FairBERT: Enhancing Equity in English Proficiency Testing with Adversarial Training
D. Jaichithra, G R Namita, Malik Bader Alazzam, K S Punithaasree, Al-Hussein Maysir Majid, Ponni Valavan M · 2025
Language proficiency tests are very important in education, employment, and immigration, but traditional assessments usually carry biases due to linguistic and socio-economic factors, thereby making unfair scores for certain demographics. This paper introduces FairBERT as a novel approach to reducing the biases in automated English proficiency assessments. The focus here is on improving the validity of proficiency measurement and minimizing the impact of irrelevant demographic factors. FairBERT makes use of adversarial training in a pre-trained BERT model, estimating the proficiency in a language while ensuring at the same time demographic parity. The model focuses on the proficiency of language rather than relying on the detection and removal of demographic markers by training an adversarial component. The effectiveness of FairBERT is demonstrated by its validation on a learner text corpus with an accuracy rate of 95.5%, as against 82.3% for basic BERT and 75.8% for traditional classifiers. These results demonstrate that FairBERT can offer more accurate, unbiased assessments without giving up on fairness. Thus, by dealing with the biases in English proficiency tests, FairBERT offers an encouraging solution for education, recruitment, and immigration, ensuring a more balanced evaluation process.