Study on Deep Hierarchical Neural Model Within the RAG Framework for Multi-Turn Response Generation

The contextual information of multi-turn dialogue tasks is always overlooked in the modeling process, and the historical information of system responses are greatly challenged. In this paper, we explore the necessity of topic coherence in multi-turn conversations, and propose a deep hierarchical neural model within the RAG framework, modeling topic information of the historical utterances. Comparing with the baseline RAG model, our work achieves promising performance in terms of topic coherence for multiturn response generation.