Exploiting Dialogue Act for Knowledge Selection and Response Generation

Abstract

Dialogue act (DA) is the description of the intention or function of a dialogue utterance. In document-grounded dialogue, correctly understanding the dialogue context is crucial for models to select knowledge and inject knowledge into responses. Leveraging dialogue act can help to understand the dialogue context and consequently assist the utilization of document information. In this paper, we propose a novel framework leveraging two different kinds of DAs (model-annotated and human-annotated) for Knowledge Selection (KS) and Response Generation (RG). The framework consists of two modules: the prediction module is trained with multi-task learning and learns to select knowledge and predict the next DA; the generation module uses the selected knowledge and the predicted DA for the RG. Our model achieves new state-of-the-art performance on three public datasets and the results verify that leveraging DA can help KS and RG. Our code and data will be released on github.com.

Tasks

Reproductions