metadata

language:
  - multilingual
  - pt
  - en
tags:
  - xlm-roberta-large
  - semantic role labeling
  - finetuned
license: Apache 2.0
datasets:
  - PropBank.Br
  - CoNLL-2012
metrics:
  - F1 Measure

XLM-R large fine-tune in English and Portuguese semantic role labeling

Model description

This model is the xlm-roberta-large fine-tuned first on the English CoNLL formatted OntoNotes v5.0 semantic role labeling data and then fine-tuned on the PropBank.Br data. This is part of a project from which resulted the following models:

For more information, please see the accompanying article (See BibTeX entry and citation info below) and the project's github.

Intended uses & limitations

How to use

To use the transformers portion of this model:

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("liaad/srl-enpt_xlmr-large")
model = AutoModel.from_pretrained("liaad/srl-enpt_xlmr-large")

To use the full SRL model (transformers portion + a decoding layer), refer to the project's github.

Limitations and bias

This model does not include a Tensorflow version. This is because the "type_vocab_size" in this model was changed (from 1 to 2) and, therefore, it cannot be easily converted to Tensorflow.
The English data was preprocessed to match the Portuguese data, so there are some differences in role attributions and some roles were removed from the data.

Training procedure

The model was first fine-tuned on the CoNLL-2012 dataset, preprocessed to match the Portuguese PropBank.Br data; then it was fine-tuned in the PropBank.Br dataset using 10-fold Cross-Validation. The resulting models were tested on the folds as well as on a smaller opinion dataset "Buscapé". For more information, please see the accompanying article (See BibTeX entry and citation info below) and the project's github.

Eval results

Model Name	F₁ CV PropBank.Br (in domain)	F₁ Buscapé (out of domain)
`srl-pt_bertimbau-base`	76.30	73.33
`srl-pt_bertimbau-large`	77.42	74.85
`srl-pt_xlmr-base`	75.22	72.82
`srl-pt_xlmr-large`	77.59	73.84
`srl-pt_mbert-base`	72.76	66.89
`srl-en_xlmr-base`	66.59	65.24
`srl-en_xlmr-large`	67.60	64.94
`srl-en_mbert-base`	63.07	58.56
`srl-enpt_xlmr-base`	76.50	73.74
`srl-enpt_xlmr-large`	78.22	74.55
`srl-enpt_mbert-base`	74.88	69.19
`ud_srl-pt_bertimbau-large`	77.53	74.49
`ud_srl-pt_xlmr-large`	77.69	74.91
`ud_srl-enpt_xlmr-large`	77.97	75.05

BibTeX entry and citation info

@misc{oliveira2021transformers,
      title={Transformers and Transfer Learning for Improving Portuguese Semantic Role Labeling}, 
      author={Sofia Oliveira and Daniel Loureiro and Alípio Jorge},
      year={2021},
      eprint={2101.01213},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}