Answer-Focused and Position-Aware Neural Network for Transfer Learning in Question Generation

Question generation aims to generate natural questions conditioned on answer and the corresponding context pairs, which has the potential value of developing annotated data sets for natural language processing (NLP) researches in education, reading comprehension and question answering. In this paper, focusing on the problem that scarce labeled data of different domains may constrain researchers training their models using supervised or semi-supervised approaches to obtain more semantically similar questions in those areas, we present two transfer learning techniques to further improve a well-performed hybrid question generation model trained on one domain and adapted to another domain. The hybrid question generation model uses an answer-focused and position-aware neural network to improve the model’s ability on source domain data, while the utilization of transfer learning approaches can further enhance the performance of the hybrid model on target domain data. We conduct experiments on two data sets and a significant improvement in the model performance is observed which confirms our hypothesis that a preferable neural network model trained on one domain, attached with transfer learning approaches can further promote the performance on another domain.