roberta-crf-boundary-news-ner

RoBERTa + CRF news NER model, extended from [Souvikbasur/roberta-crf-news-ner]({"https://huggingface.co/" + CHECKPOINT_REPO}) with:

• A boundary-guided interaction module (auxiliary start/end token classifiers gating the shared representation before the CRF classifier), inspired by Zheng et al. (2019, EMNLP-IJCNLP) and Gao et al. (2022).
• Boundary-smoothed multi-task loss (Zhu & Li, 2022, ACL) combining the CRF negative log-likelihood with the boundary auxiliary loss.
• FGM adversarial training on the token embeddings (Miyato et al., 2017).
• An optional confidence-based boundary-correction post-processing step at inference.

This is a fine-tuning extension of an existing checkpoint, evaluated with an ablation against the base CRF model — see the linked paper/repo for the actual measured numbers. Do not assume improvement without checking the evaluation cells in the training notebook.

Trained on the same 17-category custom news NER dataset and train/val/test split as Souvikbasur/roberta-crf-news-ner.

Requires the custom RobertaCRFBoundaryForNER class (see the training notebook) to load — a bare state_dict isn't loadable via AutoModel.from_pretrained alone.