Multi-Stage Prompting for Knowledgeable Dialogue Generation

Findings of the Association for Computational Linguistics: ACL 2022, pages 1317 - 1337 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Multi-Stage Prompting for Knowledgeable Dialogue Generation Zihan Liu‡∗, Mostofa Patwary§, Ryan Prenger§, Shrimai Prabhumoye§, Wei Ping§, Mohammad Shoeybi§, Bryan Catanzaro§ ‡The Hong Kong University of Science and Technology, §NVIDIA zihan.liu@connect.ust.hk, mpatwary@nvidia.com Abstract Existing knowledge-grounded dialogue sys- tems typically use finetuned versions of a pre- trained language model (LM) and large-scale knowledge bases. These models typically fail to generalize on topics outside of the knowl- edge base, and require maintaining separate potentially large checkpoints each time fine- tuning is needed. In this paper, we aim to address these limitations by leveraging the in- herent knowledge stored in the pretrained LM as well as its powerful generation ability. We propose a multi-stage prompting approach to generate knowledgeable responses from a sin- gle pretrained LM. We first prompt the LM to generate knowledge based on the dialogue context. Then, we further prompt it to gener- ate responses based on the dialogue context and the previously generated knowledge. Re- sults show that our knowledge generator outper- forms the state-of-the-art retrieval-based model by 5.8% when combining knowledge relevance and correctness. In addition, our multi-stage prompting outperforms the finetuning-based dialogue model in terms of response knowl- edgeability and engagement by up to 10% and 5%, respectively. Furthermore, we scale our model up to 530 billion parameters and show that larger LMs improve the generation correct- ness score by up to 10%, and response rele- vance, knowledgeability and engagement by up to 10%. Our code is available at: https:// github.com/NVIDIA/Megatron-LM. 1 Introduction Dialogue systems face the problem of producing bland and generic outputs that are devoid of con- tent (Wolf et al., 2019; Holtzman et al., 2019; Ma et al., 2020). Recent efforts have been made to address these concerns by grounding dialogue re- sponses on a source of knowledge (Dinan et al., ∗ This work was done when the first author was an in- tern at NVIDIA. Corresponding authors: Zihan Liu, Mostofa Patwary. Figure 1: Our proposed framework (MSDP) for the knowledgeable dialogue generation. 2018; Zhou et al., 2018; Zhao et al., 2019; San- thanam et al., 2020; Prabhumoye et al., 2021). Therefore, building a knowledgeable dialogue sys- tem has become one of the key milestone tasks in conversational research. Current knowledge-grounded dialogue systems highly rely on a massive external knowledge cor- pus for a retrieval model to obtain relevant knowl- edge (Dinan et al., 2018; Kim et al., 2019; Zhao et al., 2020), which inevitably brings several limita- tions. First, retrieval systems are constrained to the size and domains of the database, and they cannot generalize to out-of-domain topics that are not cov- ered by the database. Second, retrieval from a mas- sive corpus takes substantial resources. Reimers and Gurevych (2021) show that it is more difficult for the state-of-the-art retrieval model (Karpukhin et al., 2020) to retrieve relevant knowledge when the size of the database increases. The larger database increases the chance that …

Content truncated to 500 words