A Multi-perspective Co-matching model for Multi-choice reading comprehension
DUAN Jianyong,WEI Xiaopeng,WANG Hao
(School of Information, North China University of Technology, Beijing 100144, China)
(CNONIX National Standard Application and Promotion Laboratory, North China University of Technology, Beijing 100144, China)
[Objective] This paper proposes a model of Multi-choice reading comprehension, and explore the influence of the question type and answer length on the machine reading comprehension.
[Methods] The multi-perspective matching mechanism is used to obtain the correlation between the article, the question and the candidate answer, and then the correlation is used to multiply the article to obtain the vector representation of the question and the candidate answer. Next we extract sentence-level and document-level features, and finally select the correct answer based on these features. Categorize the data based on the question type and answer length, then test their accuracy and analyze their impact on the machine's choice of correct answers.
[Result] The accuracy of this model on the RACE-M, RACE-H and RACE datasets reached 72.5%, 63.1% and 66.1% respectively.
[Limitations] The multi-perspective matching mechanism contains four matching strategies and multiple angles, which makes the model consume a lot of memory and time at the interactive layer.
[Conclusions] The multi-perspective matching mechanism can better interact the article with the question and answer. The accuracy of the model is more affected by the type of question, not by the length of the answer.