Target-guided dialog generation with dynamic knowledge path by commonsense knowledge graph and relation prediction

초록

Conversational AI has been rapidly advancing with the development of large language models and has shown excellent performance. However, one of its limitations is a passive system that cannot ask or guide users back to ambiguous questions. To overcome this, we have implemented an active dialog system that can smoothly transition from previous conversations. Our system is a target-guided system, which means it can guide the conversation by asking the user to provide a desired response or target. This approach is knowledge-rich and challenging, as it requires achieving the target while maintaining contextual consistency. To generate responses, we dynamically construct knowledge paths through knowledge graphs and relation predictors. These play an essential role in generating diverse and logically connected responses. To achieve this, we follow a global planning method that systematically conducts conversations with a target, and constructs knowledge paths based on common sense. We perform multi-hop reasoning and bi-directional search simultaneously to increase diversity and logical connectivity. We have overcome the limitations of existing works that rely solely on knowledge graphs by reflecting the results of relation predictors along with each object's WIKI data in the path. Therefore, the consideration of the knowledge graph and the performance of the relation predictor, compared to the existing system, in completing the dynamic knowledge path and generating transition responses allowed conversations to transition more naturally. We have verified the proposed model through experiments.

키워드