Agha Ali Raza

Publications/Efficient Machine Learning: Data Selection, Pruning, Active & Federated Learning

Conference · 2025

To Label or Not to Label: Hybrid Active Learning for Neural Machine Translation

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza

COLING 2025, Abu Dhabi, UAE · CORE A

Summary

Adapting a translation system to a new domain requires human-translated sentences, so it pays to choose carefully which sentences to send for translation. Existing strategies pick either the sentences the model is most unsure about, which can be repetitive, or the most varied ones, which can be trivial. Our method, HUDS, combines the two: it scores sentences by model uncertainty, groups them, and then favors those that add diversity. On multi-domain German-English and French-English data, HUDS outperforms strong active learning baselines, prioritizing diverse, high-uncertainty sentences in the early rounds.

Cite this paper

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza. To Label or Not to Label: Hybrid Active Learning for Neural Machine Translation. COLING 2025, Abu Dhabi, UAE.

@inproceedings{azeemi2025label,
  author = {Abdul Hameed Azeemi and Ihsan Ayyub Qazi and Agha Ali Raza},
  title = {{To Label or Not to Label: Hybrid Active Learning for Neural Machine Translation}},
  booktitle = {COLING 2025, Abu Dhabi, UAE},
  year = {2025},
  url = {https://aghaaliraza.com/publications/2025--COLING--To_Label_or_Not_to_Label-Hybrid_Active_Learning_for_Neural_Machine_Translation.pdf}
}

More on this theme

All 8 works on this theme →