Abstract:Abstract Question answering(Q&A) system is important for accelerating the landing of artificial intelligence. This paper makes an improvement on the Q&A system which uses the method of retrieval-machine reading comprehension (MRC). In the retrieval phase, we use BM25 to recall some documents and split these documents into paragraphs, then we reorder the paragraphs according to the correlation with the question, so as to reduce the number of recalled paragraphs and improve the speed of MRC. In the MRC stage, we design a multi-task MRC structure, which can judge whether the paragraph contains answer and locate answer accurately. Besides, we modify the loss function to fit the sparse labels during the training. The experiments are carried out on multiple data sets to verify the effectiveness of the improved system.

Intelligent Question Answering System Based on Machine Reading Comprehension