I am trying to finetune T5 model on Xsum dataset. However, in the generation process, I am facing the hallucination problem. In fact, the model is introducing named entities that are existing in the train dataset or other named entities that are not mentionned in the text to summarize. The generated summary is different from the text content.
- Have you faced the same problem while dealing with Xsum dataset training
- Are there some remedies to this problem (Techniques to control the hallucination problem)