A deep learning solution for intent classification problem in the domain of Polish language.
- HerBERT model on HuggingFace: https://huggingface.co/allegro
Intent classification is done by using a HerBERT as a backbone language model to understand the utterances. On top of that, a custom intent classification head is used to output the utterance intent.
MASSIVE: https://arxiv.org/pdf/2204.08582.pdf
train.py- train codetest.py- test code with csv export for further usage in theanalytics.pyanalytics.py- code that extracts interesting information about the trained model, it uses the csv file exported usingtest.pycodemodels/best_model- this folder contains the config and model definition for a training setting, that achieved ~86% validation accuracy