SENTIMENT ANALYSIS FRAMEWORK: DOMAIN-SPECIFIC CLASSIFICATION THROUGH FINE-TUNED BERT

Abstract

The exponential growth of digital retail platforms made huge volumes of unstructured con sumer feedback available. Automated opinion mining has become a vital asset of modern business analytics. Traditional machine learning methods often do not understand the context of language and depend heavily on manually prepared text. This dissertation provides a rigor ous comparative study of traditional frequency based statistical models against modern neural network architectures for binary sentiment prediction. The experimental setup is derived from the Amazon Musical Instruments review dataset. Since there were no pre-defined sentiment labels for the raw data, the target variables were created algorithmically. This was done by combining Natural Language Toolkit Part of Speech tagging with the SentiWordNet lexical database to calculate net polarity scores . Then, a Logistic Regression algorithm was train on the sparse term frequency inverse doc frequency matrices to establish a baseline classification standard. To overcome the structural limitations of the baseline framework, a Bidirectional Encoder Representations from Transformers was deployed and trained the same dataset using advanced graphics processing units. The classical Logistic Regression classifier was restricted to an accuracy of seventy percent. It was very limited in performance as it was not capable of handling sequential grammar. On the other hand, the neural network model achieved an impressive accuracy of ninety four percent. The bidirectional self attention mechanism allowed the model to capture deep semantic meaning and complex domain vocabulary. Crucially, the neural architecture was extremely robust to errors in the training data and successfully identified real human sentiment that the automated dictionary missed. This study presents a compelling case that deep transfer learning is a far superior, highly scalable and highly accurate solution for processing complex consumer language.