On Controlling Fallback Responses for Grounded Dialogue Generation

Findings of the Association for Computational Linguistics: ACL 2022, pages 2591 - 2601 May 22-27, 2022 c⃝2022 Association for Computational Linguistics On Controlling Fallback Responses for Grounded Dialogue Generation Hongyuan Lu, Wai Lam, Hong Cheng, Helen M. Meng The Chinese University of Hong Kong {hylu,wlam,hcheng,hmmeng}@se.cuhk.edu.hk Abstract Dialogue agents can leverage external textual knowledge to generate responses of a higher quality. To our best knowledge, most existing works on knowledge grounded dialogue set- tings assume that the user intention is always answerable. Unfortunately, this is impractical as there is no guarantee that the knowledge re- trievers could always retrieve the desired knowl- edge. Therefore, this is crucial to incorporate fallback responses to respond to unanswerable contexts appropriately while responding to the answerable contexts in an informative manner. We propose a novel framework that automat- ically generates a control token with the gen- erator to bias the succeeding response towards informativeness for answerable contexts and fallback for unanswerable contexts in an end- to-end manner. Since no existing knowledge grounded dialogue dataset considers this aim, we augment the existing dataset with unanswer- able contexts to conduct our experiments. Au- tomatic and human evaluation results indicate that naively incorporating fallback responses with controlled text generation still hurts infor- mativeness for answerable context. In contrast, our proposed framework effectively mitigates this problem while still appropriately present- ing fallback responses to unanswerable con- texts. Such a framework also reduces the extra burden of the additional classifier and the over- heads introduced in the previous works, which operates in a pipeline manner.1 1 Introduction Building knowledge grounded dialogue agents has been an important research line (Bordes et al., 2016; Young et al., 2017; Zhou et al., 2018; Chaud- huri et al., 2019; Moon et al., 2019; Dziri et al., 2021). Such incorporation of real-world knowledge (Young et al., 2017; Zhou et al., 2018) gives rise 1Related resources can be found at https://github. com/HongyuanLuke/OCFR. Figure 1: An illustrated example for answerable and unanswerable context conditioned on retrieved knowl- edge, along with corresponding desired and undesired responses. We demonstrate an easy single-turn conver- sation for simplicity. Better viewed in colour. to consistent, informative and engaging response generation. Unfortunately, even with a high-quality knowledge retriever, there is no guarantee that the desired knowledge can always be retrieved. There is indeed even no guarantee for the existence of the desired knowledge in the knowledge database. Hence, presenting fallback responses is an essen- tial ability for grounded dialogue agents. We make use of the notion of answerability that represents whether a dialogue context is answerable or not conditioned on the knowledge retrieved. Figure 1 depicts an example to illustrate the importance of answerability in grounded dialogue response gen- eration. As in the unanswerable dialogue context, a fallback response is desirable. Conversely, as in the answerable dialogue context, the response should be as informative as possible. Although the concept of answerability has been well explored in other NLP areas such as Question Answering (Rajpurkar et al., 2018), it is underex- 2591 plored in …

Content truncated to 500 words