P-Reader: A Clue-Inspired Model for Machine Reading Comprehension

With the widespread use of web applications, a large amount of textual content is generated on the Internet continuously. In order to analyze and mine the information contained in the text, machine reading comprehension (MRC) is receiving increasingly more attention. As an important technique, MRC can boost the business value of Internet applications. Traditional MRC models use an extractive approach, resulting in poor performance on unanswerable questions. To address this problem, we propose a clue-inspired MRC model. Specifically, we mimic the human reading comprehension process through a combination of sketchy and intensive reading. Experimental results show that the proposed model achieves better performance on several public datasets, especially for unanswerable questions.