Year
2025
Role
Dataset construction, model implementation, evaluation design
Stack
- PyTorch
- Transformer
- LSTM
- GRU
- RNN
- sacreBLEU
Sequence modelling
Neural Machine Translation
Built a 12.5M-pair dataset for Vietnamese diacritic restoration and implemented RNN, GRU, LSTM and Transformer seq2seq models from scratch in PyTorch.
- Built the first large-scale Vietnamese diacritic restoration dataset (12.5M sentence pairs without → with diacritics).
- Implemented Seq2Seq models (RNN, GRU, LSTM, Transformer) from scratch using PyTorch.
- Designed and conducted comprehensive model evaluation with BLEU and ChrF++ metrics, achieving BLEU 81.31 and ChrF++ 88.57 with Transformer, significantly outperforming RNN and GRU baselines.
- Contributed across data construction, model development, and performance evaluation, leading to a robust prototype system for Vietnamese text restoration.
- BLEU
- 81.31
- Transformer
- ChrF++
- 88.57
- Transformer
- Sentence pairs
- 12.5M
- without → with diacritics
