addition, we establish a fusion inference module to further incorporate the information obtained from both coarse-grained and fine-grained reasoning, and perform structured emotion label prediction to capture the sequential dependencies among emotion labels. We conduct extensive experiments on four benchmarks, and the experimental results show that our model exhibits competitive results in the task of ERC.