spacyturk Goto Github PK

spaCyTurk is a library providing trained spaCy models for Turkish language.

Trained floret vectors for Turkish

The floret vectors were trained on the deduplicated version of OSCAR-2109 Turkish corpus. The sentence segmented (non-Turkish sentences were removed) and tokenized final corpus has a size of 30GB and 4327M tokens.

For more details, see the article describing the parameter selection and evaluation process.

training parameters: model=cbow, dim=300, minn=4, maxn=6, hashCount=2, minCount=5, ws=5, neg=10, lr=0.05, epoch=5

Two models (tr_floret_web_md, tr_floret_web_lg) are available with bucket sizes of 50000 and 200000 respectively.

Model performances were evaluated in below downstream NLP tasks.

• Named Entity Recognition, NER
• Part of Speech Tagging, POS
• Offensive Language Identificaton, OLI
• Movie Sentiment Analaysis, MSA

Evaluation metrics: micro f1-score for NER, accuracy for POS, macro f1-score for OLI and MSA.

Trained models can be installed directly from Hugging Face Hub. Alternatively, you can install spacyturk from PyPI and download models through its API. This is the recommended way since the downloader performs version compatibility checks.

pip install spacyturk

import spacyturk

# downloads the spaCyTurk model
spacyturk.download("model_name")

# info about spaCyTurk installation and models
spacyturk.info()

# load the model using spaCy
import spacy
nlp = spacy.load("model_name")

Alternatively, download models through CLI

# downloads the spaCyTurk model
python -m spacyturk download model_name