Balancing role contributions: a novel approach for role-oriented dialogue summarization
Zheng Ren · 2024
The task of role-oriented dialog summarization aims at generating appropriate summaries for the different roles in a dialogue. Although existing approaches are able to handle the dialogue content of roles, the extent of the contribution of different roles in a topic has not been analyzed in depth. In this paper, we propose a new model, RoleAdapt, which aims to balance the contributions of each role. RoleAdapt determines the type of summaries that need to be generated by recognizing the Prompt, efficiently identifies the content that contributes more to the topic, and dynamically adjusts the token-level vector representation, which is followed by a decoder that generates the summaries for each role. This approach not only evaluates the contribution of each actor, but also ensures the integrity of the contextual content. We validate RoleAdapt on two publicly available benchmark datasets, CSDS and MC. Experimental results show that the model achieves significant progress on both datasets. Further analysis shows that balancing role contributions is crucial for generating correctly structured summaries.