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 code
• test.py - test code with csv export for further usage in the analytics.py
• analytics.py - code that extracts interesting information about the trained model, it uses the csv file exported using test.py code
• models/best_model - this folder contains the config and model definition for a training setting, that achieved ~86% validation accuracy