基于渐进直推式支持向量机的Twitter文本情感分析研究

发布时间：2018-03-15 06:09

本文选题：Twitter　切入点：情感分析　出处：《合肥工业大学》2015年硕士论文　论文类型：学位论文

【摘要】：近年来,随着互联网技术的迅猛发展和移动终端的进一步普及,社交网络(Social Network Service)己逐渐深入人们的日常生活。Twitter是一个典型的社交网络及微博客服务网站,其访问量位列全球互联网网站前十位。用户可随时随地更新自己的状态,同时可以关注其他用户的动态。每天数以亿计的推文记录了用户的所见、所闻、所为、所感,分享了他们的喜、怒、哀、乐。对表达于Tweets中的情感特征进行深度挖掘在市场营销、舆情监控、突发事件应急等方面都有重要作用。情感计算是近年来人工智能领域的研究热点。情感的主观性、隐藏性、判定标准不统一性等特点致使文本情感分析有别于传统的文本分析。而Tweets的短文本性、口语化、高噪声等特性使得Twitter'情感分析更具挑战性。针对Twitter的特点,研究了Twitter情感分析的若干关键问题,主要包括Tweets的文本预处理、特征分析、基于半监督的情感分析算法等。为降低Tweets中噪声对情感分析造成的影响,研究了一系列改良的Twitter文本预处理方法,重点对URLs、否定词、重复字母等信息进行预处理,并通过对比试验验证了预处理方法对Twitter情感分析的有效性。同时,分析了不同选择标准在特征选择方面的能力,重点对比了文档频率、信息增益和卡方统计量特征选择的效果。实验结果表明,预处理及特征选择操作在提高了情感分类准确率的同时有效降低了特征空间的维度。另外,为克服大量标注数据的难获取性,避免海量未标注数据的资源浪费,研究了一种基于渐进直推式支持向量机的半监督Twitter情感分析算法。通过引入扰动因子稳步优化了Twitter情感分析效果,并且可以自适应数据分布,自动控制学习进度和训练时间。
[Abstract]:In recent years, with the rapid development of Internet technology and the further popularization of mobile terminals, Social Network Service has gradually penetrated into people's daily life. Twitter is a typical social network and Weibo guest service website. Users can update their status anytime, anywhere, and follow the movements of other users. Hundreds of millions of tweets every day record what they see, hear, do and feel. Sharing their joy, anger, sadness, joy, deep mining of emotional features expressed in Tweets in marketing, public opinion monitoring, Emotional computing has become a research hotspot in the field of artificial intelligence in recent years. The disunity of the judging standard makes the text emotional analysis different from the traditional text analysis. However, the characteristics of Tweets, such as short text nature, colloquialization and high noise, make it more challenging. This paper studies some key problems of Twitter affective analysis, including Tweets text preprocessing, feature analysis, semi-supervised affective analysis algorithm and so on. This paper studies a series of improved Twitter text preprocessing methods, focusing on the preprocessing of URLs, negative words and repeated letters, and validates the validity of the preprocessing method for Twitter affective analysis through comparative experiments. The ability of different selection criteria in feature selection is analyzed, and the effects of document frequency, information gain and chi-square statistics feature selection are compared. The experimental results show that, Preprocessing and feature selection not only improve the accuracy of emotion classification, but also reduce the dimension of feature space effectively. A semi-supervised Twitter affective analysis algorithm based on progressive direct push support vector machine (SVM) is studied. By introducing disturbance factor, the effect of Twitter affective analysis can be optimized steadily, and the data distribution can be adaptively distributed, and the learning schedule and training time can be automatically controlled.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.1

【相似文献】