Financial Text Sentiment Analysis Based on DK-BERT Model and Its Application in Market Forecasting

This paper proposes a novel model, DK-BERT, which enhances BERT with domain-specific knowledge graph for financial text analysis. DK-BERT embeds financial knowledge graphs to capture implicit relationships between professional financial terms. Using a dynamic attention mechanism, the weight is adjusted according to the correlation between vocabulary and knowledge graph to improve semantic understanding. In addition, the model integrates multimodal information such as market data and applies an adaptive learning rate optimization algorithm for faster convergence. The experimental data comes from 1.2 million financial texts collected from multiple channels between 2020 and 2024. In the sentiment analysis task, the DK-BERT model has an accuracy of 91.2% on the test set, which is 4.3% higher than FinBERT, and the F1 value is 6.7% higher in the long text scenario. In the market prediction experiment, the sentiment analysis results were combined with SVM, LSTM and other models to construct a prediction model. The sentiment factor reduced the mean square error of the yield prediction by 18.5%, the accuracy of the bull market prediction reached 82.3%, and the bear market reached 79.1%. The study shows that the DK-BERT model significantly improves the accuracy of financial text sentiment analysis and market prediction.