面向阅读理解的问题自动生成关键技术研究

问题生成近年来引起了研究者广泛的兴趣，其研究成果可以节约教育者的时间和精力、提升智能问答和人机对话等应用系统的性能。本项目面向教育领域的阅读理解，研究段落级无答案问题生成，即给定一个段落文本而无答案信息，研究如何自动生成语义相关、语法通顺的高质量问题。围绕问题生成，将着力探索以下关键技术难题：1) 基于自动文摘、机器阅读理解等数据，利用多任务学习方法来自动识别文本中值得提问的关键句；2) 融入篇章结构信息进行长文档的表示学习，显性编码句子位置索引，在图神经网络模型中引入篇章结构关系和指代信息，以期生成具有一定跨句推理能力的问题；3)提出了一个新颖的多任务学习框架，将复述生成作为子任务来辅助问题生成，以期生成具有多样化表达的问题；4) 提出将常识融入问题生成的网络模型，采用门控注意力机制在编码中融入常识并且利用双重拷贝机制使模型得以拷贝常识中的词项，以期生成具有一定常识推理能力的问题。

In recent years, question generation (QG) attracts wide attention from the NLP researchers, due to the fact that it can save educational experts’ time and labor, and it can boost the performance of question answering and dialog generation systems. This project will make a research on answer-agnostic paragraph-level question generation oriented towards reading comprehension in the educational field. Given a passage but without answer, the system aims to automatically generate high-quality questions that are fluent and semantically related with the passage. We will make an in-depth study on the following key technologies for QG. 1) Based on different datasets available for summarization and machine reading comprehension, we propose a multitask learning method to predict the question-worthy key sentences in a passage. 2) We incorporate discourse structure information to learn the passage representations to generate questions with multi-sentence reasoning, adding sentence position index, and integrating the discourse information and the coreference dependency into the graph neural network. 3) We propose a novel multitask learning framework to produce human-like questions with diverse expressions, by employing paraphrase generation as an auxiliary task to help QG. 4) We propose a novel model to leverage commonsense knowledge for QG to generate human-like questions that can do commonsense reasoning, using a gated attention mechanism to fuse commonsense into the passage encoding, and using a dual copy mechanism to allow the model to copy words from commonsense tuples.