Turkish NLP APIs

This is a repo where I plan to put my experimental models behind a REST API to be deployed in Heroku. It is a Python Flask server running behind gunicorn.

Current APIs:

1. /nlp/stemmer/q=**words**: returns the stems of words using a Seq2Seq Neural Model
   1. Params:
      • words (str): contains 1+ words. Special characters like quote, question mark will be discarded
   2. Returns:
      • stemmed words (str). Not wrapped in any JSON or similar for now.
   3. Model details:
      • Characters are one hot encoded.
      • 2 layers of Bidirectional GRUs with 32 hidden units
      • 2 layers of GRUs with 64 hidden units, each using the corresponding state of the encoder
      • 78k params
      • Model is trained in personal Colab (https://colab.research.google.com/drive/1SjuaADaHocVkNfgeY0iYsKj_c1sAm7KK)
   4. Training data details:
      • I once found a Turkish corpus from Zargan containing words, their roots, freqs, etc. Now it's not online. I used 50k word pairs for training & testing.
   5. Example Usages (already deployed):
      • https://cryptic-mountain-83365.herokuapp.com/nlp/stemmer/q=urfanın%20etrafı%20dumanlı%20dağlar
      • https://cryptic-mountain-83365.herokuapp.com/nlp/stemmer/q=kayserilileştiremediklerimizdenmişsiniz

Setup:

1. curl https://cli-assets.heroku.com/install.sh | sh # https://devcenter.heroku.com/articles/heroku-cli#standalone-installation
2. heroku login --interactive
3. python -m venv venv
4. . venv/bin/activate
5. pip install -r requirements.txt
6. heroku create
7. git push heroku main

Local testing:

1. heroku local