JointBERT for Norwegian Feed Orders

A joint intent classification and slot filling / NER model for Norwegian animal feed order queries, fine-tuned from NbAiLab/nb-bert-large.

Task

The model processes Norwegian spoken order queries and:

• Classifies the intent (6 classes): create_order, edit_order, confirm, reject, help, reorder_last
• Extracts named entities (7 entity types, IOB2): PRODUCT, QUANTITY, UNIT, DELIVERY_METHOD, DELIVERY_DATE, ADDRESS, TANK_SILO

Test Set Results

Overall

Metric	Score
NER Precision	95.69%
NER Recall	98.04%
NER F1	96.85%
Intent Accuracy	99.07%
Intent F1	99.04%
Combined F1	97.94%

Per-Entity NER F1

Entity	F1
PRODUCT	96.00%
QUANTITY	98.77%
UNIT	96.97%
DELIVERY_METHOD	100.00%
DELIVERY_DATE	91.30%
ADDRESS	97.25%
TANK_SILO	97.44%

Training

• Base model: NbAiLab/nb-bert-large (1024 hidden, 24 layers)
• Training data: 972 utterances (train + val merged after hyperparameter search)
• Hyperparameter search: Optuna (40 trials), retrained on train+val with best config
• Loss: 0.6 * intent + 0.4 * NER
• Epochs: 15
• Learning rate: 3e-05
• Batch size: 16
• Weight decay: 0.1
• Warmup ratio: 0.2
• Frozen layers: 0 (full fine-tuning)