COMPARATIVE ANALYSIS OF DEEP LEARNING AND TRANSFORMER-BASED ARCHITECTURES FOR SOCIAL MEDIA SENTIMENT ANALYSIS: A SYSTEMATIC BENCHMARKING STUDY

Abstract

Extensive analysis of sentiments of social media apps has been done in the past but the study of comments given by users is limited due to the challenges involved. This is primarily due to wide variation in language usage in expressing the same sentiments, ambiguity in expression and issues faced in computation. Traditional machine learning (SVM), deep learning (CNN, BiLSTM) and transformer-based methodologies (BERT, RoBERTa, DistilBERT) are used across five social media datasets for analysis. The experiments were performed under controlled noise settings in which lexical perturbations varied from 0 to 30 %. Synthetic sarcasm evaluation and code-switching was also included in the tests. It was found that transformer topologies can attain enhanced noise robustness (NRS = 89.8–90.9%) but also incurred significant processing costs. BERT needed more time for each inference while CNN required much lesser.   The research also introduces a decision matrix for model selection based on operational constraints for comparing theoretical performance with practical implementation. These findings gave useful information for the creation of trustworthy and effective sensitivity analysis system in resource-constrained environment.