LEVERAGING DEEP LEARNING AND WORD EMBEDDING FOR FINANCIAL SENTIMENT ANALYSIS WITH CHATGPT-BASED DATA AUGMENTATION

Abstract

Artificial Intelligence (AI) has a wide range of applications in business and finance, including the analysis of textual and unstructured documents and data. Financial Sentiment Analysis (FSA) is crucial for understanding the emotional tone and attitudes expressed in financial data presented in textual form. Financial sentiment classification models enable the automatic analysis of large market data volumes, identifying sentiment to recognize potential risks and opportunities, and can be useful tools for decision support for investors, managers, and traders. This study explores the effectiveness of combining word embedding techniques, such as Word2Vec and FastText, with deep learning architectures, including Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Unit (BiGRU), to improve financial sentiment classification accuracy. A key challenge addressed is class imbalance in the financial sentiment dataset. To mitigate this, ChatGPT is used for data augmentation, generating synthetic samples to balance the dataset and enhance model performance. The proposed methodology is tested using the Financial PhraseBank dataset, where the class imbalance ratio between negative, positive, and neutral classes is 1:2.24:4.75. Simulation results suggest that augmenting the class with the fewest instances improved model performance. The cosine similarity between the actual and augmented samples is 0.908. The best results are achieved using a model based on Word2Vec vectorization with BiGRU architecture, trained on both original and augmented data. The accuracy is 0.81, and the macro average F1 score is 0.79. The average F1 score for the minority classes increased by 12.2% compared to the model trained on the original data.