Generating Comprehension Questions Using Paraphrase - Technologies and Applications of Artificial Intelligence - page 308

Generating Comprehension Questions Using Paraphrase

Ya-Min Tseng 1 , Yi-Ting Huang 2 , Meng Chang Chen 1 , and Yeali S. Sun 2

1 Institute of Information Science, Academia Sinica, Taipei, Taiwan
{tym,mcc}@iis.sinica.edu.tw
2 Department of Information Management, National Taiwan University, Taipei, Taiwan
{d97725008,sunny}@ntu.edu.tw

Abstract. As online English learning environment becomes more and more
ubiquitous, English as a Foreign Language (EFL) learners have more choices to
learning English. There is thus increasing demand for automatic assessment
tools that help self-motivated learners evaluate their understanding and compre-
hension. Existing question generation systems, however, focus on the sentence-
to-question surface transformation and the questions could be simply answered
by word matching, even without good comprehension. We propose a novel ap-
proach to generating more challenging choices for reading comprehension ques-
tions by combining paraphrase generation with question generation. In the final
evaluation, although there is a slight decrease in the overall quality, our results
outperform the baseline system in challenging score and have a significantly
smaller percentage of statements that remain intact from the sources sentences.

Keywords: question generation, automatic assessment, reading comprehension,
e-learning, multiple choice questions, paraphrase generation, discourse relation.

1 Introduction

Online learning has become a popular choice for English learners. Reading online
news and watching talks, for example, are ways to learning English. There are all
sorts of learning material on the Internet but there are only a limited number of human
quiz creators to provide assessments based on online resources. Automatic assessment
tools could help evaluate whether the readers comprehend the text well. Aware of the
demand, several Question Generation (QG) systems have focused on the generation
of questions for reading comprehension. These work, however, tend to generate sim-
plistic questions with doubtful ability to assess comprehension. The same wording as
the source sentences are applied to the questions, like the question “ what is often
voted as the best treat in Taiwan? ” and its source “ bubble tea is often voted as the
best treats in Taiwan. ” Inevitably, such questions could be solved by searching for the
same word spans in the article, even without good comprehension.
The over-simplicity problem might result from two common characteristics of ex-
isting QG systems. Firstly, the generating approaches have mostly focused on wh -
questions or on question stems in the form of cloze . Answering these questions only
requires a single piece of information, such as a location ( where -question), a person