Multi-hop Reading Comprehension Learning Method Based on Answer Contrastive Learning

Multi-hop reading comprehension generally requires the model to give the answer and complete the prediction of supporting facts. However, previous works mainly focus on the interaction between question and context, and ignore the problem that many entities or short spans in sentences are similar to the true answer, so they do not take advantage of the differentiation information between true and plausible answers. To solve the above problems, we propose a learning method based on answer contrastive learning for multi-hop reading comprehension, which makes full use of answer judgment information to reduce the interference of confusing information to the model. Specifically, similar entity and random span data augmentation methods are proposed firstly from the perspective of answer for contrastive learning. Secondly, we implement multi-task joint learning by combining answer contrastive learning and graph neural network model through a shared encoder, and use several subtasks to mine shared information to assist in answer extraction and supporting fact prediction. Especially, the learning method forces the model to pay more attention to the true answer information through answer contrastive learning, which helps the model distinguish the start and end positions of answers. We validate our proposed learning method on the HotpotQA dataset, and the experimental results show that it performs better than the competitive baselines on several evaluation metrics.