edchengg/transfusion

Translation and Annotation Fusion for Cross-lingual transfer in low-resource languages # TransFusion (BERT-based Model) ## Data TransFusion Data ## Code Train transfusion on masakhaner with mdebertav3 (code/). ```bash bash train_masakha_ner_mdeberta.sh ``` ## EasyProject Data Generation Translation data can be found in the Google Drive. Run the following code to project labels from translation data in 'conll_nllb_3B_ft.pkl': ``` python decode_marker_conll.py ``` - We use nllb-200-3.3B-easyproject translation model that is fine-tuned to preserve special markers during translation on 5 high-resource languages (Germany, Spanish, Dutch, Chinese, Arabic). The fine-tuning data is automatically constructed (see details in Section 4.1 in EasyProject). You can access fine-tuning data and training script at EasyProject Github. ## Funding Acknowledgment This material is based in part on research sponsored by IARPA via the BETTER program (2019-19051600004).