Parametrically modified BERT for emotion prediction through sentiment analysis

Social networks are increasingly in demand for text mining applications. Text analysis has become a more popular technique used on the internet. Social media platforms provide abundant access to text data, allowing users to post comments, making the analysis of these comments crucial for various business applications. Sentiment analysis (SA) is a subfield of natural language processing that aims to automatically identify and classify opinions expressed in text as positive, negative, or neutral. It plays a crucial role in understanding public opinion, especially when applied to large-scale textual data from social media platforms. However, social media data extraction policies governed by platform-specific APIs, privacy constraints, and data usage limitations pose challenges in acquiring high-quality, representative datasets for research. This study proposes TSAPM-BERT, a parametrically modified BERT-based framework that integrates a weighted attention mechanism, information from the sentiment lexicon and optimization of the learning rate to enhance the classification of sentiment at the aspect level. We evaluate the model on a benchmark emotion dataset, comparing its performance against traditional machine learning and deep learning baselines. Experimental results demonstrate that TSAPM-BERT achieves a training accuracy of 99.37 and testing accuracy of 98.52, outperforming competing methods. This work provides a reproducible framework that bridges the gap between high-accuracy sentiment classification and the practical constraints of social media data usage.