Hallucination refers to generating text that is not 'faithful' to the source. In most cases, the hallucination occurs due to divergence between the source and reference. For example, in the context of image captioning, hallucination can be defined as generating captions that contain descriptions not present in the given image.
Also, hallucination has been observed when the system catches on to wrong correlations between different training data parts. In general higher predictive uncertainty corresponds to a higher chance of hallucination and epistemic uncertainty is more indicative of hallucination than aleatoric or total uncertainties. This is the case as the common theme across all the hallucination explanations in conditional NLG tasks is predictive uncertainty.
For example in the case of image captioning, it was shown that at higher uncertainty levels the generated objects are more likely to be hallucinated and this correlation is consistent across data-to-text generation, abstractive summarization and neural machine translation (NMT) as well.