Deep learning have recently begun to be applied to natural language processing and generation. As a result, artificial intelligence's natural language generation has advanced significantly. In addition, natural language generation is receiving great a...
Deep learning have recently begun to be applied to natural language processing and generation. As a result, artificial intelligence's natural language generation has advanced significantly. In addition, natural language generation is receiving great attention due to the surprising natural language processing ability of GPT based ChatGPT released by OpenAI. However, Because of hallucinations of natural language generation, users are confused. In the case of GPT's natural language generation task, side effects such as hallucinations occur because it does not perform an information-oriented task. Therefore, it is becoming more important for users not to unconditionally trust text data created through natural language generation, but to distinguish it from facts. In addition, in the case of pre-trained models, there is a possibility of learning with intentionally manipulated data during transfer learning such as fine-tuning. So a method to detect this must be prepared. The purpose of this paper is to find out how contamination of the training data used in the fine-tuning of GPT, a deep learning-based language model for natural language generation(NLG), affects the output. In this experiment, fine-tuning was carried out using contaminated data created by intentionally contaminating data. Using the language model created through fine-tuning, the result was output as a task for generating domain text. In addition, the probability distribution of domain words was analyzed with KL-divergence. As a result, it was confirmed that there was little difference in the word appearance probability distribution in the case of contamination below a certain level, but there was a large difference in the probability distribution in the case of high level contamination.