面向微博应用的新闻文本自动摘要研究

发布时间：2018-01-05 09:41

本文关键词：面向微博应用的新闻文本自动摘要研究　出处：《郑州大学》2017年硕士论文　论文类型：学位论文

【摘要】：Web2.0时代到来后,大数据时代接踵而至,微博等社交媒体越来越受到社会和研究者们的高度关注,面向微博应用的相关任务以及评测也在逐年增加。文本自动摘要主要是通过计算机自动提取文本的关键信息,生成反映文本中心主旨的短文,能够方便压缩文本内容、加快信息交流速度、提升信息检索效率。因此,本文的主要研究工作是面向微博应用的新闻文本自动摘要研究,通过文本自动摘要技术,微博用户可以实现对所关注新闻进行快速缩减然后转发,能够很大程度减少微博用户人工编辑新闻字数再进行转发所耗费的时间,具有很高的实用价值。在深入学习了目前文本自动摘要相关研究方法的基础上,本文提出了基于多层次特征和Ranking SVM(Support Vector Model)排序模型的面向微博应用的新闻文本自动摘要研究算法,具体研究工作如下:(1)多层次特征提取。深入挖掘新闻文本各个统计特征及语义特征即句子词频、句子位置、句子与标题相似度、句子长度、指示性词语等统计特征和主题句这一语义特征,全面分析不同特征的提取方法,积极探索基于新闻文本的表示模型,以更好地运用文本特征。(2)文本句子排序。本文首先预处理新闻文本训练数据中各个句子并将其处理成能够运用的形式,接着将其转换为dat文件作为输入,获得Ranking SVM训练模型,利用得到的模型对处理好的测试数据文本中的各个句子进行排序,最终得到句子由高到低的排序;后期使用冗余处理、通顺度处理、指代消解等相应的可读性加工原则,对排好序的句子进行处理,最终得到比较理想的低冗余度及相对通顺的摘要句集合。最后,本文将上述特征提取方法及句子排序算法在NLPCC2015面向微博新闻自动摘要评测任务的数据集上做实验,实验结果的ROUGE-1值达到50%以上,具备可行性。
[Abstract]:The arrival of Web2.0 era, the era of big data followed, micro-blog and other social media paid more and more attention by society and researchers, the relevant task for micro-blog application and evaluation also increased year by year. Automatic text summarization is the key information automatic extraction of text by computer, text generation reflects the central theme of the passage can be easily compressed the text content, accelerate the speed of information exchange, enhance the efficiency of information retrieval. Therefore, the main work of this paper is to study the news text summarization for micro-blog applications, through the automatic text summarization technology, micro-blog users can realize to the attention of the press are shrinking fast and forward, can greatly reduce micro-blog user manual editing news words again forwarding time, has very high practical value. In depth study of the current automatic text summarization related Based on the research method, this paper proposes a multi level feature based on SVM and Ranking (Support Vector Model) news automatic text summarization algorithm based on ranking model micro-blog oriented applications, specific studies are as follows: (1) extracting multi level feature. Dig the various news statistics and semantic features of that sentence word, sentence the position, title and sentence similarity, sentence length, indicative words such as statistical features and the semantic features of the topic sentence, the comprehensive analysis method to extract different features, and actively explore based on news text representation model, in order to better use the text features. (2) the text sentence ordering. Firstly, preprocessing each sentence of news text training data in the process to use the form, and then convert it to dat file as input, Ranking SVM training model, using the model of processing Each sentence in the text of the test data well sorted, finally get the sentence from high to low ranking; the use of redundant processing, smoothness, readability refers to the corresponding principle of coreference resolution processing, processing of sorted sentences, finally get the ideal low redundancy and relatively smooth Abstract sentence set. Finally, the method of feature extraction and sentence ranking algorithm for micro-blog News Summarization Evaluation task in NLPCC2015 data sets to do the experiment, the experimental results of the ROUGE-1 value reached more than 50%, with feasibility.

【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1

【相似文献】