This work is a submission to the Creole MT Creole Language Identification Shared Task conducted as a part of the Eleventh Conference on Machine Translation (WMT ’26) colocated with EMNLP 2026, investigating whether hierarchical classification can improve language identification for closely related French-lexifier Creoles. We compare expert-designed linguistic hierarchies with data-driven hierarchies learned from model confusions and representation similarity, together with both hard and soft routing strategies. All models are built on a shared character-level TF–IDF representation with linear classifiers and are evaluated against a strong flat LinearSVC baseline and the multilingual GlotLID system. The best-performing approach is a data-driven combined hierarchy with soft routing, achieving a Macro-F1 of 87.54 while maintaining one of the lowest Macro False Positive Rates (0.00485). Analysis shows that training data availability remains the strongest predictor of language-level performance, with recall significantly correlated with training set size. Manual inspection further reveals that most remaining errors arise from named entities, lexical borrowing, noisy text, and confusions between closely related Creole varieties. Overall, our results demonstrate that learned hierarchical organization provides a practical and effective alternative to expert-designed taxonomies for low-resource language identification while offering improved balance between classification performance and false-positive control.

The paper is a submission to the WMT 2025 Creole MT Task, aimed at improving translation between Seychellois Creole and English by using data from other closely related Creole languages. The approach treats all Creoles as variations of the same language, allowing the model to share knowledge across them. This strategy led to better translation quality than using separate language settings. The work highlights that related languages can effectively support each other in low-resource settings, while also noting challenges such as data imbalance, differences in language domains, and the absence of human evaluation.