Abstract
The technique of examining consumer perceptions of a product or service based on reviews is called sentiment analysis. Customer's thoughts are separated into three categories using this natural language processing technique: positive, negative, and neutral. The financial and organizational industries rely heavily on sentiment analysis to make critical business decisions. Conventional sentiment analysis approaches treat the text as a single entity and calculate the opinion based on the text's overall content. As a result, this approach falls short in providing comprehensive details regarding the key elements of the text, including the loan approval process, interest rate, and loan process, all of which are crucial elements in the financial language. To assess each significant component of the financial text independently and offer the sentiment associated with that aspect, aspect-based sentiment analysis is used. Aspect-based sentiment analysis, which can pinpoint each significant element in the financial text and deliver the sentiments independently, is used in this research. To achieve this, the Divide and Conquer kernel support vector machine learning technique is used to develop the BFDCS (BiLSTM-FinBERT-Divide Conquer SVM), which uses pretrained models like BiLSTM and FinBERT. The financial text dataset, which was taken from Kaggle, has two characteristics: sentiments, which might be neutral, negative, or positive. Financial text written by consumers expressing their thoughts regarding the services provided by the financial sector is another feature of this dataset. The text dataset absences a correct format; it is preprocessed using several techniques, including lowercasing, stop word removal, special character removal, and lemmatization. After that, the data is run through the BiLSTM model, which analyses the text in both directions and extracts the key points. This text is then sent through FinBERT, which is trained and specifically intended for financial text and can use CLS to produce a vector for it. The results of these two models are then sent into DCKSVM, which can categorize each aspect's outcome as either positive, negative, or neutral. The performance of the model is then assessed using various metrics, including accuracy, precision, and recall. With 95% accuracy, the proposed approach is shown to be the best model when compared to the existing approaches.