S-Net: From Answer Extraction to Answer Generation for Machine Reading Comprehension

In this paper, we present a novel approach to machine reading comprehension the MS-MARCO dataset. Unlike the SQuAD dataset that aims to answer a with exact text spans in a passage, the MS-MARCO dataset defines the as answering a question from multiple passages and the words in the answer not necessary in the passages. We therefore develop an-then-synthesis framework to synthesize answers from extraction. Specifically, the answer extraction model is first employed to predict most important sub-spans from the passage as evidence, and the answer model takes the evidence as additional features along with the and passage to further elaborate the final answers. We build the extraction model with state-of-the-art neural networks for single reading comprehension, and propose an additional task of passage to help answer extraction in multiple passages. The answer synthesis is based on the sequence-to-sequence neural networks with extracted as features. Experiments show that our extraction-then-synthesis outperforms state-of-the-art methods.