【作者】 刘柏嵩；

【导师】 高济；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2007， 博士

【摘要】 语义Web提出以来，本体(Ontology)正在成为人工智能和知识工程中一种重要的工具，在知识的获取、表示、分析和应用等方面具有重要的意义。从本体开发的角度来说，由于借助本体编辑器(如Protégé)手工开发本体是一个繁重和棘手的任务，极易导致知识获取的瓶颈。因此，Web本体的可获取性已被学术界公认为是制约语义Web成功的巨大挑战之一。从现有信息源，包括文本、词典、遗留知识库、WWW文档等，获取领域知识、以自动方式构建或扩充本体，即所谓的本体学习(OntologyLearning)，是开发本体的有效途径。 目前国际上在本体学习方面的研究非常活跃，虽然已经提出了很多本体学习方法，但大部分方法都不理想。由于缺乏统一的本体学习体系结构概念和方法，虽然开发了一些本体学习方法，但这些方法难以被其他系统重用。当前本体学习系统工具多是原型系统，不能大规模实时处理网上海量信息源，也缺乏中文语料处理能力。同时，目前还缺乏有效的评价本体学习结果的标准和方法，不利于本体学习方法和工具的进一步发展。 本文就是在开放的网络环境下，综合运用机器学习和自然语言处理方法，按照分层技术原理，提出了一种新的分层本体学习方法体系，并实现了一种基于Web的多策略本体学习工具GOLF，然后讨论了本体演化和评价方法，并对本体学习框架GOLF进行了实验和评价。本文的主要研究内容如下： (1)提出了一种分层本体学习方法体系，其中包括术语自动抽取、概念学习、实例学习和分类关系学习和非分类关系学习多种关键技术。在对现有学习方法做大量改进的基础上，完全实现了本体学习全过程的无缝集成，并且在本体学习系统中集成了本体评价模块。 (2)开发了基于Web的多语种通用本体学习工具GOLF，并采用Web文档作为本体学习源，进行了跨领域、多语种实验。同时，实现本体学习过程中的本体演化管理，并对学习结果进行评价和反馈。 (3)在本体学习中引入多策略学习方法，以提高学习质量。各学习算法的组合框架采用概率组合分布，可根据不同的语料特征为每个算法设定权值，从而增强了对不同领域语料的适应性。通过实验对比分析，在学习结果的准确率和召回率方面，GOLF系统都比著名的Text2Onto系统更好。 (4)实现了中文语料的本体学习。本文所提出的方法和工具能够很好地处理多语种语料，与同类系统(大多只能处理西文文本)相比，对中文的处理能力明显加强；特别针对中文的语言学特征，引入HowNet语义词典，添加了对应于中文文本的语言学模式和停用词表，性能有明显改善。 (5)提出了一种新的基于贝叶斯决策理论的本体评价方法RiMOE，并采用RiMOE更多还原

【Abstract】 Since the Semantic Web had been proposed, ontology is become an important tool in the artificial intelligence and knowledge engineering. And it is of great significance to the acquisition, representation, analysis and application of knowledge areas. An issue named "ontology bottleneck" , the lack of efficient ways to build ontologies, has been coming up to generate ontologies. Therefore, it is an urgent task to improve the methodology for rapid development of more detailed and specialized domain ontologies. A framework of automatic extract ontology knowledge from the existing source of information, which can reduce the cost, is an effective way of ontology rapid-development.At present, the research of Ontology learning is a trend in the computer science dispciline. A lot of ontology learning methods have been proposed, but most of them are not perfect. The existing ontology learning methods are all in need of manual work, and the fully automatic approach is unrealistic in the short term. However, due to the massive nature of Web resources, we still need to further improve the degree of automation, and reduce the participation of users. In addition, most of the ontology learning tools are very limited, because they can only handle certain types of data sources, or capture some objects, but can’ t process Chinese corpus. Due to the limitation of the existing ontology learning methods, these tools are still very immature; and some of the latest research results have not been used.In this paper, we combine the NLP and machine learning methods in the open network environment. Firstly, we in-depth discuss the key technologies of ontology learning, and propose a web-based multi-strategy Ontology learning framework (called GOLF). And then we discuss the way of ontology evaluation and evaluate the GOLF by several experiments. The main research contents of the dissertation contains as follows:1) According to layered approach, we propose a layered ontology learning framework, including the automatic extraction of terms, domain concepts learning, instances learning, taxonomy and non-taxonomic relations learning. In order to achieve a seamless integration of ontology learning process, we improve these technologies which are also applied in our ontology learning framework. And originately the ontology evaluation module is integrated in更多还原