A New Multi-level Knowledge Retrieval Model for Task-Oriented Dialogue

One of the main challenges in task-oriented dialogue systems is how to retrieve accurate knowledge from external knowledge bases. Existing methods usually retrieve knowledge and entire entity by utilizing dialogue context, while the correlations between dialogue context and entity attributes are overlook, leading suboptimal knowledge retrieval. Therefore, we introduce a Multi-Level knowledge retrieval model for Task-Oriented Dialogue (MLTOD) consisted of an entity retriever, an attribute retriever, a ranker and a response generator. The entity retriever retrieved entities from knowledge bases and the attribute retriever extracts relevant attributes respectively. The ranker dynamically combines the results from the retrievers to select the most relevant knowledge entities. Then the response generator generates final system response based on the ranking result. In addition, this paper introduces a novel multi-level retrieval mechanism. It considers both entity level and attribute level relevance for coarse to fine knowledge retrieval. Experiments on two publicly available datasets show that our MLTOD model outperforms existing state-of-the-art baseline approaches, validating its effectiveness for task-oriented dialogue.