微博短文本细粒度情感分析

发布时间：2018-05-11 19:08

本文选题：微博短文本 + 机器学习　；参考：《北京邮电大学》2016年硕士论文

【摘要】：微博作为是移动互联网起步最早也发展最快的业务之一,在经历了最初几年的高速增长和热捧之后,一度有些沉寂。但是,伴随着4G网络、WiFi网络大规模覆盖,移动终端智能设备的极速增长,越来越多的人频繁地使用微博。移动端用户的增长,让微博有了日活跃用户超过1亿的基础。基于该庞大的用户群体,利用微博文本对用户进行情感分析,不仅有利于新生代商业模式探索、社会舆情监控与分析,而且对人工智能发展的积极意义也不容小觑。文本情感分析为人工智能在人类情感领域的探索做出了积极的贡献。微博具有典型的网络语言特点,语法规范性差、口语化、新词汇层出不穷、大量的表情符号和文本噪声等。基于传统规范书面语言词典的微博情感分析效果差强人意。为了解决这个问题,本文一方面通过对传统情感词典进行网络语言的扩充,构建了新的适用于微博文本分析的词典,并通过实验验证了词典的有效性;另一方面,使用对特殊情感词汇依赖性小的递归自动编码机(Recursive Autoencoder,RAE)模型,性能得到了提高。含有否定词的文本情感判别较.为复杂,否定词的加入可能使原文本情感极性发生转变,也可能极性保持不变,程度有所削弱。传统本文传统的情感分析方法通常对文本模型进行简化,假设一个词语仅和其前一个词语有关,对含有否定词的文本的积极/消极判别较差。递归神经张量网络(Recursive Neural Tensor Network,RNTN)模型没有进行类似简化,保持了词向量之间的强相关性。本文通过对RNTN模型的训练,使其在含有否定词中文文本情感极性判别上达到了 80.7%的精度。RNTN模型对词语或短语级细粒度情感分析上表现出色,而且不需要大量的人工标注;RAE模型作为较为成熟的深度学习模型,抛弃了传统的词袋模型,利用层次结构和成分语义来进行情感分类;SVM作为传统机器学习方法,通用性强,适用范围广。结合以上方法的特点以及情感分类的任务目标,本文设计了微博短文本的双极性、细粒度情感分类方案,首先使用SVM对微博篇章进行主客观分类,RAE和RNTN对判定为主观情感的文本的每一个句子进行正负极性判断,并选出极性最强的句子,该句子的情感极性即为整个微博的情感极性。若RAE和RNTN评判语句极性一致,即为最终结果。本文设计方案F值、召回率、准确率均高于基于融合特征的方法。通过实验证明了系统的可行性和高效性。
[Abstract]:As one of the earliest and fastest growing mobile Internet services, Weibo has been quiet for a while after its rapid growth and popularity in the first few years. However, with the large-scale coverage of 4G networks and the rapid growth of smart devices in mobile terminals, more and more people frequently use Weibo. The growth of mobile users has given Weibo a daily base of more than 100m active users. Based on this huge group of users, the use of Weibo text for emotional analysis of users is not only conducive to the exploration of the new generation of business models, monitoring and analysis of social public opinion, but also positive significance for the development of artificial intelligence can not be underestimated. Text emotion analysis has made a positive contribution to the exploration of artificial intelligence in the field of human emotion. Weibo has the characteristics of typical network language, poor grammar, colloquial, new vocabulary, a large number of emoji and text noise and so on. The effect of Weibo affective analysis based on the traditional standard written language dictionary is unsatisfactory. In order to solve this problem, on the one hand, this paper constructs a new dictionary suitable for Weibo text analysis by extending the traditional emotional dictionary, and proves the validity of the dictionary through experiments; on the other hand, The performance is improved by using the recursive autocoder rae model, which has small dependency on special emotional vocabulary. The text emotion discriminates with negative words. For complexity, the addition of negative words may change the emotional polarity of the original text, or keep the polarity unchanged, and weaken the degree of the original text. The traditional affective analysis method in this paper usually simplifies the text model, assuming that a word is only related to the previous word, and the positive / negative discrimination of the text containing negative words is poor. Recursive Neural Tensor Network (RNT) model of recurrent neural Zhang Liang network is not simplified, and the strong correlation between word vectors is maintained. In this paper, the RNTN model is trained to achieve a precision of 80.7% in judging the affective polarity of the Chinese text with negative words. The model performs well in the fine-grained emotional analysis of words or phrases. Moreover, it does not need a large number of manual tagged rae model as a more mature in-depth learning model, abandoned the traditional word bag model, using hierarchical structure and component semantics to carry out emotion classification and SVM as the traditional machine learning method, it has strong generality. The scope of application is wide. Combining the characteristics of the above methods and the task goal of emotion classification, this paper designs a bipolar, fine-grained emotional classification scheme for Weibo short text. Firstly, SVM is used to classify Weibo texts subjectively and objectively. Rae and RNTN judge each sentence as subjective and negative, and select the most polar sentence. The emotional polarity of the sentence is the emotional polarity of the whole Weibo. If the polarity of RAE and RNTN judgment statements is the same, it is the final result. In this paper, F value, recall rate and accuracy are higher than those based on fusion feature. The feasibility and efficiency of the system are proved by experiments.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1;TP393.092

【相似文献】