Algorithm Optimization of Emotional Polarity Integrating Prompt Learning and Virtual Adversarial Training

Abstract:
The research on emotional polarity takes the customer service of Meituan as the scene. The research on negative emotions is often used as an important test for the service attitude of the agents. Through online monitoring or offline quality inspection, the behavior of the agents can be regulated. Based on the BERT model, this paper proposes two algorithm optimizations for emotional polarity. On the one hand, by constructing a Prompt to verify its classification effect in the case of a small sample of emotional polarity, you can set a Prompt template on the sentence, and then pass the model predicts the emotional answer, and finally converts the predicted answer into a label for sentiment polarity classification through the mapping relationship. On the other hand, in order to improve the generalization ability of the model, we added virtual adversarial training and established two models, one model was trained with labeled data, and the other was trained with unlabeled data.