基于全局语义学习的文本情感增强方法研究

为解决弱情感倾向语料影响文本情感分类的问题,提出基于全局语义学习的文本情感增强方法。首先设计语料划分方法,将语料划分为强情感倾向语料与弱情感倾向语料,然后,从文本处理全过程及整体语义学习的角度出发,构造均值抽取与最大值抽取的语义提取方式及文档信息向量,改进基于循环神经网络的变分自编码器的语义学习过程,并用于学习强情感倾向语料中文本的词语序列特征与语义特征。基于此,对弱情感倾向语料进行重构,实现情感增强目标,最后,将经过情感增强的语料替换掉原来的弱情感倾向语料,再进行情感分类模型的训练与测试。结果表明:提出的文本情感增强方法能够提升情感分类效果,并使得Bert分类器对IMDb影评数据集的情感分类精确率达到了93.03%。

Research on Text Sentiment Enhancement Based on Global Semantic Learning

In order to solve the issues that weak sentimental tendency corpus affects the results of text sentiment classification, an approach of text sentiment enhancement based on global semantic learning was proposed. First, text corpus was divided into strong sentimental tendency corpus set and weak sentimental tendency corpus set, then, from the perspective of processing whole text process and learning whole semantic, two kinds of semantic extracting forms such as mean extracting and maximum extracting were constructed as well as document information vector. The semantic learning process of Variational Auto-Encoder based on Recurrent Neural Network was improved, and it was used to learn the word sequence features and semantic features of strong sentimental tendency corpus texts. On the basis of these, the weak sentimental tendency corpus texts were rebuilt based on the word sequence features semantic features learned from strong sentimental tendency corpus texts and the Variational Auto-Encoder based on Recurrent Neural Network. So the sentiment enhancement of weak sentimental tendency corpus is realized in the end. And then, the original weak sentiment tendency text corpus is replaced by these sentiment-enhanced text corpus. Integrating them with original strong sentimental tendency corpus set forms new text corpus. And train and test process for text sentiment classification were redone based on new text corpus. The results shown that the text sentiment enhancement methods proposed in this paper can improve the effects of text sentiment classification.