AlexisBal/specialized-nmt-framework

Multi-Stage Domain Adaptation for Neural Machine Translation: A Hybrid Learning Framework for Specialized Terminology and Low-Resource Language Pairs # Multi-Stage Domain Adaptation for Neural Machine Translation A hybrid learning framework for specialized terminology and low-resource language pairs, demonstrated on Chinese religious text translation. ## Overview This framework addresses domain adaptation challenges in neural machine translation through a multi-stage computational approach that combines: - **Automated terminological extraction** for all language pairs - **Neural fine-tuning** for mBART-supported modern languages - **Systematic terminology management** across diverse resource scenarios ## Key Features - **Multi-stage processing pipeline**: Dictionary matching → Similarity analysis → Statistical dominance → Proper noun detection - **Hybrid approach**: Pure terminological learning for ancient languages (Hebrew, Greek, Latin) + neural enhancement for modern languages - **Scalable architecture**: Modular design adaptable to other specialized domains - **Comprehensive evaluation**: Multiple metrics including BLEU, ROUGE-L, and COMET scores ## Performance Results - **1.13x to 1.68x improvement** in terminological learning compared to baseline dictionary approaches (see dataset infra) - **Consistent gains** in BLEU/ROUGE/COMET scores for neural-enhanced language pairs - **Effective processing** of both resource-constrained ancient languages and modern languages ## Dataset ### Biblical Corpora | Language Pair | Source Text | Target Text | Units | Batches | |---|---|---|---|---| | Hebrew → Chinese | Leningrad Codex (Genesis) | Chinese Union Version 2010 | 1,533 | 17 | | Greek → Chinese | SBL Greek New Testament | Studium Biblicum Version 2019 | 4,748 | 55 | | Latin → Chinese | Nova Vulgata (Leviticus) | Studium Biblicum Version | 859 | 9 | | English → Chinese | NIV (Gospels + Acts) | Chinese Union Version 2010 | — | — | ### Terminological Extraction Results | Language Pair | Stage 1 (Dict) | Stage 2 (+Sim) | Stage 3 (+Stat) | Stage 4 (+PN) | Total | Total Gain | |---|---|---|---|---|---|---| | H …