A cascaded retrieval-while-reasoning multi-document comprehension framework with incremental attention for medical question answering

Clinical Machine Reading Comprehension (MRC) is challenging due to the need for medical expertise and comprehensive reasoning chains for diagnosis. This paper introduces a novel cascade retrieval-while-reasoning framework for clinical MRC that incrementally retrieves and processes multiple supporting documents to effectively address the complexities of answering medical questions. The proposed cascade system is designed in such a way that easier questions are processed by shallower network layers with fewer documents, while more difficult ones are handled by deeper layers with more documents. In the proposed system, a retriever is designed to provide knowledge documents incrementally according to the comprehended difficulty level of each question, which interacts with the reader via query updating for each retrieval and modifies its search direction to better exploit the knowledge bank. To handle the supporting information from incrementally retrieved documents, a progressive attention mechanism is designed to extract cross-document features for better reasoning. The attentional information from multiple supporting documents is then aggregated for the final decision. The proposed method is compared with state-of-the-art models for medical MRC tasks on a large medical QA dataset. Experimental results show that the proposed model effectively combines multiple knowledge documents to solve challenging real-world clinical diagnosis problems. It significantly outperforms the previously best-performing model by 1.84%, reaching an accuracy of 63.00%.