NERGL

NERGL finds the named entities in the title of an ukiyo-e print, types them, and grounds each one in a region of the print, or says that it is not depicted. It is a BART encoder-decoder (ku-nlp/bart-large-japanese) that reads the title together with a knowledge stream written by an LLM, and points to one of the VinVL region proposals of the print.

Code: NERGL on GitHub. Data: BohaoWu/Ukiyoe_MNERG.

This checkpoint

nergl_vinvl_f0_s16_bestgmner.bin is the PyTorch state_dict of one run of the main configuration:

• VinVL region proposals (18 boxes), GPT-6.1 Sol knowledge stream, box labels linked to the title
• fold 0 of the five-fold split, seed 16, 50 epochs
• the epoch with the best validation GMNER F1 (epoch 43, validation GMNER F1 69.61)

Use

git clone https://github.com/BohaoWu/NERGL.git && cd NERGL
pip install -r requirements.txt
python3 script/download.py --model     # dataset, images, BART and this checkpoint

Citation

@InProceedings{Wu2026NERGL,
  author    = {Wu, Bohao and Maeda, Akira},
  editor    = {Balke, Wolf-Tilo and Golub, Koraljka and
               Manolopoulos, Yannis and Stefanidis, Kostas and
               Zhang, Zheying and Aalberg, Trond and Manghi, Paolo},
  title     = {NERGL: Named Entity Recognition and Grounding
               with Large Language Models for Ukiyo-E Artworks},
  booktitle = {New Trends in Theory and Practice of Digital Libraries},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  pages     = {3--13},
  isbn      = {978-3-032-06136-2},
  doi       = {10.1007/978-3-032-06136-2_1}
}