基于深度学习的Web信息抽取研究与实现

发布时间：2018-04-03 22:41

本文选题：信息抽取　切入点：深度学习　出处：《浙江大学》2017年硕士论文

【摘要】：近年来,海量的信息数据随互联网时代的飞速发展不断注入到Web当中,因此从Web中精确地获取所需信息的重要性也日益凸显。但由于Web页面中信息种类的繁多、结构格式之间的差异等原因,使得传统利用手工编写规则进行信息抽取的方法过分依赖于具有专业领域知识的人的劳动并且通常具有较差的可移植性。循环神经网络在文本语义理解以及处理序列化数据中的优势使其十分适用于对Web页面进行信息抽取,有鉴于此,本文构建了一种基于双层循环神经网络的模型用于Web页面信息的抽取,其中第一层循环神经网络用于理解Web页面中每个文本节点中字词序列的语义,而第二层网络用于理解页面中多个文本节点相互间的序列化描述关系。在模型成功建立之后,本文使用TensorFlow深度学习编程系统对模型进行了实现,并在实验中对模型中各个主要参数逐一进行调整对比,在优化了模型的同时也得到了不同参数对于模型整体识别准确率的影响,并给出一般性的结论。本文构建的信息抽取模型经实验分析及优化后,对于同网站的Web页面识别准确率可达到900%以上。在此基础上,本文又对模型做了改进,使其可以对跨网站同类Web页面进行信息抽取,且信息抽取的准确率同样可以达到较为理想的水平,为提升Web信息抽取的效率及移植性提供了一种新的方法。
[Abstract]:In recent years, with the rapid development of the Internet era, massive information data has been continuously injected into the Web, so the importance of accurately obtaining the required information from the Web has become increasingly prominent.However, due to the variety of information in Web pages and the differences in structure and format,The traditional method of information extraction by manual writing rules is too dependent on the labor of people with professional domain knowledge and usually has poor portability.The advantage of circular neural network in semantic understanding of text and processing serialized data makes it very suitable for information extraction from Web pages.In this paper, a model based on double-layer cyclic neural network is constructed for extracting the information of Web pages. The first layer of cyclic neural networks is used to understand the semantics of word sequences in each text node in a Web page.The second layer network is used to understand the serialization relationship between multiple text nodes in the page.After the successful establishment of the model, this paper uses the TensorFlow depth learning programming system to implement the model, and adjusts and compares the main parameters of the model one by one in the experiment.At the same time, the influence of different parameters on the recognition accuracy of the model is obtained, and a general conclusion is given.After the experimental analysis and optimization of the information extraction model constructed in this paper, the recognition accuracy of Web pages of the same website can reach more than 900%.On this basis, the model is improved so that it can extract information from similar Web pages, and the accuracy of information extraction can reach an ideal level.It provides a new method to improve the efficiency and portability of Web information extraction.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181;TP393.09

【相似文献】