【作者】 陈超；

【导师】 贺樑；

【作者基本信息】 华东师范大学 ， 计算机应用技术， 2013， 硕士

【摘要】 随着信息技术的快速发展和日益成熟,互联网得到了充分的发展并取得了长足的进步。越来越多的信息先于传统媒体被发布到互联网上从而带来了互联网数据的极大繁荣。然而,数据的急剧增加并没有带来知识的快速增长,反而为知识的获取带来极大的不便。因此,人们迫切地希望找到一种自动化的工具来对海量的数据和信息进行处理,从而实现知识的快速检索和定位。信息抽取正是在这种背景下产生的。而实体关系抽取作为信息抽取的一个重要环节和关键性子任务,也受到了众多研究者的青睐。目前,实体关系抽取技术主要分为基于知识工程的方法和基于机器学习的方法。基于知识工程的方法由于需要大量的人工参与,并且具有较强的领域相关性,因此,不适于在互联网等场合采用。而基于机器学习的方法按所需人工参与的多少分为监督的学习方法、半监督的学习方法和无监督的学习方法。其中,半监督的学习方法由于所需要的人工参与较少,并且具有较高的准确率和召回率,从而得到了广泛地应用。二元实体关系抽取问题是实体关系抽取中的基本问题,即运用自然语言处理的相关技术抽取两个实体间的关系。本文重点研究基于互联网的二元实体关系抽取问题,并探索了一种基于自举技术的实体关系抽取方法：首先,该方法对现有的基于自举技术的实体关系抽取过程中的两个关键的处理环节进行了改进；其次,该方法对现有的抽取过程进行了扩展。本文的研究主要包括：1.对现有的关系描述模式的构成和获取技术进行了分析和研究,并通过与信息增益的相关理论和研究成果相结合,提出并实现了一种基于信息增益的关系描述模式获取方法。通过该方法获取的关系描述模式能够有效地表达实体对间的关系。2.对现有的获取新关系元组的方法进行了研究,结合基于信息增益的关系描述模式获取方法,设计并实现了一种基于模式匹配的新关系元组获取方法。该方法能够有效地提高新关系元组获取的准确率和召回率。3.通过对中文中二元关系的表达方式进行深入的观察和统计分析,提出并实现了一种基于最小覆盖的关系元组验证方法。该验证方法能够大幅提高返回的关系元组的准确率。最后,本文以互联网作为基础语料库对提出的各种方法进行了性能评测,实验结果表明,本文提出的方法能够有效地提高二元实体关系抽取的准确率和召回率,从而验证了所提出的各个方法的有效性。更多还原

【Abstract】 With the rapid development and becoming more mature of information technologies, Internet experienced continuous development and gained great achievements. More and more information is published on the Internet prior to traditional media, which brings prosperities to the Internet. However, rapid increment of information did not lead to faster growth of knowledge rather made it more difficult to acquire valuable knowledge. Therefore, in order to retrieve and locate our requested knowledge, we urgently need a tool to help us deal with massive data and information. Information extraction exactly emerged under this background. As a key component of information extraction, entity relationship extraction attracted attention of many researchers.Generally speaking, two well-known paradigms namely Knowledge Engineering paradigm and Machine Learning paradigm have been proposed to address this problem currently. Approaches based on Knowledge Engineering need substantial human involvements and are domain related. Therefore, it cannot be applied to open corpora such as the Internet. Approaches based on Machine Learning can be classified into supervised approaches, weakly supervised approaches and unsupervised approaches according to the quantity of human involvements in one approach. Weakly supervised approaches, which need a handful of human involvements and can get upper precision rate and recall rate, have been widely researched and deployed.Binary relation extraction is the basic problem of entity relation extraction, namely adopting related technologies of NLP (Natural Language Processing) to extract relationship between two entities. This paper focuses on the research about the binary relation extraction problem basing on the World Wide Web and explores a weakly supervised approach basing on bootstrapping techniques. This approach extends the previously existent prototype system bases on bootstrapping techniques and enhances two key components of it. The research is mainly comprised of:1. Through analyzing compositions and extraction methods of existing relation patterns combined with theory and research results of information increment, this paper proposes a method to extract relation patterns basing on information increment. Patterns generated by this method can be applied to describe relationship between two entities effectively.2. Through researching existent traditional approaches used to obtain new relation triples and taking advantages of method proposed to extract relation patterns, this paper designs and implements an approach to gain new relation triples bases on pattern matching. The method can improve the precision and recall rate of gaining new relation triples.3. By means of deep research about expressions of Chinese binary relation, this paper proposes a minimum cover template based approach to assess new triples. The proposed approach observably improves the precision rate of output triples.Finally, this paper designs a lot of experiments to evaluate performances of the proposed approaches taking World Wide Web as the corpus. The experimental results showed that approaches proposed in this paper are able to effectively improve the precision and recall rate of relation extraction, thereby, the effectiveness of proposed methods were validated.更多还原