rbg-research/EMNLP-2025

RBG-AI: Benefits of Multilingual Language Models for Low-Resource Languages # EMNLP-2025 RBG-AI: Benefits of Multilingual Language Models for Low-Resource Languages ## Contents | S.No | Division | Description | Link | |:----:|:---------------------:|:----------------------------------------------------------------------------:|:-----------------------------------------------------------------------------------------------:| | 1 | Model Selection | Qualitative benchmarking against multilingual translation (NLLB, MADLAD-400) | Link | | 2 | Zero-Shot Inference | Zero-Shot Inference with MADLAD-400 | Link | | 3 | Result and Analysis | Analysis on the results of WMT 2025 Training Corpus | Link | ```commandline . ├── data │   ├── README.md │   └── WMT_INDIC_MT_Task_2025.zip ├── LICENSE ├── notebooks │   ├── 0.Model-Selection.ipynb │   ├── 1.Zero-shot-Inference-Training.ipynb │   ├── 2.Observations.ipynb │   └── translation_analysis_plots ├── README.md ├── requirements.txt └── results ├── madlad_translation_batch.csv └── nllb_translation_eval_combined.csv ``` ## Observations | BLEU Score w.r.t Source Family | BLEU Score w.r.t Target Family | |:---------------------:|:--------------------------------------------------------------------| | | | * The analysis by language family reveals additional nuances in the zero-shot translation performance patterns. As a source language family, Indo-Aryan achieves the highest mean BLEU score, with a wide spread indicating both strong successes (e.g., Assamese→English) and weaker cases. * Tibeto-Burman and Indo-European source families achieve more modest but consistent BLEU scores, while Austroasiatic scores are notably lower, reflecting the challenges in leveraging these …