TDCIER: A Topic-Driven and Dialogue-Interactive Emotion Recognition Model for Multi-Speaker Dialogue
Journal of Environmental Science Computer Science and Engineering & Technology · 2024
For the task of multi-speaker dialog emotion recognition, we propose a Topic-Driven and Dialogue-Interactive Emotion Recognition model, TDCIER.The model contains two key modules: Topic Module and Dialogue-Interactive Module.The topic module first learns the word embeddings of the input text using the BERT model, and then combines the syntactic structure information and topic information to represent the sentences, to capture the semantic information and potential derivations.Meanwhile, the topic module also uses a variational autoencoder to extract the topic information of the text, which is used for dialog segmentation to solve the long-range dependency problem in multi-person dialogs.The dialog interaction module simulates the emotional interactions between speakers by constructing a dialog graph, in which nodes represent sentences and edges represent speaker relationships, and graph convolutional networks are used to update node features.Experiments on two datasets, MELD and EmoryNLP, show that the method in this paper achieves optimal results compared with several benchmark models, effectively simulates the dynamic changes of emotions in multiperson conversations, and proves the effectiveness of the model in the task of multiperson conversation emotion recognition.