Python library for word stress detection
Install
pip install russ
or
git clone https://github.com/IlyaGusev/russ
cd russ
pip install -r requirements.txt
python setup.py
Usage
Colab: link
from russ.stress.predictor import StressPredictor
model = StressPredictor()
model.predict("корова")
>>> [3]
Dataset
- Train/val/test split: link
Metrics
- CPU, AMD EPYC 7282, batch size = 2048: 1495 μs for 1 word
- GPU, 1x RTX 3090, batch size = 2048: 173 μs for 1 word
- Test accuracy: 89.73%
- All accuracy: 97.95%
Commands
download.sh
Script for downloading datasets:
- ru_custom.txt: 885 words
- zaliznyak.txt: 86839 lexemes
- espeak.txt: 804909 words
- ruwiktionary-20221201-pages-articles.xml: articles from ruwiktionary, update to a new dump
scripts/prepare_data.py
Preparing data for training
| Argument | Default | Description |
|---|---|---|
| --wiktionary-dump-path | None | path to downloaded wiktionary dump |
| --espeak-dump-path | None | path to espeak dump |
| --custom-dict-path | None | path to file with custom words |
| --inflected-dict-path | None | path to downloaded file with lexemes |
| --inflected-sample-rate | 0.3 | part of inflected dict to use |
| --split-mode | lexemes | how to split into train, val and test files: "sort", "lexemes" or "shuffle" |
| --all-path | data/all.txt | path to output train file |
| --train-path | data/train.txt | path to output train file |
| --val-path | data/val.txt | path to output validation file |
| --test-path | data/test.txt | path to output test file |
| --val-part | 0.05 | part of validation file |
| --test-part | 0.05 | part of test file |
| --lower | Fasle | lowercase all words |
