WEIGHTED LOSS STRATEGY FOR BERT-BASED TWITTER SENTIMENT ANALYSIS WITHOUT SYNTHETIC OVERSAMPLING

Abstract

The widespread adoption of ChatGPT has generated extensive public discourse across social media, necessitating robust sentiment analysis to understand collective opinions. Traditional approaches frequently employ the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance; however, its effectiveness on short-text data remains an open question. This study develops an optimized sentiment classification model and evaluates whether competitive performance can be achieved without synthetic data augmentation. The methodology encompasses comprehensive Natural Language Processing (NLP) preprocessing and stratified data partitioning to preserve distributional characteristics. A BERT-base architecture is fine-tuned using a class-weighted Cross-Entropy loss combined with weighted random sampling, deliberately avoiding SMOTE-based oversampling. The model is trained with the AdamW optimizer (learning rate: 3 × 10⁻⁵), batch size 32, and mixed-precision training for four epochs. On 198,639 preprocessed tweets, the proposed approach achieves 93.81% accuracy, with weighted precision, recall, and F1-score of 0.9365, 0.9381, and 0.9380 respectively, outperforming the baseline by 1.75 percentage points. Per-class analysis reveals strong performance for negative (F1-score: 0.96) and positive sentiment (F1-score: 0.94), with lower neutral classification (F1-score: 0.89), attributable to the inherent heterogeneity of neutral expressions. The training-validation gap remains below 5%, consistent with adequate regularization. These findings provide empirical evidence that, within the present experimental configuration, a properly optimized weighted loss strategy offers a viable and computationally efficient alternative to synthetic oversampling for BERT-based Twitter sentiment classification. Further controlled ablation studies and statistical validation are needed to establish generalizability.