基于大规模知识库的问答系统关键技术研究

As the emerging of very large structured knowledge bases (KBs), question answering over large KB has attracted increasing research efforts from both academia and industrial community. In this proposal, we will address the natural language question answering systems that are based on open-domain structured knowledge bases (KB-QA). Different from traditional information retrieval based question answering, large scale open-domain KBs provide more accurate and easy-access knowledge facts than pure text resources, and their inherent nature of structures can surely improve the capturing of users' query intention, handling of structured queries, reasoning over multiple knowledge facts over evidences. . We will focus on utilizing large scale open-domain structured KBs to answer natural language questions. Specifically, we will first investigate how to model users' query intention inherent in natural language questions, and further how to ground the query representation with the KBs, which will be in turn converted into structured queries and executed against structured KBs. Meanwhile, we will examine how to explore the structured KBs to analyse complex questions and perform reasonings over the KBs. Finally, we investigate the possibility of using structured KBs to improve traditional IR based question answering systems.

随着大规模结构化知识库的兴起,基于大规模开放域知识库的问答系统已成为学术界和工业界的关注热点。本项目以基于结构化知识库的问答系统为主要研究对象,分别从问题意图解析、结构化查询的构造、推理问题解答等角度研究以开放域结构化知识库为基础的自然语言问答技术。不同于以往以信息检索为基础的问答系统,大规模开放域结构化知识库能提供具准确的语义信息,同时其天然的图结构属性又为问题语义建模、高效的结构化查询及推理问题的解答提供了便利条件,从而更好地满足用户的知识检索需求。本项目围绕大规模开放域结构化知识库,集中研究以下四个方面:基于结构化知识库的查询意图分析方法,基于结构化知识库回答事实性问题的方法,基于结构化知识库推理的方法以及如何利用结构化知识库提升现有基于文本的问答系统。

本项目以基于知识驱动的语义分析方法为核心，设计了面向多源异构知识资源的自然语言问答新方法，并创新性的将自然语言问答系列技术应用于面向智能人机交互 的对话理解及回复生成领域，提出了一系列基于多层次文本匹配、用户意图理解、 多轮对话策略、高质量回复生成的新方法，在知识问答、多轮对话等领域取得了 较好的研究成果。项目执行期间，在自然语言处理、人工智能领域国内外重要国 际会议上发表论文 20 余篇(包括 AAAI、IJCAI、ACL、SIGIR 等)。