Reference Knowledgeable Network for Machine Reading Comprehension

Multi-choice Machine Reading Comprehension (MRC) as a challenge requires models to select the most appropriate answer from set of candidates with given passage and question. Most existing researches focus on modeling specific tasks or complex networks, without explicitly referring relevant credible external knowledge sources, which are supposed greatly make up for deficiency passage. Thus we propose novel reference-based enhancement model called xmlns:xlink="http://www.w3.org/1999/xlink">Reference xmlns:xlink="http://www.w3.org/1999/xlink">Knowledgeable xmlns:xlink="http://www.w3.org/1999/xlink">Network (RekNet) , simulates human reading strategies refine critical information quote explicit in necessity. In detail, xmlns:xlink="http://www.w3.org/1999/xlink">RekNet refines fine-grained defines it xmlns:xlink="http://www.w3.org/1999/xlink">Reference Span then quotes quadruples by co-occurrence candidates. The proposed is evaluated three multi-choice MRC benchmarks: RACE, DREAM Cosmos QA, obtaining consistent remarkable performance improvement observable statistical significance level over strong baselines. Our code available at https://github.com/Yilin1111/RekNet .