基于深度学习的多文档自动文摘研究

发布时间：2018-07-07 11:32

本文选题：多文档自动文摘 + 玻尔兹曼机　；参考：《吉林大学》2017年硕士论文

【摘要】：互联网迅速发展,每天产生大量文本、图像、视频等类型的数据。其中文本数据是最常见的,用户想查询和了解自己关注的话题需要花费大量时间进行选择文章和阅读文章。自动文摘技术为我们提供了一种快速了解相关话题的方式,该技术可以快速地对文档进行总结,用户只需阅读短短十几句或几十句的总结便可以了解相关信息。用户的需求驱动使得很多类型的文摘方法应运而生,例如:基于主题模型的,基于词袋的等。这些模型初步解决了单文档文摘问题,然而多文档涉及话题较多、特征提取困难,很多模型不能得到令人满意的效果。近年来深度学习在文本处理领域取得了长足的进步,最先进的神经机器翻译模型在多种语言上超过了传统的算法模型,因此本文考虑将深度学习的方法应用到多文档自动文摘中。受限波尔兹曼机是深度学习中的经典模型,可以对数据进行编码,广泛应用在特征降维以及神经网络权重初始化中。由于文本数据特征提取困难,很多特征的作用很难理解,所以本文使用由多个受限波尔兹曼机构建的多层网络模型进行特征降维,这种多层网络结构使我们更容易的获取特征并且使特征包含的文本信息更加完整;然后使用支持向量机模型对文档中的句子进行评分,从多篇文档中抽取代表性句子并进行句子冗余控制,依次选择单位长度质量评分最高的句子生成文摘集;最后对文摘集进行排序,以文摘句在文档中的相对顺序为基准,把同话题的句子聚集到一起,使文摘集合的顺序更合理。主要流程如下:(1)关于多文档的信息表示,我们通过对自然语言处理以及文摘方法的研究,了解到文本理解一般在词-句子-文章-多个文档集合这几个层次表示文档信息。我们使用基于空间向量模型的方式表示文本,尽可能在文档的多个层次提取特征,例如:句子中词的重要性信息,句子内容信息,句子在文档中位置的重要性,句子与文章标题或者查询词的相似度等。(2)关于特征降维,我们使用一层局部特征提取层和两层受限波尔兹曼机模型组合成的多层网络结构进行特征降维,从收集到的大量特征中学习到更加抽象的特征。(3)关于文摘集生成以及文摘句排序,首先我们使用支持向量机模型获取句子评分,将评分较高的句子组合成候选文摘,然后对每个候选文摘句求单位长度文摘质量评分,依次获取评分最高的文摘句组成文摘集,我们称之为增量式的生成文摘方案,这种方法使我们的文摘结果在覆盖度和冗余度都表现良好,最后我们对文摘结果进行文摘句排序使其在逻辑上更合理,我们以遵循句子在文章中的相对顺序为基准,然后把话题类似的句子聚集到一起。
[Abstract]:The rapid development of the Internet, daily production of a large number of text, images, video and other types of data. Among them, text data is the most common. It takes a lot of time for users to select articles and read articles. Automatic summarization provides us with a way to quickly understand the relevant topics. This technology can quickly summarize the documents, and users can understand the relevant information only by reading only a dozen or dozens of summary sentences. Many kinds of abstracting methods come into being because of the user's demand driving, such as theme-based model, word bag based and so on. These models have solved the problem of single document abstracting initially, however, many documents involve more topics, feature extraction is difficult, and many models can not get satisfactory results. In recent years, deep learning has made great progress in the field of text processing. The most advanced neural machine translation model has surpassed the traditional algorithm model in many languages. In this paper, we consider applying the method of deep learning to multi-document automatic abstracting. Limited Boltzmann machine is a classical model in depth learning. It can encode data and is widely used in feature dimensionality reduction and neural network weight initialization. Because the feature extraction of text data is difficult and the function of many features is difficult to understand, this paper uses a multi-layer network model built by multiple constrained Boltzmann mechanisms to reduce the feature dimension. This multi-layer network structure makes it easier for us to get features and make the text information contained in the features more complete. Then we use the support vector machine model to evaluate the sentences in the document. The representative sentences are extracted from multiple documents and the redundant sentences are controlled, and the sentences with the highest score of unit length and quality are selected in turn to generate abstracts. Finally, the summarization sets are sorted, and the relative order of abstract sentences in the document is taken as the benchmark. Bring together sentences on the same topic to make the order of abstracts more reasonable. The main flow is as follows: (1) through the research of natural language processing and abstracting methods, we know that text understanding generally represents document information at the level of word, sentence, article and multiple document sets. We use a spatial vector model to represent the text and extract features at as many levels of the document as possible, such as the importance of the word in the sentence, the content of the sentence, the importance of the sentence's position in the document. (2) about feature dimension reduction, we use a local feature extraction layer and a two-layer constrained Boltzmann machine model to reduce the feature dimension. We learn more abstract features from a large number of features collected. (3) on the generation of abstracts and the ranking of abstract sentences, we first use support vector machine model to obtain sentence scores, and combine the higher-scoring sentences into candidate abstracts. Then for each candidate summary sentence, the unit length summarization quality score is obtained, and the summarization sentence with the highest score is obtained in turn. We call it an incremental summarization scheme. This method makes our summary results perform well in coverage and redundancy. Finally, we sort the abstract sentences to make them more logical. We follow the relative order of sentences in the article as a benchmark. Then bring together sentences similar to the topic.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1

【相似文献】