在互联网时代,越来越多的财务公司选择在财经新闻平台上发表自己的见解,这些评论文本作为舆情的载体,可以充分反映财务公司的情绪,影响公众的投资决策和市场走势.情感分析为分析海量的经济类文本情感类型提供了有效的研究手段.但是,由于特定领域文本的专业性和大标签数据集的不适用性,经济类文本情感分析给传统的情感分析模型带来了巨大的挑战.当将一般情感分析模型应用于经济等特定领域时,模型在准确率与召回率上表现较差.为了克服这些挑战,文章针对财经新闻平台上的经济类文本的情感分析任务,从词表示模型出发,提出了基于知识蒸馏方法的双路BERT(Two-way BERT based on knowledge distillation method)情感分析模型,与文本卷积神经网络(Text-CNN)、卷积递归神经网络(CRNN)、双向长时和短时记忆网络(Bi-LSTM)等算法进行对比实验,结果得出该改进方法相较于其他算法在准确率、召回率和F1值均提升了 1％～3％,具有较好的泛化性能.
Sentiment Analysis of Economic Texts Based on Improved Two-Path BERT Based on Knowledge Distillation
In the international age,more and more financial companies choose to give their views on the financial and economic news platform.As the carrier of public opinion,these comments can fully reflect the emotions of financial companies and affect the public's investment decisions and market trends.Emotional analysis provides an effective research tool for analyzing the emotional types of massive economic texts.However,due to the professionalism of texts in specific fields and the inapplicability of large label data sets,the emotional analysis of economic texts brings a huge challenge to the traditional emotional analysis model.When the general sentiment analysis model is applied to specific fields such as economy,the model performs poorly in accuracy and recall.In order to overcome these challenges,this paper builds a two-way BERT(based on knowledge decomposition method two way BERT model)emotional analysis model based on knowledge distillation method for the emotional analysis task of economic texts on the fi-nancial news platform,Text convolution neural network,Convolutional recurrent neural network,Two way long term and short term memory networks(Bi-LSTM)and other algorithms are compared and tested.The results show that the accuracy,recall andF1value produced by this improved method improved by 1％～3％compared with other algorithms,and it has good generalization perform-ance.
NETL
NSTL
万方数据
