Fusion of Topic Models for Conversational Emotion Recognition in Graph Neural Networks

Emotion recognition in conversations（ERC） aims to predict emotional categories of utterances within a conversation. Presently, graph neural network-based ERC methods predominantly employ fixed hyperparameters to determine the connections among graph edges, lacking adaptive strategies for edge construction tailored to diverse data and ignoring thematic relationships between statements. Furthermore, during the training process of graph neural networks, these methods often utilize a summation superposition approach to aggregate node information, limiting the model’s non-linear capabilities. To address these limitations, this paper integrated topic modeling with graph neural networks and proposed a novel edge construction method. Firstly, a topic model was employed to extract the thematic distribution of statements within a conversation, followed by the connection of statements sharing same themes. Meanwhile, the SwiGLU gated unit was introduced to regulate the flow of information between layers in the graph neural network. Considering differences in character information, the edge types were carefully tailored to better capture intrinsic and extrinsic factors influencing emotional changes. Through extensive experiments conducted on four publicly available datasets（IEMOCAP, MELD, EmoryNLP, DailyDialogue）, our approach demonstrates significant improvements over advanced ERC methods, achieving F1 score enhancements of 1. 69%, 0. 27%, and 0. 38% on the first three datasets, respectively. Moreover, our adaptive method exhibits a 2. 11% improvement on long conversations, surpassing the 0. 8% gain on short conversations. The introduction of the SwiGLU unit effectively mitigates over-smoothing phenomena in the graph neural network. Consequently, the proposed approach, which combines adaptive graph construction with topic modeling and integrates SwiGLU gated units into graph neural networks, proves to be effective in enhancing dialogue emotion recognition, thereby reinforcing the model’s robustness.