NorBERT Multitask Classifier

A multitask fine-tune of NbAiLab/nb-bert-base that classifies Norwegian customer service messages across three dimensions in a single model pass.

Background

Customer service teams handling Norwegian-language messages need to triage incoming requests quickly — routing by urgency, understanding tone, and categorizing the issue type. Most approaches train separate models for each task. This project handles all three with one fine-tuned BERT model, reducing latency and infrastructure overhead.

What it classifies

Task	Labels
Sentiment	negativ, nøytral, positiv
Priority	lav, normal, høy
Category	leveringsproblem, feil vare, fakturaspørsmål, generell henvendelse, reklamasjon

Approach

Training data was generated synthetically using structured LLM prompts. This was a deliberate choice — real customer service data carries privacy constraints, and synthetic data allowed rapid iteration on label balance and edge cases. The trade-off is reduced real-world noise; a production deployment would benefit from fine-tuning on actual labelled messages.

Example output

Input:  Jeg har blitt trukket dobbelt på kortet. Kan dere rydde opp?
Output: kategori=fakturaspørsmål  prioritet=høy  sentiment=negativ

Project structure

norbert-multitask-classifier/
├── data/
│   └── norbert_synthetic_sample.jsonl
├── examples/
│   └── sample_predictions.txt
├── src/
│   ├── model.py          # BERT with 3 classification heads
│   ├── prepare_data.py   # Tokenization pipeline
│   ├── train_all.py      # Multitask fine-tuning
│   └── predict_all.py    # CLI prediction
├── requirements.txt
└── README.md

Usage

pip install -r requirements.txt
python src/prepare_data.py   # tokenize dataset
python src/train_all.py      # fine-tune model (~600MB output)
python src/predict_all.py    # run predictions

Note: trained weights are not included. Run train_all.py to generate, or save your own with model.bert.save_pretrained().

License

MIT