This is a two-stage fine-tuned BART-large model for multi-label classification of bibliographic metadata into BK classification codes. The model achieved state-of-the-art performance on the K10plus library catalog dataset.

Performance

Subset Accuracy: 25.7%
Matthews Correlation Coefficient (MCC): 0.498
F1-Score (Micro): 47.9%
F1-Score (Macro): 21.4%
Precision (Micro): 66.1%
Recall (Micro): 37.6%

Training Approach

The model uses a two-stage fine-tuning approach:

Stage 1: Train on parent BK categories (48 labels)
Stage 2: Fine-tune on all BK codes (1,884 labels) using Stage 1 as initialization

This approach outperformed both standard fine-tuning and hierarchical joint training.

Dataset

Source: K10plus German library catalog (2010-2020)
Total Records: 250,831 bibliographic entries
Labels: 1,884 unique BK classification codes
Input Fields: Title, Summary, Keywords, LOC Keywords, RVK codes

Usage

python

import torch
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("mrehank209/bk-classification-bart-two-stage")

# Load model components (see repository for full inference code)
# This model requires custom loading due to the classifier head

Citation

If you use this model, please cite:

bibtex

@misc{bk-classification-bart,
  title={Automatic BK Classification using Two-Stage BART Fine-tuning},
  author={Khalid, M. Rehan},
  year={2025},
  howpublished={Hugging Face Model Hub},
  url={https://huggingface.co/mrehank209/bk-classification-bart-two-stage}
}

Contact

Author: M. Rehan Khalid
Email: m.khalid@stud.uni-goettingen.de
Affiliation: University of Göttingen

License

MIT License