Lightweight and Explainable Neural Models for Multilingual Movie Script Certification

Automated film certification remains an underexplored regulatory challenge, requiring scalable yet transparent models capable of handling full-length multilingual scripts. This paper presents a unified framework that delivers lightweight, explainable, and calibrated neural classifiers for multilingual movie script certification in English, Hindi, and Marathi. Unlike prior studies that operate on short snippets or monolingual text, our approach models entire scripts through chunk-level transformer encoding, knowledge distillation, and file-level temperature calibration, coupled with explainability-guided rule mapping for interpretable decision refinement. The proposed pipeline systematically integrates six stages—baselines, teacher modeling, distillation, calibration, explainability, and rule enrichment—yielding a compact yet trustworthy system. Experiments show that the distilled students retain over 85% of teacher accuracy while being 3× smaller, and temperature scaling substantially improves reliability (English Expected Calibration Error 0.303→0.086, Brier 0.684→0.540). Faithfulness analysis using deletion Area Under Curve confirms interpretable token attributions (0.157, 0.239, and 0.258 for English, Hindi, and Marathi respectively). Moreover, rule integration improves accuracy (English 0.581→0.587) while offering human-auditable rationales. All models are deployment-feasible, exported to ONNX/TorchScript with 3.5× compression (545 MB→150 MB) and no performance loss. Together, these results establish a reproducible, end-to-end pipeline that works multilingual long-document modeling, calibration, and interpretability for film certification—advancing trustworthy Artificial Intelligence in regulatory Natural Language Processing. To our knowledge, this is the first work to build a unified, multilingual, and explainable pipeline for movie script certification using full-length scripts across MPAA and CBFC regulatory settings.