Sentiment Analysis of Financial Texts Based on Attention Mechanism of FinBERT and BiLSTM

More and more studies are using sophisticated textual sentiment models to help us better understand behavioral finance patterns across financial market participants. Most of the existing sentiment analysis methods are oriented to general fields. Most text representation extraction methods use fixed token encoders. Generally, sentiment analysis models are invalid for financial applications. To overcome these challenges, a text sentiment analysis model (BBiLSTM-Attention) for financial fields is proposed. The model uses the pre-training language model FinBERT as a feature extractor to dynamically obtain the context information of comments, and combinate BiLSTM and multiple attention mechanisms to extract the sentiment of financial comments. Experiments is performed using financial field commentary dataset. The results show improved accuracy and generalization ability, accuracy reached 79.33%, and F1-score reached 0.8068.