【作者】 傅魁；

【导师】 聂规划；

【作者基本信息】 武汉理工大学 ， 管理科学与工程， 2007， 博士

【摘要】 本体能够支持人机之间、机器之间的信息交换、知识共享与重用，而得到越来越广泛的重视、研究和应用。然而，领域本体的匮乏却是困扰本体理论研究与现实应用的最主要瓶颈之一，本体学习应运而生，它能够以自动或半自动化的机器学习方式从多种不同的数据源中获取本体。相比国外较多本体学习研究而言，中文环境下本体学习刚刚拉开序幕。本文通过对基于Web的本体学习的研究，为具有实用价值的中文本体学习系统的研发提供理论方法基础。论文在借鉴国外现有的本体学习理论、方法和技术的基础上，结合中文自然语言处理的研究成果，对中文环境下领域本体的概念获取、继承关系学习、属性关系学习和本体实例获取的理论方法展开研究。论文主要研究内容如下：(1)通用本体学习系统的体系结构。设计了一种通用本体学习系统体系结构，由建立在资源层基础上的五大功能模块构成，分别为：资源管理模块、通用资源读写模块、数据预处理模块、本体抽取模块和本体评价与编辑模块。本文所提出的基于Web的本体学习的方法可组件式地无缝集成到该体系结构中。(2)多策略领域概念获取。提出了一种融合信息抽取、中文自然语言处理、语言学和统计等多种策略的领域概念获取算法。能根据页面块特征判定结果自适应选择信息抽取或基于隐马模型和候选名词短语约简的术语获取方法，研究了基于搜索引擎的术语间同义词关系识别方法以及领域概念的过滤算法。(3)继承关系学习。提出了基于Web分类目录判定的继承关系学习方法和基于语境自学习的继承关系学习方法。前者主要包括网页中Web分类目录判定算法、分类目录标注规则、隐式分类目录模式发现机制、标注文档合并中的歧义消解算法以及继承关系映像规则。后者主要包括继承关系语境的自学习机制和基于语境的继承关系获取算法。两种方法各有优缺点，具有互补性。(4)基于知网的属性关系学习。属性关系具有重要作用，但研究很少。首先采用基于语境自学的方法获取候选属性集合；分析认为候选属性集合由非属性词汇、无效属性和有效属性构成，提出了利用知网中属性义原所描述的上下位关系实现非属性词汇过滤和利用属性一一宿主关系实现无效属性过滤的算法；研究基于领域概念树的实现属性关系映射与修剪的基本规则，设计了相应的算法。(5)本体实例的获取。分析了Web网页中个体知识表示的主要形式，提出了基于本体的网页主题概念和个体知识表示特征的判定算法，重点设计了Web表格中本体实例获取的规则，包括本体实例表格识别规则、属性单元识别规则、属性值单元识别的基本规则和扩展规则、实例名称识别规则，给出了算法总体描述。更多还原

【Abstract】 Ontology can support information exchange, knowledge sharing and reuse between human and machine, machine and machine, therefore, it gained increasing attention, research and application. However, the scarcity of domain ontology is one of the main bottlenecks which plagued ontology theoretical research and practical application, so ontology learning emerged as the times require. It is able to acquire ontology from various different data sources, using automatically or semi-automatically machine learning methods. Compared to many ontology studies abroad, ontology learning in the Chinese environment is just started off. This dissertation researches on the basic theory and methodology of web-based ontology learning which we hope will provide the fundamental support in theory and methodology for developing a practical Chinese ontology learning system.Based on the existing ontology learning theory, methods and techniques from abroad, the dissertation combines the research fruits of Chinese natural language processing field, studies the theory and methodology of domain concept acquisition, is-a relation learning, attributive relation learning and ontology instance acquisition in the Chinese environment. The detailed work is as follows:(1) General architecture of ontology learning system. A general architecture of ontology learning system is designed, which consists of five functional modules: resource management module, universal resource reading and writing module, data preprocessing module, ontology extraction module and ontology evaluation and editing module. Our proposed web-based ontology learning methods in this dissertation can be seamlessly integrated into the architecture.(2) Multi-strategy domain concept acquisition. This dissertation presents a multi-strategy domain concept acquisition algorithm which combines information extraction technique, Chinese natural language processing technique, linguistics method and statistical analysis, etc. According to the features of page block, our method can self-adaptively choose the term acquisition method which is based on information extraction technique or based on hidden markov model and reduction of NP candidates. The search engine-based method of synonymous relation identification among the terms and the filtering algorithm of domain concept are discussed in detail.(3) IS-A relation learning. This paper proposes two methods for is-a relation learning: method based on the discrimination of web directory and method based on self-learning context. The former includes web directory discriminant algorithm, tagging rules for directory, implicit web directory pattern discovery mechanism, disambiguation algorithm in the process of merging taged documents and mapping rules from web directory to is-a relation. The latter includes research of inheritance context self-learning mechanisms and is-a relationship acquisition algorithm based on inheritance context. Both methods have advantages and disadvantages, are complementary.(4) HowNet-based attributive relation learning. Attributive relations are important, but have been few studied. Firstly, the candidate attributes aggregation is obtained using our methods based self-study context. Secondly, the constitution of candidate attributes aggregation is analyzed and subdivided into three types: non-attribute vocabulary, invalid attribute, and valid attribute. The HowNet-based filtering Algorithms are presented which filter the non-attribute vocabulary using the hyponymous relation described by the attribute sememe and filter the invalid attribute using the attribute-host relation described by the attribute sememe in HowNet. Thirdly, this dissertation studies the basic rules for attribute mapping and attributive relation pruning based on the domain concept tree. The corresponding algorithm is proposed at last.(5) Ontology instance acquisition. Firstly, the main types of individual knowledge in web pages are analysed. Secondly, the ontology-based discriminant algorithms of web theme concepts and characteristics of the individual knowledge representation are proposed. Thirdly, we focus on the design of rules for ontology innstance acquisition from the web tables, including of ontology-instance form’s identification rules, attribute unit identification rules, the basic rules and expansion rules for attribute-value unit identification, and instance name recognition rules. In the end, the overall description of the algorithm is described.更多还原