Multi-Round Reasoning Incorporating Prior Knowledge in Conversational Sentiment Analysis
Linying Zhang, Xiao Jing Yang, Qiuxian Chen · 2023
Sentiment analysis in conversation is an important method for human-computer interaction systems. Current models of conversational sentiment analysis lack the ability to integrate sentiment cues. So it causes the models not to understand deep semantics. This paper proposes a model of Multi-round Reasoning Incorporating Prior Knowledge (MRIPK). The model is divided into three parts: knowledge fusion part, multi-round reasoning part and emotion classification part. Firstly, the knowledge fusion part uses a knowledge base and a sentiment lexicon to enrich the semantics of the dialogue. Secondly, the multi-round reasoning part uses a hierarchical self-attention mechanism to obtain context-level and speaker-level global context. It uses the global context to integrate context-related cues in multiple rounds. Finally, the emotion classification part uses softmax to derive the final results of classification. To validate the model, we do some experiments on three datasets MELD, IEMOCAP and EmoryNLP. The F1 values improve by about 2%, 3% and 1% on the three datasets respectively. The accuracy of the model has some improvements.