基于无监督知识提取和多关系表示学习的自动问答关键技术研究

Question-Answering（QA) on a large scale of domain-oriented texts is an important discipline within the fields of artificial intelligence and big data application at present. Its main challenges are how to extract the domain knowledge from web-scale natural texts and how to understand the questions released by users. This project aims at building QA systems with the ability of unsupervised knowledge learning and multi-relational question semantical representation. Firstly, by constructing the neural sparse topic model and introducing the prior information of relation mentions, we extract the relations without supervision and generate the knowledge as the candidate answers for QA systems. Secondly, by considering the multi-relational inference path of the questions, we propose a multi-relational translation model TransGP, and then construct an entity annotated word co-occurrence graph to uncover the context information. We therefore devise an end-to-end neural attention network based on bi-directional LSTM to leverage the embeddings of multi-relational questions. Finally, we present an answer retrieval model based on EP-K-D tree to figure out the relational subgraphs of target entities with low computational complexity. On the basis of a shortest path search method MRSP, we hence enrich the answer entities with natural language descriptive sentences. The innovation results of this project can not only effectively reduce the dependency of QA systems on limited knowledge, but also provide users with efficient and accurate QA service, thus it will greatly expand the applications of QA systems.

面向海量领域文本的自动问答技术是当前人工智能和大数据应用的一个重要分支，主要难点是如何提取海量的领域文本知识以及怎样理解用户提出的问题。本项目旨在实现具有无监督知识学习能力和问句多关系语义表征的自动问答技术。主要包括：1）通过构建神经稀疏主题模型，结合关系指称先验信息，无监督地提取实体关系，生成作为自动问答答案备选的知识。2）考虑问句的多关系推断路径，提出不确定性多关系翻译表示模型TransGP，然后结合领域文本构建实体标注词汇共现图，生成问句的上下文信息，设计基于双向LSTM以及注意力机制的神经网络，得到多关系问句的嵌入表示。3）构建基于实体优先级的EP-K-D树的答案搜索模型，结合基于神经网络的最短路径方法MRSP，生成较低计算复杂度的答案实体关系子图，并概括为自然语言描述语句。研究成果可有效减少自动问答对已有知识的依赖，为用户提供高效准确的问答服务，并极大地扩展自动问答的应用领域。

面向海量领域文本的自动问答技术是当前人工智能和大数据应用的一个重要分支，主要难点是如何提取海量的领域文本知识以及怎样理解用户提出的问题。本项目旨在实现具有无监督知识学习能力和问句多关系语义表征的自动问答技术。主要包括：1）通过构建神经稀疏主题模型，结合关系指称先验信息，无监督地提取实体关系，生成作为自动问答答案备选的知识。2）考虑问句的多关系推断路径，提出不确定性多关系翻译表示模型TransGP，然后结合领域文本构建实体标注词汇共现图，生成问句的上下文信息，设计基于双向LSTM以及注意力机制的神经网络，得到多关系问句的嵌入表示。3）构建基于实体优先级的EP-K-D树的答案搜索模型，结合基于神经网络的最短路径方法MRSP，生成较低计算复杂度的答案实体关系子图，并概括为自然语言描述语句。项目已圆满完成研究目标和研究计划，取得一系列软件、论文、专利等研究成果。研究成果可有效减少自动问答对已有知识的依赖，为用户提供高效准确的问答服务，并极大地扩展自动问答的应用领域。