Retrieval-Free Knowledge-Grounded Dialogue Response Generation with Adapters

To diversify and enrich generated dialogue responses, knowledge-grounded dialogue has been investigated in recent years. The existing methods tackle the knowledge grounding challenge by retrieving the relevant sentences over a large corpus and augmenting the dialogues with explicit extra information. Despite their success, however, the existing works have drawbacks on the inference efficiency. This paper proposes KnowExpert, an end-to-end framework to bypass the explicit retrieval process and inject knowledge into the pre-trained language models with lightweight adapters and adapt to the knowledge-grounded dialogue task. To the best of our knowledge, this is the first attempt to tackle this challenge without retrieval in this task under an open-domain chit-chat scenario. The experimental results show that KknowExpert performs comparably with some retrieval-based baselines while being time-efficient in inference, demonstrating the potential of our proposed direction.