XLM-RoBERTa large for QA on Vietnamese languages (also support various languages)

Overview

• Language model: xlm-roberta-large
• Fine-tune: deepset/xlm-roberta-large-squad2
• Language: Vietnamese
• Downstream-task: Extractive QA
• Dataset: mailong25/bert-vietnamese-question-answering
• Training data: train-v2.0.json (SQuAD 2.0 format)
• Eval data: dev-v2.0.json (SQuAD 2.0 format)
• Infrastructure: 1x Tesla P100 (Google Colab)

Performance

Evaluated on dev-v2.0.json

  exact: 136 / 141
  f1: 0.9692671394799054

Evaluated on Vietnamese XQuAD: xquad.vi.json

  exact: 604 / 1190
  f1: 0.7224454217571596

Author

An Pham (ancs21.ps [at] gmail.com)

License

MIT