arXiv 2021 · Second author · Jul 2021
Back-Translated Task Adaptive Pretraining
01Background & Goals
- A study that combines translation-based augmentation with task-adaptive pre-training to improve BERT-style models when data is scarce.
02Contributions
Augmentation and experiments
- Implemented back-translation (translate to a pivot language and back), with nucleus sampling for diversity
- Implemented and compared augmentation methods such as EDA, TF-IDF and embedding swapping
- Produced the augmentation seminar video and co-wrote the paper
03Tech Stack
- Framework / Platform
- PyTorch, Fairseq, Hydra
- Methodology
- Back-translation, Task-adaptive pre-training, Data augmentation
04Results
- Released as an arXiv preprint (2021, second author)