intent_classifier

Korean Intention classifier with pytorch lightning ⚡

About

This is an intention classifier that was learned using 3i4k dataset to identify the intonation intention of Korean.

• implemented using phytorch lighting. ⚡
• trained based on Transformers' distil-KoBERT PLM and has approx. 90% accuracy.
• use it as a web API powerd by flask-restplus.

Installation

• clone
• pip install -r requirements.txt
• download model file
• copy into /model/intent_distilkobert.ckpt

Train

• bash train_distil.sh to use distil-KoBERT model

• bash train.sh to use KoBERT model

Inference

• python app.py to start web server
• open {SERVER_IP}:41063 in web browser
• using swagger UI to try out

• response types
  • FRAG : Fragments
  • STAT : Statements
  • QUES : Questions
  • COMM : Commands
  • RHEQ : rhetorical questions
  • RHEC : rhetorical commands
  • INTO : Intonation-dependent utterances

References

• 3i4k dataset to identify the intonation intention of Korean
• monologg/distil-kobert