snjev310/SrcMix

Repository for the EACL 2026 paper "SrcMix: Mixing of Related Source Languages Benefits Extremely Low-resource Machine Translation". Contains the dataset and code for SrcMix, a controlled source-language mixing approach for extremely low-resource machine translation. # SrcMix: Mixing of Related Source Languages Benefits Extremely Low-resource Machine Translation 📢 **Accepted to EACL 2026 (Findings)** Official implementation of **SrcMix**, a simple and effective training-time strategy for improving machine translation in **Extremely Low-Resource Languages (ELRLs)**. ------------------------------------------------------------------------ ## Motivation Extremely low-resource languages (ELRLs) often suffer from: - Fewer than 6K parallel sentences - Limited supervision - High typological diversity - Weak transfer from massively multilingual models Standard multilingual training (naïve concatenation or many-to-many setups) Often performs poorly in ELRL settings due to: - Negative transfer - Cross-lingual interference - Fragmented decoder supervision **SrcMix** addresses this by introducing multilinguality only on the *source side*, while keeping the decoder specialized to a single target ELRL. ------------------------------------------------------------------------ ## What is SrcMix? Let: - Target ELRL: n1 - Related source languages: n2, n3, ..., nN - High-resource language: H To train H → n1, we mix: H → n1 n2 → n1 n3 → n1 ... nN → n1 Key design choices: - Remove explicit source language identifiers - Force implicit structural transfer - Preserve decoder specialization Unlike naïve multilingual mixing, SrcMix prevents decoder fragmentation and improves performance under extreme data scarcity. ------------------------------------------------------------------------ ## 🧡 First Public Angika MT Dataset This work introduces the **first publicly available machine translation training and evaluation dataset for Angika**, an Indo-Aryan language spoken in eastern India. Dataset available here: huggingface.co The dataset includes: - Parallel training data - Evaluation splits - FLORES-aligned dev/test sets - Clean preprocessing for reproducibility - …