QDSE: Question-Driven Semantic Enhancement for Logical Reasoning

In recent years, pretrained language models have demonstrated outstanding performance across various reading comprehension datasets. However, the logical reasoning task demands a higher level of logical reasoning ability. In many existing research methods, there is a tendency to concentrate on uncovering the logical structure between context and options while overlooking the crucial role of ‘questions.’ This paper proposes Question-Driven Semantic Enhancement (QDSE) to augment the semantic expression of context and options by utilizing questions as filters. Furthermore, our method can serve as available auxiliary modules for other methods based on token embedding. Experiments indicate that our method enhances the semantic information of token embedding, improving the performance of the baseline model.