An Emotion Recognition Method in Conversations Based on Knowledge Selection and Fuzzy Fingerprints

In the context of large-scale pre-trained language models such as BERT or RoBERTa, they exhibit outstanding performance in natural language processing tasks, but they suffer from a lack of interpretability and the potential introduction of irrelevant information when incorporating external knowledge into sentiment recognition tasks (ERC). This paper presents an innovative ERC sentiment analysis model that combines large-scale pre-trained language models and knowledge pruning techniques while introducing a fuzzy fingerprint module. The workflow of this model consists of several steps: firstly, through the knowledge pruning module, irrelevant external knowledge that does not align with the sentiment of words is removed to enhance the accuracy of sentiment analysis. Next, each utterance in the dialogue as well as the previous dialogue is transformed into contextual embedding representations. Finally, these representations are fed into the fuzzy fingerprint classification module for sentiment recognition. Experimental results demonstrate that this model outperforms other models significantly in sentiment classification tasks.