用户评论情感分类的方法与实证研究

发布时间：2018-04-16 18:51

本文选题：词向量 + 情感词典　；参考：《兰州大学》2017年硕士论文

【摘要】：文本情感分析,又被叫做情感极性计算,主要包括意见抽取,意见挖掘,情感挖掘,主客观分析等研究方向,旨在对含有人类主观性态度的文本数据进行分析,抽取,挖掘以研究用户对篇章级,句子级或词语级文本的情感态度,是近年来文本挖掘领域比较热门的方向.本文主要研究评论文本的正负向情感分类并在真实评论数据的基础上进行基于细粒度的意见挖掘.在情感分类方面,目前主要的研究方法有:基于情感词典匹配的方法,基于机器学习的方法以及基于深度学习的方法,本文在前辈研究者提出算法的基础上进行改进,考察了一种新的基于词向量和词典的情感分类算法以优化在无标注数据集训练模型情况下网站评论数据的情感分类效果,并将其与基于机器学习的方法如支持向量机,逻辑回归,以及基于深度学习的方法如卷积神经网络,长短期记忆网络进行比较,分析不同模型方法的优劣并提出在实际应用中的可行性建议.在基于细粒度的意见挖掘方面,我们的主要目的是进行基于细粒度的词对(属性词,评价词)抽取以收集并分析用户对产品不同特征属性的评价情况.本文使用的数据集为某品牌笔记本电脑真实评论数据,采用基于依存句法分析结合语法规则进行抽取和筛选的方法,并依据抽取结果进一步分析细粒度层面用户评论的正负情感倾向.
[Abstract]:Text emotion analysis, also called affective polarity calculation, mainly includes opinion extraction, opinion mining, emotion mining, subjective and objective analysis and so on.Mining is a hot topic in the field of text mining in recent years to study users' emotional attitudes towards text at text level, sentence level or word level.This paper mainly studies the positive and negative emotion classification of comment text and carries out fine-grained opinion mining based on the real comment data.In the aspect of emotion classification, the main research methods are as follows: based on affective dictionary matching, machine learning and deep learning.A new affective classification algorithm based on word vectors and dictionaries is investigated to optimize the classification effect of website comment data without annotated dataset training model, and it is compared with machine learning-based methods such as support vector machine (SVM).Logical regression, as well as the methods based on deep learning, such as convolution neural network and long-term and short-term memory network, are compared to analyze the merits and demerits of different model methods and put forward some feasible suggestions in practical application.In the aspect of fine-grained opinion mining, our main purpose is to extract word pairs (attribute words, evaluation words) based on fine granularity to collect and analyze users' evaluation of different feature attributes of products.The data set used in this paper is the real comment data of a brand notebook computer, which is extracted and filtered based on dependency syntactic analysis and grammar rules.According to the extraction results, the positive and negative emotional tendency of fine-grained user comments is further analyzed.
【学位授予单位】：兰州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1;F274

【参考文献】