面向高可靠的开放域问答研究

In recent years, due to the explosive growth of Internet data, the rapid improvement of hardware computing ability and the great progress of natural language processing and deep learning technology, intelligent question-and-answer method has made rapid progress in recent years. With the popularity of smart phones and the wide application of various intelligent assistants, the open-domain question answering system draws close to the people's daily life even more, and the demand is gradually increasing. Therefore, it has become an exigent task to build a reliable and secure system. In fact, there are several key challenges in open-domain question answering: Firstly, the forms and intentions of the questions vary significantly across different questions. Secondly, the answer to the question is of great uncertainty as many of the question often have no answers. Thirdly, it requires the answers able to be interpreted especially for those high-risk questions. In this proposal, we focus on the following three research questions: deep semantic question understanding, robust answer generation, and visual evidence inference and interpretation. In this way, we aim to establish the basic theory and key technologies for a highly reliable Open-domain question answering system.

近年来，由于互联网数据的爆发式增长、硬件计算能力的飞速提高以及自然语言处理和深度学习技术的长足进步，智能问答方法在近年来取得了突飞猛进的发展。随着智能手机的普及以及各类智能助手的广泛应用，开放域的智能问答系统也越来越贴近人们的生活，人们对于问答系统的依赖性也逐渐增强，因此构建安全可靠的问答系统是一个迫切需要解决的问题。事实上，在开放域问答中存在着几个关键特性：首先，问题的形式多变且意图具有多样性；其次，问题答案存在不确定性；再次，高风险问题答案的可解释性。这些特性为高可靠问答系统的构建带来严峻的挑战。针对该问题，本项目围绕开放域问答中深层次的问题语义理解、鲁棒性的问题答案生成、以及可视化的证据推断解释三个方面展开研究，旨在建立面向高可靠的开放域问答成体系的基础理论和关键技术。

针对开放域问答系统在实际应用中面临的问题形式多变与意图多样、问题答案不确定以及答案缺乏可解释的挑战，本项目致力于面向高可靠的开放域问答技术研究，实现意图理解、模型构建以及答案生成三个层面的技术突破，解决开放域问答在问题意图深度理解、匹配模型鲁棒构建以及答案可解释生成上的关键科学问题。本项目从如下四个方面展开研究工作，并取得了一系列原创性的理论与技术成果，具体包括：（1）基于对比式的查询意图理解与表示学习：为了支撑面向复杂检索意图的查询理解与表示，我们主要研究了对比式查询意图生成、基于语义代表性的预训练检索模型以及基于描述的实体分布语义表示；（2）鲁棒可靠的问答匹配模型构建：为了保障问答匹配模型的可靠性，我们主要研究了神经检索模型训练中的标注偏差纠正、基于标签分布增强的极大似然估计方法以及阅读理解模型的持续领域适配；（3）可解释的模型理解与答案生成：为了支撑模型的可解释理解和答案的可解释生成，我们主要研究了相关性匹配模型的深度分析与理解、基于对抗攻击的神经检索模型脆弱性分析、开放域问答系统中的答案可信性研究，以及面向事实验证的生成式证据检索；除了上述的研究成果外，我们还将项目中形成的理论方法成果整合发布成了开源深度文本匹配工具MatchZoo，同时还集成了当前信息检索领域最佳的文本匹配模型，验证了模型的有效性和可用性。