Monolingual Coreference Retraining
Selected monolingual coreference resolution checkpoints from the MSc thesis Coreference in Context: Portuguese Evaluation of Multilingual and Monolingual Models, University of Porto, 2026.
Spanish AnCora — BETO
- Encoder:
dccuchile/bert-base-spanish-wwm-cased - Training: 40 epochs
- DEV-selected epoch: 29
- DEV-selected mention threshold: 0.0
- DEV CoNLL F1: 48.76
- TEST CoNLL F1: 47.74
- Checkpoint:
spanish-ancora-beto/checkpoint_epoch_29.pt
French ANCOR — CamemBERT
- Encoder:
almanach/camembert-base - Training: 40 epochs
- Dataset:
french_ancor_lm - DEV-selected epoch: 35
- DEV-selected mention threshold: 0.0
- DEV CoNLL F1: 30.66
- TEST CoNLL F1: 25.78
- Checkpoint:
french-ancor-camembert/checkpoint_epoch_35.pt
The French low-memory dataset preserves the original flattened token sequences, gold clusters, and token-level speakers, while changing sentence segmentation to avoid excessively long model inputs.
French Democrat — CamemBERT
- Encoder:
almanach/camembert-base - Dataset:
french_democrat_lm - Training: 40 epochs, independently from French ANCOR
- Selected checkpoint: epoch 20
- Selected mention threshold: 0.8
- DEV CoNLL F1: 10.34
- TEST CoNLL F1: 7.69
- TEST MUC F1: 8.16
- TEST B3 F1: 6.30
- TEST CEAF-e F1: 8.62
- Original thesis TEST CoNLL F1: 6.50
- Difference: +1.19 absolute CoNLL F1 points
- Published checkpoint:
french-democrat-camembert/checkpoint_epoch_20.pt
Low-memory sentence segmentation — French experiments
French ANCOR and French Democrat were retrained using the
french_ancor_lm and french_democrat_lm dataset variants.
These variants use shorter sentence segmentation to accommodate GPU memory and model input-length constraints. The original token sequences and gold coreference spans and clusters were preserved.
However, changing sentence boundaries may influence candidate-span generation and predictions. Comparisons with the original thesis results should therefore be interpreted with caution.
No gold coreference annotations were removed or relabelled.
Italian OntoCorefIT — UmBERTo
- Encoder:
Musixmatch/umberto-commoncrawl-cased-v1 - Dataset:
OntoCorefIT - Training: 40 epochs
- DEV-selected epoch: 38
- DEV-selected mention threshold: 0.0
- DEV CoNLL F1: 72.86
- TEST MUC F1: 78.82
- TEST B3 F1: 69.72
- TEST CEAF-e F1: 69.36
- TEST CoNLL F1: 72.63
- Original thesis TEST CoNLL F1: 73.15
- Selected checkpoint:
italian-ontocorefit-umberto/checkpoint_epoch_38.pt
Catalan AnCora — RoBERTa-ca v2
- Encoder:
projecte-aina/roberta-base-ca-v2 - Dataset:
catalan_ancora - Training: 40 epochs
- DEV-selected epoch: 28
- DEV-selected mention threshold: 0.0
- DEV CoNLL F1: 53.00
- TEST MUC F1: 52.31
- TEST B3 F1: 46.76
- TEST CEAF-e F1: 48.78
- TEST CoNLL F1: 49.28
- Original thesis TEST CoNLL F1: 33.26
- Selected checkpoint:
catalan-ancora-roberta-ca-lowmem/checkpoint_epoch_28.pt
Catalan memory-constrained configuration
The original-parameter retraining attempt failed with CUDA out-of-memory during span-pair scoring. The successful run retained the dataset, encoder, model implementation, 40 training epochs, maximum sequence length of 384, and maximum span width of 20.
The following candidate limits were reduced:
| Parameter | Original configuration | Retraining |
|---|---|---|
| Top-span ratio | 0.4 | 0.1 |
| Maximum top spans | 3900 | 800 |
| Maximum antecedents | 50 | 10 |
These changes reduce GPU memory requirements but may affect candidate selection and coreference predictions. Therefore, the original and retrained TEST results are not strictly like-for-like comparisons. No Catalan sentence resegmentation or gold annotation changes were introduced in this retraining.
Evaluation
The selected checkpoints and thresholds were determined using the development sets. Final evaluation was carried out on the test sets using the CorefUD scorer with exact matching.
CoNLL F1 is the average of MUC, B-cubed, and CEAF-e F1.
Usage and reproducibility
These are custom PyTorch span-coreference checkpoints. They are not directly loadable using AutoModel.from_pretrained().
Use the corresponding coreference implementation and prediction scripts: https://github.com/camiladuart/coref
The source code and training configurations are versioned on GitHub.
Related model releases
- Corrected Portuguese Coref-PT: https://huggingface.co/camilaalves/bertimbau-corefpt-cp1252
- Corrected Romance multilingual model: https://huggingface.co/camilaalves/romance-coreference-mdeberta-v3-corefpt-fixed
Citation
Camila Alves (2026). Coreference in Context: Portuguese Evaluation of Multilingual and Monolingual Models. Master's thesis, University of Porto.