Bert在微博短文本情感分类中的应用与优化

微博短文本是一种典型的用户生成数据(user generate data),蕴含了丰富的用户情感信息,微博短文本情感分类在舆情分析等众多应用中具有较强的实用价值.微博短文本具有简洁不规范、话题性强等特征,现有研究表明基于有监督的深度学习模型能够显著提升分类效果.本文针对广播电视领域微博文本展开情感分类研究,实验对比了多种文本分类模型,结果表明基于Bert的情感分类方法准确率最高.深入分析实验结果发现,Bert模型对于困难样本的分类错误率较高,为此本文引入Focal Loss作为Bert模型的损失函数,提出一种基于Bert与Focal Loss的微博短文本情感分类方法(简称为Bert-FL方法),使得Bert模型能够更容易学习到困难样本的类别边界信息,实验表明Bert-FL方法的分类准确率绝对提升了0.8%,同时对困难样本的分类准确率也有显著提升.

Application and Optimization of Bert in Sentiment Classification of Weibo Short Text

Weibo short text is a typical type of user generated data,which contains rich user sentiment information.Weibo short text sentiment classification has strong practical value in many applications such as public opinion analysis.The short text of Weibo has the characteristics of conciseness,irregularity and strong topicality.Existing research shows that supervised deep learning models can significantly improve the classification effect.In this paper,sentiment classification research is performed on Weibo texts in the field of broadcasting and television.The experiments compare a variety of text classification models.The results show that Bert’ s sentiment classification method has the highest accuracy rate.In-depth analysis of the experimental results illustrate that the Bert model has a higher classification error rate for hard samples.Therefore,this paper introduces Focal Loss as the loss function of the Bert model,and proposes a Weibo short text sentiment classification method based on Bert and Focal Loss(referred to as Bert-FL method),which makes it easier for the Bert model to learn the class boundary information of hard samples.Experiments show that the classification accuracy of the Bert-FL method has improved by 0.8% absolutely,and the method can significantly improve the classification accuracy of hard samples.