Recent progress in NLP has spurred the development of technologies capable of handling code-switched data. Despite the initiation of Code-Switching research several years ago, progress within the research community has been sluggish. The primary challenges in addressing this issue arise from the insufficient availability of data. A limited number of languages, such as Spanish-English, Hindi-English, or Chinese-English, dominate research and resources in CSW. Nevertheless, numerous countries and cultures that extensively use CSW remain underrepresented in NLP research.
The purpose of this project is to extend code-switching translation capabilities of models on low resource Kazakh - Russian language pair.
Here is Code-Switching examples:
| Sentence |
|---|
| Кадымгы заварка (одноразовый емес) жылы суга шыгынын шыгарып аласыздар и сеуып койсаныз, иыс кетеди |
| Ой тегі, любой адам солай істейді ғой, не болды сонша?! |
| Айына 10-15кг ға дейін арықтау. Фигурная болғың келсе маған кел, Мен көмектесемін |
Wandb project: https://wandb.ai/maksim-borisov-2013/kk-ru-csw
We employ different types of data augmentation to create code-switching (CSW) training data. Namely,
- cs-1: Replace Kazakh word with Russian one in normal form.
- cs-2: Replace Kazakh word with Russian one's stem with ``Kazakh ending.''
- cs-3: Replace Kazakh word with Russian one in random form.
- cs-4: Replace Kazakh word with Russian word aligned using fastalign
- cs-5: Replace Kazakh word with Russian word aligned using SimAlign
- Low resource:
- Adding translated RTC corpus
- Code switching
- A new method of data augmentation for csw
- A method that works on real data
Fine-tuning of existing models
CSW modelling
