Abstract: With a broad area of its applications, the task of emotion recognition in conversation has increasingly attracted attention. The text in the dialogue contains information about the speakers and links closely with the preceding ones, thus a particular word order and structural features are represented by it. Excellent results have been obtained in studies on emotion recognition in conversation using transformer-based pre-training models. However， its traditional classification approaches cannot take into account conversational word order and structural feature. And a mismatch will occur between the downstream task and the pre-trained task. learning can narrow the gap between them by reconstructing downstream tasks. Therefore, the PERC Roberta model is proposed. This model first learns word order and structural features of the dialogue by predicting masked texts and then reconstructs the downstream task through ing learning, thus a richer dialogue knowledge distributed in the pre-training model can be further stimulated. The experiments conducted on two public data, MELD and EmoryNLP, demonstrate the superior performance of the proposed PERC Roberta model. Further, the ablation experimental results also prove the effectiveness of each step in the PERC Roberta model. The code is publicly available on GitHub repository1.

Key words: natural language processing, emotional recognition in conversation, learning