Cross Dataset Fairness Evaluation of Transformer Based Sentiment Models

Abstract

With the growing exploration of Natural Language Processing (NLP) systems in decision-making environments, it is essential to evaluate technical and ethical aspects of the dataset and the NLP model to improve fairness. To assess fairness, the thesis examines demographic imbalances in sentiment classification models by evaluating transformer-based models fine-tuned on the Stanford Sentiment Treebank version 2 dataset (SST-2) against the demographically annotated Comprehensive Assessment of Language Model dataset (CALM). This work identifies performance disparities in sentiment prediction across demographic groups by examining sensitive attributes such as gender and race. The study evaluates both the RoBERTa and MentalBERT transformer models using a complete set of fairness metrics consisting of Statistical Parity Difference (SPD), Equal Opportunity Difference (EOD), False Positive Rates (FPR), False Negative Rates (FNR), Jensen-Shannon Divergence (JSD), and Wasserstein Distance (WD). The analysis examines both group-vs-rest and pairwise subgroup comparisons, including gender and ethnicity. Results show that applying adversarial mitigation reduced fairness disparities across demographic subgroups, with the most notable improvements observed for non-binary and Asian users. The observed disparities emphasize the challenge of reducing performance gaps across demographic subgroups in sentiment classification tasks. The thesis introduces a practical framework for evaluating demographic dis- disparities, extends fairness analysis, and assesses the impact of mitigation techniques in cross-dataset sentiment classification. This research proposes a framework that demonstrates a path toward creating inclusive NLP systems and establishes the groundwork for upcoming ethical Artificial Intelligence (AI) studies.