3R: Reading - Ranking - Recognizing for Multi- Passage Reading Comprehension

multi-passage reading comprehension aims to answer questions about a set of passages. The current models include selecting the most related passage for reading using reinforcement learning or using all passages for reading and then re-ranking answer candidates. But these models can't deal with some noise passages properly. Besides, the answer ranking methods cannot take all valuable information into consideration. So, we propose to select all informative passages, then reading these passages to extract answer candidates, finally re-ranking these answer candidates so that the top ranked answer candidate is the best answer. In the answer re-ranker module, we consider both answer-to-answer verify and question-to-answer verify. Also, we add a no answer recognition section to detect no answer cases.