Knowledge based neural question generation from text

文本自动生成问题由于其在会话系统、智能辅导系统和阅读理解评估等各种应用中的实用性而受到越来越多的关注。已经开发了两种不同的策略来从文本生成问题：基于规则的方法和神经序列到序列（seq-to-seq）模型。基于规则的方法利用从文本中提取的语言特征，通过人类专家设计的模板规则将句子转化为问题。另外，使用人工神经网络将输入序列（例如，答案）转换为问题的seq-to-seq模型，由于能够从数据中学习复杂的映射函数，最近在问题生成中显示出良好的性能。然而，产生高质量的问题仍然是一个公开的挑战，因为目前的方法产生的问题往往是模糊的、无意义的和缺乏多样性的，因此需要进行重大改进，以便在工业产品中使用。我们将与行业合作伙伴iNAGO公司合作，开发新的技术，通过将领域内和领域外的知识整合到序列到序列的模型中，从而从文本数据中产生高质量的问题，从而改进问题生成。我们将研究语言资源的使用，结合应用领域中为问题生成设计的规则，来指导序列到序列的方法。我们还将研究领域自适应问题，即在一个领域中可用的规则集和/或训练样本集在另一个领域中使用，以解决神经网络方法的有限训练数据问题。该项目将增强iNAGO的Netpeople会话助手平台产品，使其能够自动生成知识库（由问题和答案组成），用于在互联汽车及其相关服务等领域构建会话助手系统。

Automatic generation of questions from text has gained increasing attention due to its usefulness in various applications, such as conversational systems, intelligent tutoring systems, and reading comprehension assessment. Two different strategies have been developed to generate questions from text: rule-based methods and neural sequence-to-sequence (seq-to-seq) models. Rule-based methods employ linguistic features extracted from text to transform sentences into questions through template rules designed by human experts. Alternatively, seq-to-seq models, which convert an input sequence (e.g., an answer) into a question using artificial neural networks, have recently shown good performance in question generation due to their ability to learn complex mapping functions from data. However, generating high-quality questions is still an open challenge since the questions generated by current approaches are often vague, meaningless and lack of diversity, and thus significant improvements are needed in order for them to be used in industrial products.Working with industry partner iNAGO Inc., we will develop novel techniques to improve question generation by incorporating in-domain and out-domain knowledge into seq-to-seq models to produce high-quality questions from text data. We will investigate the use of linguistic resources, in combination with the rules designed for question generation in the application domain, to guide seq-to-seq approaches. We will also investigate the problem of domain adaptation, where the sets of rules and/or training samples are available in one domain are used in another domain to solve the limited training data problem of neural network methods.This project will enhance iNAGO's Netpeople conversational assistant platform product by enabling it to automatically generating the knowledge base (consisting of questions and answers) for building conversational assistant systems in domains such as connected automobiles and their related services.