camilaalves/romance-coreference-mdeberta-v3-base

Romance Multilingual Coreference Resolution Model

This repository contains the selected checkpoints of the Romance multilingual coreference resolution model developed for the MSc thesis Coreference in Context: Portuguese Evaluation of Multilingual and Monolingual Models.

Model description

The model follows a span-ranking coreference resolution architecture and uses microsoft/mdeberta-v3-base as its pretrained multilingual Transformer encoder.

The model was jointly trained on six Romance-language coreference datasets:

• Coref-PT (Portuguese)
• Spanish AnCora
• Catalan AnCora
• French ANCOR
• French Democrat
• OntoCorefIT (Italian)

The same jointly trained model is evaluated on each dataset. The checkpoint used for each final evaluation was selected on the corresponding development set.

Selected checkpoints

Dataset	Selected epoch	Threshold
Coref-PT	38	0.0
Spanish AnCora	39	0.0
Catalan AnCora	39	0.0
French ANCOR	20	0.0
French Democrat	28	0.0
OntoCorefIT	17	0.0

Because Spanish AnCora and Catalan AnCora selected the same epoch, five unique checkpoint files are provided.

Checkpoint files

• checkpoint_epoch_17.pt
• checkpoint_epoch_20.pt
• checkpoint_epoch_28.pt
• checkpoint_epoch_38.pt
• checkpoint_epoch_39.pt

Usage

These files are PyTorch checkpoints produced by the custom coreference resolution implementation used in the thesis. They are not intended to be loaded directly with AutoModel.from_pretrained().

The corresponding implementation and experimental configuration should be used to reproduce inference and evaluation.

Evaluation

Coreference resolution is evaluated using MUC, B³, CEAF-e and their average, CoNLL F1.

Citation

If you use this model, please cite:

Camila Alves. Coreference in Context: Portuguese Evaluation of Multilingual and Monolingual Models. MSc thesis, University of Porto, 2026.