Improving the robustness of machine reading comprehension via contrastive learning

Abstract Pre-trained language models achieve high performance on machine reading comprehension task, but these lack robustness and are vulnerable to adversarial samples. Most of the current methods for improving model based data enrichment. However, do not solve problem poor context representation model. We find that plays a key role in model, dense space results robustness. To deal with this, we propose Multi-task Reading Comprehension learning framework via Contrastive Learning. Its main idea is improve encoded by through contrastive learning. This special proposed called Learning Context Representation Space(CLCRS). CLCRS samples sentences containing information from as positive negative samples, expanding distance between answer sentence other context. Therefore, has been expanded. The can better distinguish correct answers misleading sentence. Thus, improved. Experiment datasets show our method exceeds comparison achieves state-of-the-art performance.