【作者】 张磊；

【导师】 王珊；

【作者基本信息】 中国科学院研究生院（计算技术研究所） ， 计算机应用技术， 2002， 博士

【摘要】 因特网的迅速发展导致了信息的爆炸式增长，面对海量的信息，如何使用户能准确快捷地找到所需的信息(What You Get Is What You Want)成为一个迫切需要解决的问题。本文讨论了个性化信息分发及概念检索的技术，在现有语义资源(知网、WordNet)的基础上，提出多视图概念网络模型(Concept Network-Views Model)作为文档表达和用户兴趣模型，利用概念间的联系来提高信息检索的精度。同时综合考虑个性化中的诸多要素，对用户访问模式和反馈进行聚类分析，设计和开发了系统原型框架，从而使个性化信息分发系统更为有效。在概念检索的研究中，我们则是利用领域本体将半结构化数据和结构化数据关联起来，为用户的查询提供概念级的检索结果。实验表明，利用概念间的联系可以有效提高信息检索的精度。 多视图概念网络模型是本文的核心内容，它的生成可以划分为两个阶段：(1)从词空间到概念空间的变换；(2)从概念到多视图概念网络的扩展。在第一阶段中我们给出了基于统计和规则的分词及词性标注、词义排歧方法，提出了扩展短语的定义及其序列模板挖掘算法，以解决目前分词方法的不足；在第二阶段中提出了能量传播的ConceptRank算法和枢纽概念识别算法，用于准确抽取描述用户兴趣和文档的主题概念，并降低枢纽概念产生的噪声。在此模型的基础上提出两种相关度的计算方法：能量消减算法和概念向量的余弦距离。 对个性化信息分发的研究包括三个方面：用户兴趣建模、潜在兴趣的发现和协同过滤技术。(1)本文提出了基于多视图概念网络的用户兴趣建模方法，阐述和分析了个性化的影响因素。(2)通过分析用户的访问模式可以挖掘出用户的潜在兴趣，本文给出潜在兴趣挖掘的流程和算法，从客户端和服务器端收集个性化数据，进行数据预处理后利用Web日志对用户访问过的内容聚类，并将挖掘出的用户潜在兴趣同样用多视图概念网络表示出来，用于信息推荐过程。(3)协同信息过滤技术同样可以支持个性化信息分发，本文提出了基于用户反馈的用户兴趣小组发现算法，使用ISODATA聚类算法对用户的反馈结果进行聚类，调整用户与文档的兴趣度。作为用户兴趣模型的补充，以解决用户兴趣描述的不完备性。 文档的表达同样以多视图概念网络模型为基础，并通过对HTML文件格式的分析来改善网页内容表达的有效性，给出高性能的词典访问方法和倒排文件索引方法。 要实现概念级的检索，就要解决传统检索方法中的如下问题：(1)传统的信息检索是基于关键词的匹配，面临着一词多义和一义多词的问题。(2)由于Internet上包含了半结构化和结构化信息，在检索过程中如何将它们关联集成起来也是非常必要的。本文对半(无)结构化信息进行了分级表达，从原始信息中抽取出词、短语和概念，并引入领域本体以支持文本与数据库的信息关联，从而实现概念检索。文本分类是信息检索中的重要基础，本文提出基于概念的分类方法，修正向量空间模型中单词正交性的错误前提。更多还原

【Abstract】 Information retrieval is concerned with selecting documents from a collection that will be of interest to a user with a stated information need or query. How to let user gets the information he (she) wants is becoming increasingly important in wide-area information system. This dissertation describes a new Concept Network - Views (CN-V) model to represent documents and user’s interests, together with user’s access pattern analysis and ontology theory, which can be used to overcome several limitations both in traditional user’s interests modeling and information retrieval.CN-V model is the kernel of this dissertation, which can be divided into two phase: (1) Transform text from word space into concept space; (2) Generate CN-V model from concepts. In the first phase, we combine statistics technique and rule based method to resolve these problems, and we use wordnet and how-net to disambiguate word sense. Extended phrase mining algorithm is presented to extract semantic units in documents. In the second phase, we give ConceptRank algorithm to extract the topic of documents and user’s interests. At the same time, the concepts that have high ConceptRank but don’t give contribution to the topic will be distinguished as hub concepts, in order to avoid them affect the efficiency. At last, we present two similarity measures based on CN-V: energy decreasing algorithm and cosine measure of concept vectors.Our research on personalized information dissemination include three parts: (1) We give a new user’s interests modeling method based on CN-V model, analyze the main factors in personalization; (2) By analysis on user’s access pattern, we give the potential interests mining algorithm. Firstly we collect personal data, these data will be clustered and represented by Concept Network - Views model after preprocessing, which can be used in information recommendation. (3) Collabratative information filtering technique can support personalized information dissemination too. We use ISODATA algorithm to cluster user’s feedback, and the result can be used to discover user groups that have similar interests.Also, Document representation is based on CN-V model. Further more, the tags of HTML files can be used to represent document efficiently. And we give high efficiency methods for accessing dictionary and invert file indexing.Traditional information retrieval methods are based on keyword matching, which face two difficulties: multivocal and thesaurus. At the same time, how to integrate the semi-structured information and structured information during information retrieval is also an important problem. We represent texts in multi levels and use domain ontology to associate them with the data in database, in order to achieve conceptual retrieval.更多还原