语义网自动构建中句子成分识别的研究

发布时间：2018-11-02 15:27

【摘要】：互联网的迅速发展，使得世界的信息化程度越来越高。但是现在所使用的万维网无法让人满意，主要问题体现在搜索引擎智能化程度还不够。作为解决方案之一，一种新兴的主要针对于搜索数据的解决方法——语义网，它很好地体现了智能化的优点。为了构建语义网，则需要对资源描述框架进行构成。对于中文来说，构成资源描述框架的工作就是把一个句子的主谓宾成分进行提取后，利用此三元组进行语义网的自动构建工作。依存句法分析是句法分析技术的一种。它利用语言模型、Eisner算法和最大生成树算法来构成依存句法分析的主体。通过找到和谓语动词相依存的几个组块，进一步找到句子的主语和宾语成分。利用这种方式，可以在短时间内获取结构简单句子的主谓宾成分。在此基础之上的语义角色标注是浅层语义分析的一种表现方式，通过合并句子中的各个组块，来分析它们在句子中的起到的主要作用。利用这种方法来分析句子成分有着更高的精确度。其中有关于特征的选择是两种方法的重要组成部分，对最终结果的影响占有很大比例。本文把重点放在特征的挑选上，并选择了合适的对最终结果有效的特征进行使用。利用训练好的语言模型，，进行特征的选择和匹配，使两种处理搭配得当，可以很好地找到句子中的主谓宾三元组。本系统将两种方法进行综合，改善了单纯语义角色方法所带来时间上的问题。使用此系统在公开的未标注的语料库上进行试验，并与单纯的语义角色标注和单纯的依存语法分析的方法进行了对比，正确率高于单纯的依存分析方法，时间上低于语义角色标注的方法。
[Abstract]:With the rapid development of the Internet, the information level of the world is getting higher and higher. But the current use of the World wide Web is not satisfactory, the main problem lies in the search engine intelligence is not enough. As one of the solutions, the semantic Web, a new solution for searching data, embodies the advantages of intelligence. In order to construct the semantic web, it is necessary to construct the resource description framework. For Chinese, the task of constructing a resource description framework is to extract the subject, predicate and object components of a sentence and use the triple to construct the semantic web automatically. Dependency parsing is one of syntactic parsing techniques. It uses language model, Eisner algorithm and maximal spanning tree algorithm to construct the main body of dependency parsing. By finding several blocks that are dependent on the predicate verb, the subject and object components of the sentence are further found. In this way, the subject, predicate and object components of a simple sentence can be obtained in a short time. Semantic role tagging based on this is a representation of shallow semantic analysis. The main role of semantic role tagging in sentences is analyzed by merging the blocks in sentences. Using this method to analyze sentence components has higher accuracy. The selection of features is an important part of the two methods and has a great influence on the final results. This paper focuses on feature selection and selects suitable features that are valid for the final result. The trained language model is used to select and match the features so that the two types can be collocated properly and the subject-predicate triples in sentences can be well found. In this system, the two methods are integrated to improve the time problem caused by the simple semantic role method. The system is tested on the open unannotated corpus and compared with the methods of simple semantic role annotation and simple dependency syntax analysis. The correct rate of this system is higher than that of the simple dependency analysis method. A method in which the time is lower than the semantic role annotation.
【学位授予单位】：沈阳工业大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.1

【参考文献】