Multi-View Knowledge Retrieval with A New Memory Tracking Mechanism for Task-Oriented Dialogue

Task-oriented dialogue systems have gained considerable attention for their capability to accomplish various tasks that satisfy user requirements using natural language. The essence of the systems is the retrieval of pertinent knowledge from external knowledge bases for generating informative responses. However, the difficulty of accurately retrieving knowledge according to dialogue context frequently leads to the generation of sub-optimal responses. To tackle this issue, this paper proposes a multi-view retriever (MUER) from two different aspects of graph view and edge view to enhance the performance of knowledge retrieval. Specifically, the multi-view retriever includes a graph selector and an edge pruner. According to the dialogue context, the graph selector selects knowledge graphs and the edge pruner prune irrelevant edges within these graphs. For consistent knowledge retrieval, a memory tracking mechanism is proposed to integrate and manage recently accessed knowledge graphs, reducing computations by swiftly retrieving knowledge graphs from historical interactions. Experiment results on three public benchmarks show that our MUER achieves superior performance in terms of both accuracy and consistently. Moreover, the results demonstrate the promising adaptability of the multi-view retriever and the memory tracking mechanism in incorporating into large language models.