GE-Blender: Graph Based Knowledge Enhancement for Blender

Abstract

In spite of the considerable accomplishments in open-domain dialogue generation, the presence of unseen entities can substantially impact the task, resulting in a performance decline in model-driven dialogue generation. While previous research has utilized retrieved knowledge of seen entities to augment model representation, a comprehensive exploration of unseen entities, including their potential co-occurrence and semantic associations within the entity category, remains largely uncharted. In this study, we propose an approach to address this challenge. We construct a graph by extracting entity nodes, enhancing the contextual representation of unseen entities by leveraging the 1-hop surrounding nodes of associated entities. Furthermore, we introduce a named entity tag prediction task to address the issue of unseen entities not present in the constructed graph. We conduct our experiments on an open dataset Wizard of Wikipedia and the empirical results indicate that our approach outperforms the state-of-the-art approaches on Wizard of Wikipedia.