Detecting Hallucinated Content in Conditional Neural Sequence Generation
Neural sequence models can generate highly fluent sentences, but recent studies have also shown that they are also prone to hallucinate additional content not supported by the input.
BibTex:
