智能问答(Question Answering)是自然语言处理领域中一个比较热门的方向，它旨在理解用户由自然语言表达的问题并给出准确合理的答案。而基于知识的智能问答则是在理解和回答问题的同时还要检索和参考相关的知识。根据知识的组织形式可以将其进一步划分为基于结构化知识和非结构化知识的问答。随着信息化速度的加快以及5G网络的推广，每天都会有海量的信息产生，如何通过问答的方式让机器快速准确地给出我们所需要的知识，是基于知识的智能问答所要解决的关键问题。所以它不仅在学术界有很大的研究价值，还对工业界有着非常重要的影响，比如智能客服、对话系统亦或是搜索引擎等。虽然深度学习的快速发展以及数据的激增让智能问答取得了较高的性能，甚至在一些特定的数据集上超越了人类的水平，但是想要达到人类的语言理解和思考能力，仍然有很多的问题与挑战值得我们来进一步探索：（1）如何有效地用图建模表格等结构化的知识，特别是解决复杂关系的上下文表示与相对孤立节点的信息更新；（2）如何在开放域多篇章的场景下快速筛选到所需要精读的段落并解决不同问题类型对答案选择的影响；（3）如何从多个角度高效地去考虑对话历史的深层信息；（4）如何融合结构化与非结构化的知识来回答那些需要先验知识才能回答的问题，以及如何融合多种网络结构来获得更综合的信息。所以针对以上的这些问题，本文将从结构化知识的表格问答(Table-Based Question Answering, TBQA)和非结构化知识的文档问答(Document-Based Question Answering, DBQA)入手，针对单轮与多轮的问答交互形式，对基于知识的智能问答进行进一步的探究。本论文的研究内容与贡献如下：针对结构化知识的问答，即复杂多表知识的问答任务，我们提出了交互流图神经网络，将循环神经网络作为聚合函数来获得同级兄弟节点的上下文关系并加快度较少节点的更新。此外，父子节点的信息与不同层的推理信息也参与到图的交互中。针对非结构化知识的单轮问答，即开放域多篇章阅读理解任务，我们提出了一个整合流水线和模型置信度的系统方法。它根据相应的问题类型选择不同的子模型，并通过高层次的指针网络有效地选择段落进行后续的阅读理解。针对非结构化知识的多轮问答，即对话式阅读理解任务，我们提出了多视角卷积立方体模型。它可以有效地建模对话历史间潜在的语义信息以及并行地处理不同角度的信息。针对结合结构化知识与非结构化知识的问答，即融入外部知识的阅读理解任务，我们提出了带有知识的深宽交互网络。它通过知识注意力机制融入结构化知识图谱中的信息，并通过多通道融合局部与全局的信息。

Question Answering is a popular direction in the field of natural language processing. It aims to understand the questions expressed by users in natural language and give accurate and reasonable answers. And knowledge-based intelligent question answering aims to retrieve relevant knowledge while understanding and answering questions. According to the organizational form of knowledge, it can be further divided into structured and unstructured knowledge question answering. With the acceleration of informatization and the promotion of 5G networks, a large amount of information is produced every day. How to make the machine quickly and accurately give the knowledge we need through question answering is the key problem to be solved by knowledge-based intelligent question answering. Therefore, it not only has great research value in academia, but also has a very important influence on industry, such as intelligent customer service, dialogue systems, or search engines. Although the rapid development of deep learning and the proliferation of data greatly improved the performance of the intelligent question answering system, even surpassing the level of humans in some specific datasets, there are still many problems and challenges to be explored to reach the ability of human language understanding: (1) How to model structured knowledge, like tables, with graph effectively, especially to solve the contextual representation of complex relationships and information updates of relatively isolated nodes; (2) How to quickly select the required intensive paragraphs in the open-domain multi-passage scenario and solve the impact of different types of questions on the choice of answers; (3) How to efficiently consider the deep information of dialogue history from multiple perspectives; (4) How to integrate structured and unstructured knowledge to answer questions that require prior knowledge, and how to integrate multiple network structures to obtain more comprehensive information. So for these issues, this thesis will focus on Table-Based Question Answering (TBQA) and Document-Based Question Answering (DBQA) in the interactive form of single round and multi round, to further explore the knowledge-based intelligent question answering. The main contents and contributions of this thesis are as follows: For the question answering of structured knowledge, that is, the question answering task of complex multi-table knowledge, we propose cross flow graph neural network to obtain the context relationship of sibling nodes at the same level through RNN aggregation functions and accelerate the update of isolated nodes. In addition, the information flow of parent-child nodes in the same layer and the reasoning flow between different layers are also considered. For the single-round question answering of unstructured knowledge, that is, the task of open-domain and multi-passage reading comprehension, we propose a systematic approach to integrate pipeline and model confidence. It uses different sub models according to the corresponding question types, and effectively selects paragraphs for subsequent reading comprehension through a high-level pointer network. For the multi-round question answering of unstructured knowledge, that is, the task of conversational machine reading comprehension, we propose a multi-perspective convolutional cube. The model can efficiently model the latent semantic information between conversation histories and process information from different perspectives in parallel. For the question answering of structured knowledge and unstructured knowledge, that is, the reading comprehension task that incorporates external knowledge, we propose a deep and wide interactive network with knowledge. It integrates the information in the structured knowledge graph through the knowledge attention mechanism, and fuses local and global information through multiple channels.