Reconstructed Option Rereading Network for Opinion Questions Reading Comprehension

Delai Qiu, Liang Bao, Zhixing Tian, Yuanzhe Zhang, Kang Liu, Jun Zhao, Xiangwen Liao

Abstract: Multiple-choice reading comprehension task has seen a recent surge of popularity, aiming at choosing the correct option from candidate options for the question referring to a related passage. Previous work focuses on factoid-based questions but ignore opinion-based questions. Options of opinion-based questions are usually sentiment phrases, such as “Good” or “Bad”. It causes that previous work fail to model the interactive information among passage, question and options, because their approaches are based on the premise that options contain rich semantic information. To this end, we propose a Reconstructed Option Rereading Network (RORN) to tackle it. We first reconstruct the options based on question. Then, the model utilize the reconstructed options to generate the representation of options. Finally, we fed into a max-pooling layer to obtain the ranking score for each opinion. Experiments show that our proposed achieve state-of-art performance on the Chinese opinion questions machine reading comprehension datasets in AI challenger competition.