基于FinBERT-CNN的股吧评论情感分析方法

A Sentiment Analysis Method Based on FinBERT-CNN for Guba Stock Forum

我国股市波动受投资者情绪变化影响较大,通过对股吧等金融交流平台上投资者的评论进行情感分析,能够帮助投资者更好地了解股票市场的变化。现有的情感分析方法是利用模型对股票评论集进行分析,但缺少优质的股票评论标注数据集用于模型训练,且单一模型提取股票评论特征较为片面,模型的准确性有待提高。该文针对股吧平台上的评论数据,提出一种基于FinBERT-CNN的股吧评论情感分析方法,该方法通过FinBERT预训练模型学习股吧评论数据语义特征,解决缺乏股吧评论标注数据集的问题,并利用卷积神经网络学习股吧评论的局部特征,使模型充分学习股吧评论特征,提高模型情感分类的准确性。实验结果表明,基于FinBERT-CNN的股吧评论情感分析方法均优于现有情感分析方法。此外,通过基于股吧评论情感的股票市场关联分析,验证了股吧评论情感变化与股市波动存在相关性。

The fluctuation of the stock market greatly depends on investors’ sentiment-based factors.Sentiment analysis of investors’ reviews on financial exchange web platforms such as Guba stock forum(guba.com.cn), can help stockholders to understand the stock market more effectively. However, due to the unavailability of high-quality labeled datasets and deficient features of stock comments extracted by a single model, the accuracy of the existing sentiment methods still requires further improvements. This paper proposes a method that utilizes the FinBERT-CNN-based sentiment model for Guba comments. The semantic features of Guba comments are extracted by using the FinBERT pre-training model. Meanwhile,a convolution neural network(CNN) is applied to learn the local features of Guba comments. It enables the proposed method to learn features more precisely and improve the emotion classification’s accuracy significantly. Experiments show that the proposed method outperforms the existing models. Furthermore,the correlation analysis on Guba comments and stock market data demonstrates a relationship between the investors’ emotions and stock market volatility.