Scientific Reading Comprehension with Sentences Selection and Ranking

Jialei Chen, Weihua Wang, Shuai Shao

Abstract: Scientific Machine Reading Comprehension (SMRC) aims to understand scientific long text by providing answers for the given questions. Most existing methods trend to answer the question using Transformer-based models. However, in the scientific domain, the original text is longer than the general domain. In this paper, we proposed a model that consists of a content retrieval module and a pre-trained model module. The content retrieval module finds the most semantically relevant sentences from the text and re-rank them. The seleted sentences and question will be input into the pre-trained model to get the answers. This model could overcome the length limitation of Transformer model length while achieving impressive results. Our model achieved 0.45 score of RougeL, resulting in the second place in the NLPCC2023 Shared Task2.