Enhancing Knowledge Selection for Grounded Dialogues via Document Semantic Graphs

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2810 - 2823 July 10-15, 2022 ©2022 Association for Computational Linguistics Enhanced Knowledge Selection for Grounded Dialogues via Document Semantic Graphs Sha Li1∗, Madhi Namazifar2, Di Jin2, Mohit Bansal2, Heng Ji2, Yang Liu2, Dilek Hakkani-Tur2 1University of Illinois at Urbana-Champaign, 2 Amazon Alexa AI shal2@illinois.edu {madhinam, djinamzn, mobansal, jihj, yangliud, hakkanit}@amazon.com Abstract Providing conversation models with back- ground knowledge has been shown to make open-domain dialogues more informative and engaging. Existing models treat knowledge se- lection as a sentence ranking or classification problem where each sentence is handled indi- vidually, ignoring the internal semantic con- nection among sentences in background doc- ument. In this work, we propose to automati- cally convert the background knowledge doc- uments into document semantic graphs and then perform knowledge selection over such graphs. Our document semantic graphs pre- serve sentence-level information through the use of sentence nodes and provide concept con- nections between sentences. We apply multi- task learning for sentence-level knowledge se- lection and concept-level knowledge selection jointly, and show that it improves sentence- level selection. Our experiments show that our semantic graph based knowledge selec- tion improves over sentence selection base- lines for both the knowledge selection task and the end-to-end response generation task on HollE (Moghe et al., 2018) and improves gen- eralization on unseen topics in WoW (Dinan et al., 2019).1 1 Introduction Natural language generation models have seen great success in their ability to hold open-domain dialogues without the need for manual injection of domain knowledge. However, such models of- ten degenerate to uninteresting and repetitive re- sponses (Holtzman et al., 2020), or hallucinate false knowledge (Roller et al., 2021; Shuster et al., 2021). To avoid such phenomena, one solution ∗Work done as an intern at Amazon Alexa AI. 1See https://www.amazon.science/publica tions/enhanced-knowledge-selection-for-g rounded-dialogues-via-document-semantic- graphsfor an updated paper with information about code and resources. is to provide the conversation model with rel- evant knowledge to guide the response genera- tion (Parthasarathi and Pineau, 2018; Ghazvinine- jad et al., 2018; Dinan et al., 2019). Figure 1 illus- trates such knowledge grounded generation. Relevant knowledge is often presented in the form of documents (Moghe et al., 2018; Zhou et al., 2018b; Ghazvininejad et al., 2018; Dinan et al., 2019; Gopalakrishnan et al., 2019) and the task of identifying the appropriate knowledge snippet for each turn is formulated as a sentence classification or ranking task (Dinan et al., 2019). Although more advanced methods have been proposed by model- ing knowledge as a latent variable (Kim et al., 2020; Chen et al., 2020), or tracking topic shift (Meng et al., 2021), they abide by the setting of sentence- level selection. This setting has two inherent draw- backs: (1) it ignores the semantic connections be- tween sentences and (2) it imposes an artificial constraint over the knowledge boundary. A document is not simply a bag of sentences, in fact, it is the underlying semantic connections and …

Content truncated to 500 words