Learning to generate complex question with intent prediction from long passage.

Generating questions from the long passage is an important and challenging task. Most of the recent works focus on generating questions whose answers are consecutive text spans in the given passage. However, realistic questions are more complicated and their answers are always inductive and summative. In this paper, we focus on a complex form of question generation task, in which the answer is implied in the long passage. It means that we cannot directly find sentences relevant to the question in the passage anymore. To this end, we first construct a dataset that meets our needs on top of RACE. Based on this, we propose an Intent-aware Complex Question Generation model (ICQG). It first encodes the long passage, which exploits a gated mechanism to fetch the valuable information for elaborating the question. And then, both the passage and answer are used to support the question decoding by modeling their interaction. Finally, an intent classifier is designed to predict what kinds of questions tend to be asked, which is used to guide the question decoding. We conduct both qualitative and quantitative evaluations, and the experimental results demonstrate that the proposed model is effective on this task and superior to the competitor methods. [ABSTRACT FROM AUTHOR]