dv-muril

This is an experiment in transfer learning, to insert Dhivehi word and word-piece tokens into Google's MuRIL model.

This BERT-based model currently performs better than dv-wave ELECTRA on the Maldivian News Classification task https://github.com/Sofwath/DhivehiDatasets

Training

• Start with MuRIL (similar to mBERT) with no Thaana vocabulary
• Based on PanLex dictionaries, attach 1,100 Dhivehi words to Malayalam or English embeddings
• Add remaining words and word-pieces from BertWordPieceTokenizer / vocab.txt
• Continue BERT pretraining

Performance

• mBERT: 52%
• dv-wave (ELECTRA, 30k vocab): 89%
• dv-muril (10k vocab) before BERT pretraining step: 89.8%
• previous dv-muril (30k vocab): 90.7%
• dv-muril (10k vocab): 91.6%

CoLab notebook: https://colab.research.google.com/drive/113o6vkLZRkm6OwhTHrvE0x6QPpavj0fn