AI-Driven Multilingual Grading Tools for Indian School Boards

Abstract

The burgeoning diversity of India’s student population—speaking hundreds of languages and dialects—poses a formidable challenge to the fairness, consistency, and scalability of board examination grading. Manual evaluation of descriptive answers across multiple languages is labor-intensive, time-consuming, and susceptible to variability arising from individual examiner biases, inconsistent application of rubrics, and differing proficiencies in regional scripts. These factors contribute to delays in result publication, erode confidence in assessment fairness, and place undue stress on educators. In response, recent advances in natural language processing (NLP) and deep learning offer the promise of AI-driven grading systems capable of evaluating open-ended responses at scale, across many languages, with speed and consistency that rival or exceed human examiners. This manuscript presents a comprehensive study on the design, development, and evaluation of an AI-driven multilingual grading tool tailored for Indian school boards. Leveraging a transformer-based architecture fine-tuned on a corpus of 50,000 anonymized board-exam scripts in English, Hindi, Tamil, Marathi, and Bengali, our prototype generates numeric scores mapped to established board rubrics. We conducted a mixed-methods evaluation involving: (1) a blind grading comparison on 1,000 held-out responses— each graded by two human examiners and the AI system—to assess inter-rater reliability using Cohen’s κ and Pearson’s correlation; (2) a structured survey of 200 educators and examiners across five states to gauge perceptions of consistency, fairness, turnaround time, and willingness to adopt AI support; and (3) semi-structured interviews with board officials and senior teachers to explore usability, transparency, and governance concerns. Quantitatively, the AI tool achieves a Cohen’s κ of 0.82 and Pearson’s r of 0.88 with the human consensus score, indicating strong agreement. Performance remains robust across languages (κ ≥ 0.78), though dialectal variations in Tamil and Marathi yield marginally lower alignment. Qualitatively, 74% of surveyed participants express readiness to integrate AI-assisted grading— provided mechanisms for human review and explainability exist—while 68% acknowledge persistent inconsistencies in peer-based manual grading. The AI system slashes mean grading time from 4.5 minutes per script to 0.5 minutes, projecting a 90% reduction in overall evaluation workload. We identify key technical and operational challenges: OCR errors on handwritten or non-standard scripts, model biases toward data-rich dialects, and the necessity for continuous retraining to prevent performance drift. Policy and governance recommendations emphasize student data privacy under India’s Personal Data Protection framework, an ethics oversight committee for AI deployment, examiner training modules for the AI interface, and periodic bias audits. Our study demonstrates that contextualized AI grading can enhance equity, efficiency, and transparency in India’s high-stakes examinations, laying the groundwork for broader multilingual and multimodal assessment innovations.