External knowledge document retrieval strategy based on intention-guided and meta-learning for task-oriented dialogues

Abstract

The external knowledge augmented task-oriented dialogues system is designed to cover the user requests beyond the predefined data base (DB) and application-specific interface (APIs). However, the existing methods mainly focus to improve the retrieval accuracy of the external knowledge but ignore the modeling of the entity-aware dialogue intention, which can’t fast locate the document-level external API in the practical scenarios. Therefore, an external knowledge document retrieval strategy based on the intention-guided and meta-learning for the task-oriented dialogue has been proposed in this paper. First, the entity-aware dialogue intention information is introduced and then a two-stage training method for the intention-guided knowledge document retrieval model is presented, which is able to effectively enhance the retrieval accuracy of the Bi-encoder retrieval model in the task-based dialogue system. Moreover, to further improve the generalization ability of the retrieval model based on the intention-guided, a meta-task based on the domain offset is constructed and make the Bi-encoder retrieval model get the more robust semantic representation through the meta-learning, which can improve the generalization capability of the proposed model in the invisible domains. Experimental results tested on the augmented MultiWOZ 2.1 dataset are shown to prove the correctness and effectiveness of the proposed strategy.