Multi-Level Causal Reasoning for Emotion Recognition in Conversations

Emotion recognition in conversation (ERC) has always been a research highlight in human-computer interaction (HCI). Existing models, including large-scale language models (LLMs), exhibit their advantages in capturing semantic correlations in utterance embeddings. However, they do not consider the causal relationship between utterances, which can be directly located to the cause and avoid unnecessary noise interference. Causal relationships are powerful in seizing the generation and evolution of emotions, and are also important for understanding the spread and changes of emotions in dialogue. To address this limitation, we propose a novel Multi-Level Causal Reasoning Model (MLCRM), which explores how to distinguish different causal relationships of a simulation embedding by independent conditioning. Specifically, we put forward a novel causal graph architecture to deal with unstructured conversation data, aiming to reveal causal links between sentences in a conversation. We not only focus on causal relations at the local level, but also on semantic embedding analysis at the global level. We commit to achieving comprehensive capture of causal relationships in conversation data by devising a coarse-grained semantic embedding module to assist the fine-grained causal inference. Furthermore, we raise a mix of experts module to enhance the exchange of information between the coarse-grained and fine-grained modules and to ensure the effective fusion of information for both. We conduct extensive experiments on four benchmarks, and the experimental results show that our model exhibits competitive results in the task of ERC.