Interpersonal relationships between speakers play an important role in multi-party dialogue. The relationship between the responder and the addressee can significantly influence the topics, communication styles, and speaker's persona in response. Howe...
Interpersonal relationships between speakers play an important role in multi-party dialogue. The relationship between the responder and the addressee can significantly influence the topics, communication styles, and speaker's persona in response. However, previous works don't consider the importance of relationships between speakers.
To address this, we propose MIRROR (Multi-party dialogue generation based on Interpersonal Relationship-awaRe persOna Retrieval). First, We model the dialogue structure and interpersonal relationships by inferring discourse relations and extracting interpersonal dimensions from each utterance. Then, we integrate each information into heterogeneous graphs. Second, we retrieve the personas of the responder and the addressee which are relevant to the interpersonal relationship and dialogue history.
Experiments on two datasets, Multilight and HLA-Chat++, show that our method outperforms the baselines and effectively improves persona consistency and relationship appropriateness in generated responses.