An Intelligent Sentiment Analysis Model Based on Fine-Tuned BERT Architecture for Context-Aware Opinion Mining and Emotional Polarity Classification

Full abstract

Read the full abstract

The rapid expansion of user generated text through social media, online reviews, and feedback mechanisms has made consequent and rapid development of sentiment analysis as a key challenge in the field of Natural Language Processing (NLP) the social media text, online reviews and feedback. In the automated decision support systems, sentiment analysis, as a part of the system, brand/ product/ service monitoring systems, and public sentiment systems/analyst applications, picking positive/ negative emotion from the text is a must. Although a lot of research has been done, the earlier Machine Learning (ML) and Deep Learning (DL) models are still facing issues of contextual dependence, polysemy, and lack of resilience in various linguistic phenomena. To overcome the above research gaps, this paper presents an adaptive intelligent-aware framework for sentiment analysis using fine-tuned Transformers models/architectures whereby RoBERTa and DistilBERT are suggested as main/primary models. The proposed methodology will include data preprocessing, subword embedding tokenization, and sentiment polarity classification through supervised fine-tuning of the pre-trained language models. The use of RoBERTa will be since RoBERTa is known for its pre-training and contextual representation. DistilBERT will be used to provide the balance of other models, and to allow the framework to be designed for use in time-constrained and low-resource computing environments. The proposed system will be tested using the available datasets for sentiment analysis using the standard metrics, and to evaluate its performance, the accuracy, precision, recall, and the F1 score will be tested. The results show that the fine-tuned RoBERTa model can outperform the baseline BERT-based and traditional deep learning models, in terms of classification accuracy and managing context-dependent sentiment expressions. Even though DistilBERT is less accurate, it has a lower inference time and memory usage, creating a balance between performance and efficiency. The comparative analysis exhibits the proposed models’ scalability and robustness, despite differing lengths of text and sentiment distributions. This study advocates the proven utility of the transformer-based language models for sentiment classification and recognizes RoBERTa and DistilBERT as intelligent, efficient, and accurate sentiment analysis models. The proposed framework can be adapted for future work, such as sentiment analysis for multiple languages or opinion mining in specific domains