Encoding Paraphrase Information Generated by ChatGPT and Introducing Interrogative Word Control Mechanism for Question Generation

Answer-aware question generation means generating a relevant question based on a passage of text and an answer, where the answer is a subsequence within the text. In fact, different expressions of the same thing can affect human understanding, and the same holds true for models. Therefore, in this paper, we use paraphrase information to mimic different perspectives on the same fact and propose a new question generation model. Specially, we utilize ChatGPT to obtain paraphrase of the context. What sets our approach apart from previous work is that we directly model both the context and its paraphrase information using a shared encoder, enabling the model to capture semantic information from different perspectives and fully understand the context. Furthermore, we propose a new self-guided conditional layer normalization module (SCLN) that introduces interrogative words as conditional information into the encoder for the first time. This module adaptively maps interrogative word information to context information, enhancing the model's understanding of the context and controlling the interrogative word that generate question, thereby enhancing the answerability of the question. We conduct experiments on the SQuAD dataset and the NewsQA dataset. The experimental results show that our proposed model achieves significant performance improvements compared to several baseline models. Further human evaluation demonstrates that the questions generated by our model can comprehensively capture contextual information to improve question answerability.