Knowledge-Driven Multi-dimensional Dialogue Rewriting Model

Abstract

Traditional multiround dialogue systems have problems such as colloquial expression, co-referential resolution, and information default. These problems lead to misunderstandings of human intentions by the system and poor dialogue quality. In order to improve the quality of dialogue, this paper proposes a knowledge-driven entity rewriting model (ERM) at the entity level and proposes a multi-dimensional dialog rewriting model (multi-dimensional dialog rewrite model, MDRM). The model first uses the common sense knowledge graph for entity disambiguation to rewrite the entities in the historical and current dialogue texts to remove the ambiguity spoken expression; Then, it models the historical dialogue, rewrites the current dialogue, and solves the common reference resolution and information default problems of the current dialogue. The comparative experiment shows that the model can effectively improve the quality of dialogue, which verifies the feasibility and effectiveness of the model.