Achieving Conversational Goals with Unsupervised Post-hoc Knowledge Injection

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3140 - 3153 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Achieving Conversational Goals with Unsupervised Post-hoc Knowledge Injection Bodhisattwa Prasad Majumder♣ Harsh Jhamtani♦ Taylor Berg-Kirkpatrick♣ Julian McAuley♣ ♣Department of Computer Science and Engineering, UC San Diego {bmajumde, tberg, jmcauley}@eng.ucsd.edu ♦School of Computer Science, Carnegie Mellon University jharsh@cs.cmu.edu Abstract A limitation of current neural dialog mod- els is that they tend to suffer from a lack of specificity and informativeness in generated re- sponses, primarily due to dependence on train- ing data that covers a limited variety of sce- narios and conveys limited knowledge. One way to alleviate this issue is to extract rele- vant knowledge from external sources at de- coding time and incorporate it into the dialog response. In this paper, we propose a post- hoc knowledge-injection technique where we first retrieve a diverse set of relevant knowl- edge snippets conditioned on both the dialog history and an initial response from an exist- ing dialog model. We construct multiple can- didate responses, individually injecting each retrieved snippet into the initial response us- ing a gradient-based decoding method, and then select the final response with an unsu- pervised ranking step. Our experiments in goal-oriented and knowledge-grounded dialog settings demonstrate that human annotators judge the outputs from the proposed method to be more engaging and informative com- pared to responses from prior dialog systems. We further show that knowledge-augmentation promotes success in achieving conversational goals in both experimental settings. 1 Introduction Generic responses which lack specificity have been a major issue in existing dialog models (Hosseini- Asl et al., 2020; Dinan et al., 2019a). The issue in part stems from bottlenecks in dialog models due to a limited scope of scenarios and access to limited knowledge available during training. On the other hand, encoding all possible world knowl- edge at training time is not feasible, and even un- desirable in cases where knowledge sources are dynamically varying (Ghazvininejad et al., 2018; Majumder et al., 2020b; Zhao et al., 2020; Bruyn et al., 2020; Kim et al., 2020; Prabhumoye et al., 2021). One possible approach is to incorporate There are plenty of museums to visit around Cambridge. If you love hiking, you can enjoy the trails alongside the river. Some of my friends like to go the centre of the town and catch a movie. Find me something fun to do around Cambridge area in daytime! You can go for a movie. Is there anything else that your prefer? Many prefer to visit museums. You can do hiking around the river if you love nature. Or you can watch a movie. Which one do you prefer? Retrieved Knowledge Initial Response Final Response Dialog Context 🤖 🤖 Figure 1: Augmenting initial response from an existing di- alog model with relevant external knowledge leads to more engaging and informative responses improving the success in achieving the conversational goal (here, finding a fun activity). relevant knowledge at …

Content truncated to 500 words