Enhancing Seq2Seq Models for Role-Oriented Dialogue Summary Generation Through Adaptive Feature Weighting and Dynamic Statistical Conditioninge

Role-oriented dialogue summarization techniques are designed to generate corresponding summaries for different roles in a dialogue. Although existing techniques can handle structured dialogues, it remains a challenge to effectively adapt to role transformation and real-time adjustment of discourse importance in changing dialog scenarios. To this end, we propose a new method called Adaptive Feature Weighting with Dynamic Statistical Conditioning (ADSC), which dynamically adjusts the weights of the roles and discourses in a dialog, and flexibly responds to real-time changes in the dialogue content, especially in terms of role transformation and discourse importance adjustment. In addition, in order to adapt to the input of uncertain number of roles, we introduce Role Prompts to guide the model to generate corresponding summaries according to different roles. We validate the effectiveness of the ADSC method on two publicly available datasets, CSDS and MC, and demonstrate the advantages of combining the method with the Seq2Seq model. The experimental results show that the ADSC method significantly improves the performance of these models on the dialog summarization task, and especially performs best when combined with the BART model.