Generating Comprehension Questions Using Paraphrase - Technologies and Applications of Artificial Intelligence - page 309

( who -question) and time ( when -question). On the other hand, due to the fact that in
reading comprehension quizzes, the article is usually visible when the test takers
attempt to answer the questions, it'd be hard for the automatically generated questions
to reflect their comprehension rather than their test-taking skills. Most work concen-
trate on the surface transformation from declarative sentences to questions and barely
discuss how different the resulting questions would look. While these questions are
helpful in guiding the reading process and testing elementary English learners, the
same might not be for more advanced ones. Self-motivated online learners tend to
have higher English proficiency level, which enables them to learn independently
without subscribing to any material and without human instructors.
Fig. 1. Example question and choices
We approach the problem by developing generating approach for multiple-choice
(non-) factual questions, as Fig. 1. The question form is selected because it's common
in formal reading comprehension tests and it could be the container of different ques-
tion types by casting each question into a statement with its answer. Fig. 1 (A) is
transformed from the what -question that would be generated by many QG systems:
“ what focuses on keeping the body in balance and in harmony with nature? ” along
with its answer choice “ Chinese medicine ”. We decode the task into generating
true/false statements for these choices. By doing so, we could shift our focus from
sentence-to-question transformation to increasing the difficulty of test choices. Our
aim is to generate choices that test deeper knowledge and look different from the
source sentences.
In this work, we present a new approach to generating more challenging choices
for multiple choice questions. The novelty of this work lies in how we design choice
generation and paraphrase generation towards the mutual goal and how to locate the
best-quality choices among numerous variations, nice or erroneous. The Choice Gen-
eration System extracts and rewrites the sentences from the question generation as-
pect. We manually designed transformation rules, which use discourse relations as
trigger, to bind up each generated statement with a specified testing purpose. The
Paraphrase Generation System then moves on to enlarge the superficial difference by
paraphrasing lexically, syntactically and referentially. We merged features from ques-
tion generation and paraphrase generation to train the Acceptability Ranker, which