Informative and Coherent: Plug-and-Play Incorrect Knowledge Rejection for Knowledge-Grounded Dialog Generation

Knowledge-grounded dialog generation aims to generate informative responses with the help of external knowledge. Selecting the right knowledge is crucial for the quality of responses. Most of the existing knowledge selection methods focus on how to accurately choose either the top-1 or top-k from a given knowledge pool. However, they rarely consider the situation that the selected knowledge is incorrect, which possibly happens due to the knowledge pool being retrieved from search engines. In this work, we first design a pre-experiment to investigate the effects of incorrect knowledge, suggesting that it is harmful to the coherence and informativeness of dialog. Even without knowledge outperforms using the incorrect one. Inspired by this, we aim to improve the quality of responses via rejecting incorrect knowledge. Specially, we propose a benchmark task for incorrect knowledge rejection (IKR) that checks whether the selected knowledge is needed or not. We then design a multi-level interaction network for IKR. In the experiments, the effects of IKR are extensively discussed through both automatic and human evaluation. The results demonstrate the superiority of IKR for informative and coherent dialog generation.