Key points are not available for this paper at this time.
Recently neural response generation models have leveraged large pre-trained models and knowledge snippets to generate relevant and informative. However, this does not guarantee that generated responses are correct. In this paper, we examine factual correctness in-grounded neural response generation models. We present a human setup to identify three different response types: responses that are consistent with respect to the input knowledge, responses that hallucinated knowledge, and non-verifiable chitchat style responses. We this setup to annotate responses generated using different stateof-the-art, knowledge snippets, and decoding strategies. In addition, to facilitate development of a factual consistency detector, we automatically create a corpus called Conv-FEVER that is adapted from the Wizard of Wikipedia and includes factually consistent and inconsistent responses. We the benefit of our Conv-FEVER dataset by showing that the models on this data perform reasonably well to detect factually inconsistent with respect to the provided knowledge through evaluation on our annotated data. We will release the Conv-FEVER dataset and the human responses.
Santhanam et al. (Mon,) studied this question.