基于深度学习的文本情感分析

发布时间：2018-02-28 22:18

本文关键词： 文本情感分析深度学习卷积神经网络循环神经网络　出处：《吉林大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着Web 2.0时代的蓬勃发展,互联网上出现了大量的数据。人们在博客,微博,产品评论,电影评论,网络讨论群等区域留下了非常多的文本信息。这些非结构化的文本中包含了作者的思想,情感,观点以及看法。如果能够从这些非结构化的数据中提取出情感数据,将会推动自动抉择支持、网络舆情风险分析、信息预警、商品销售的发展,在科研以及实际应用中具有非常大的价值。传统的用于解决文本情感分析问题的方法包括基于知识的方法,基于统计的方法以及混合的方法。在数据量不大或者语义不够丰富的时候,这些方法能够取得一定的效果。但是随着数据量越来越大,表达方式越来越丰富,传统的方法已经无法有效地解决这一类问题,新的方法亟待提出。深度学习自2006年以来获得了人学术界以及工业界广泛的关注。虽然在整体架构上,基于深度学习的方法与传统的神经网络相似,但是由于采用了不同的数据表示方式以及训练方式,梯度扩散、过拟合等问题得到了有效地解决。目前,在图像识别,语音识别等领域,基于深度学习的方法已经取得了比传统的机器学习方法更好的效果。卷积神经网络和循环神经网络是深度学习中两个比较有效的模型,前者适合从数据中提取出局部特征,而后者能够有效地分析时序数据。单独地使用这两个模型中的一个难以在文本情感分析任务中取得令人满意的效果,因此出现了由二者共同构成的混合模型。本文针对混合模型中存在的缺陷做出了三点改进:优化输入向量序列,将文本转化为等长的输入向量序列;提出一种新的激活函数,有效缓解了梯度消失的问题并提高了模型的泛化能力;使用Max Pooling技术提取局部特征的最大值。从Yelp2015数据集的实验结果可以看出,本文提出的三点改进是有效的。此外,本文针对模型中比较重要的参数做了多组对比实验,研究了这些参数对于模型的影响。
[Abstract]:With the rapid development of the Web 2.0 era, there is a lot of data on the Internet. People on blogs, Weibo, product reviews, movie reviews, Web discussion groups and other areas leave a lot of text information. These unstructured texts contain the author's thoughts, feelings, opinions and opinions. If emotional data can be extracted from these unstructured data, Will promote the development of automatic choice support, network public opinion risk analysis, information early warning, commodity sales, It has great value in scientific research and practical application. The traditional methods used to solve the problem of text emotion analysis include knowledge-based methods. Methods based on statistics and mixed methods. When the amount of data is small or semantic is not rich enough, these methods can achieve certain results. But as the amount of data increases, the expression becomes more and more abundant. Traditional methods have not been able to solve such problems effectively, and new methods need to be proposed. Since 2006, in-depth learning has received extensive attention in both academia and industry, although in the overall framework, The method based on depth learning is similar to the traditional neural network, but the problems such as different data representation and training, gradient diffusion and over-fitting are solved effectively. In the field of speech recognition, the method based on depth learning has achieved better results than traditional machine learning methods. Convolution neural network and cyclic neural network are two more effective models in depth learning. The former is suitable for extracting local features from the data, while the latter can effectively analyze temporal data. In this paper, three improvements are made to the defects of the mixed model: optimizing the input vector sequence, transforming the text into the equal length input vector sequence, and proposing a new activation function. The problem of gradient disappearance is effectively alleviated and the generalization ability of the model is improved. The maximum value of local feature is extracted by using Max Pooling technique. The experimental results of Yelp2015 dataset show that the three improvements proposed in this paper are effective. In this paper, the effects of these parameters on the model are studied.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1

【相似文献】