Unified BERT-LSTM framework enhances machine learning in fraud detection, financial sentiment, and biomedical classification

Abstract

The current paper proposes a hybrid framework based on the bidirectional encoder representations from transformers (BERT) and long short-term memory (LSTM) networks for classification tasks in three diverse domains: credit card fraud detection (CCFD), financial news sentiment analysis (FNSA), and biomedical paper abstract classification (BPAC). The model leverages the strengths of BERT regarding the learning of contextual embeddings and those of LSTM in capturing sequential dependencies, thus setting the new state-of-the-art performance in each of the three domains. In the CCFD use case, the model was able to achieve an accuracy of 99.11%, considerably outperforming all the competing systems in fraud transaction detection. The BERT-LSTM model achieved a performance of 96.74% for FNSA, improving significantly in sentiment analysis. Finally, the use case of BPAC was robust, with 88.42% accuracy, which clearly classified biomedical abstract sections correctly. It is evident from the findings that this framework generalizes to a wide range of tasks and hence is an adaptable but strong tool in combating challenges of cross-domain classification.