Task-oriented dialogue agents are built to assist users in completing various tasks. Generating appropriate responses for satisfactory task completion is the ultimate goal. Hence, as a convenient and straightforward way, metrics such as success rate, inform rate etc., have been widely leveraged to evaluate the generated responses. However, beyond task completion, there are several other factors that largely affect user satisfaction, which remain under-explored. In this work, we focus on analyzing different agent behavior patterns that lead to higher user satisfaction scores. Based on the findings, we design a neural response generation model EnRG. It naturally combines the power of pre-trained GPT-2 in response semantic modeling and the merit of dual attention in making use of the external crowd-sourced knowledge. Equipped with two gates via explicit dialogue act modeling, it effectively controls the usage of external knowledge sources in the form of both text and image. We conduct extensive experiments. Both automatic and human evaluation results demonstrate that, beyond comparable task completion, our proposed method manages to generate responses gaining higher user satisfaction.
