On the Diversity and Limits of Human Explanations

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 206 - 218 July 10-15, 2022 ©2022 Association for Computational Linguistics Knowledge-Grounded Dialogue Generation with a Unified Knowledge Representation Yu Li∗2, Baolin Peng1, Yelong Shen1, Yi Mao1, Lars Liden1, Zhou Yu2, Jianfeng Gao1 1Microsoft Research, Redmond, WA 2Columbia University, New York, NY {bapeng,yeshe,maoyi,laliden,jfgao}@microsoft.com {yl5016, zy2461}@columbia.edu Abstract Knowledge-grounded dialogue systems are challenging to build due to the lack of training data and heterogeneous knowledge sources. Existing systems perform poorly on unseen topics due to limited topics covered in the train- ing data. In addition, it is challenging to gener- alize to the domains that require different types of knowledge sources. To address the above challenges, we present PLUG1, a language model that homogenizes different knowledge sources to a unified knowledge representation for knowledge-grounded dialogue generation tasks. We first retrieve relevant information from heterogeneous knowledge sources (e.g., wiki, dictionary, or knowledge graph); Then the retrieved knowledge is transformed into text and concatenated with dialogue history to feed into the language model for generating responses. PLUG is pre-trained on a large- scale knowledge-grounded dialogue corpus. The empirical evaluation on two benchmarks shows that PLUG generalizes well across dif- ferent knowledge-grounded dialogue tasks. It achieves comparable performance with state- of-the-art methods in the fully-supervised set- ting and significantly outperforms other ap- proaches in zero-shot and few-shot settings. 1 Introduction Recent work has shown that conversational mod- els can be trained in an end-to-end fashion (Gao et al., 2019; Roller et al., 2020; Zhang et al., 2019; Adiwardana et al., 2020). Though such models can generate coherent and natural responses con- sistent with conversation history, there is still a clear gap between conversational AI agents and hu- mans. The primary reason is that existing dialogue systems lack knowledge of the subject and thus cannot deep dive into specific topics with humans. ∗Work was done when Yu Li was interning at MSR 1Pre-trained Language model with a Unified knowledge representation for knowledge-Grounded dialogues. Dataset Knowledge % Topics Open-domain Wizard of Wikipedia articles 0.02% CMU_DoG articles 0.04% Recommendation REDIAL tables 15.0% OPENDIALKG graph 7.5% Table 1: Knowledge representation and topic cover- age statistics of existing knowledge-grounded dialogue datasets. % Topics means the portion of topics or facts in the knowledge database covered by the dataset. In order to better incorporate knowledge into dia- logue, knowledge-grounded dialogue systems have become increasingly popular. Knowledge-grounded dialogue generation aims to generate informative and meaningful responses based on both conversation context and external knowledge sources. Thus far, researchers have col- lected knowledge-grounded dialogues for various tasks using crowdsourcing platforms, for instance, open-domain dialogues (Dinan et al., 2019; Zhou et al., 2018) and conversational recommendation dialogues (Li et al., 2018; Moon et al., 2019; Hay- ati et al., 2020). Workers are asked to base their replies on knowledge from structured knowledge bases (Moon et al., 2019; Tuan et al., 2019) or un- structured documents (Dinan et al., 2019; Zhou et al., 2018; Feng …

Content truncated to 500 words