spa_mBERT

Model Card for SpaBERT for Semantic Role Labeling (cased)

This model is fine-tuned on a version of multilingual BERT and is one of 24 models introduced as part of this project.

Model Details

Model Description

SpaBERT for Semantic Role Labeling (SRL) is a transformers model, leveraging mBERT's extensive pretraining on 104 languages to achieve better SRL predictions for Spanish. It was fine-tuned on Spanish with the following objectives:

• Identify up to 16 verbal roots within a sentence.
• Identify available arguments and thematic roles for each verbal root.

Labels are formatted as: r#:tag, where r# links the token to a specific verbal root of index #, and tag identifies the token as the verbal root (root) or an individual argument (arg0/arg1/arg2/arg3/argM) and it's thematic role (adv/agt/atr/ben/cau/cot/des/efi/ein/exp/ext/fin/ins/loc/mnr/ori/pat/src/tem/tmp)

• Developed by: Micaella Bruton
• Model type: Transformers
• Language(s) (NLP): Spanish (es)
• License: Apache 2.0
• Finetuned from model: multilingual BERT

Model Sources

• Repository: GalicianSRL
• Paper: To be updated

Uses

This model is intended to be used to develop and improve natural language processing tools for Spanish.

Bias, Risks, and Limitations

The Spanish training set lacked highly complex sentences and as such, performs much better on sentences of mid- to low-complexity.

Training Details

Training Data

This model was fine-tuned on the "train" portion of the SpanishSRL Dataset produced as part of this same project.

Training Hyperparameters

• Learning Rate: 2e-5
• Batch Size: 16
• Weight Decay: 0.01
• Early Stopping: 10 epochs

Evaluation

Testing Data

This model was tested on the "test" portion of the SpanishSRL Dataset produced as part of this same project.

Metrics

seqeval is a Python framework for sequence labeling evaluation. It can evaluate the performance of chunking tasks such as named-entity recognition, part-of-speech tagging, and semantic role labeling. It supplies scoring both overall and per label type.

Overall:

• accuracy: the average accuracy, on a scale between 0.0 and 1.0.
• precision: the average precision, on a scale between 0.0 and 1.0.
• recall: the average recall, on a scale between 0.0 and 1.0.
• f1: the average F1 score, which is the harmonic mean of the precision and recall. It also has a scale of 0.0 to 1.0.

Per label type:

• precision: the average precision, on a scale between 0.0 and 1.0.
• recall: the average recall, on a scale between 0.0 and 1.0.
• f1: the average F1 score, on a scale between 0.0 and 1.0.

Results

Citation

BibTeX:

@inproceedings{bruton-beloucif-2023-bertie,
    title = "{BERT}ie Bott{'}s Every Flavor Labels: A Tasty Introduction to Semantic Role Labeling for {G}alician",
    author = "Bruton, Micaella  and
      Beloucif, Meriem",
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.671",
    doi = "10.18653/v1/2023.emnlp-main.671",
    pages = "10892--10902",
    abstract = "In this paper, we leverage existing corpora, WordNet, and dependency parsing to build the first Galician dataset for training semantic role labeling systems in an effort to expand available NLP resources. Additionally, we introduce verb indexing, a new pre-processing method, which helps increase the performance when semantically parsing highly-complex sentences. We use transfer-learning to test both the resource and the verb indexing method. Our results show that the effects of verb indexing were amplified in scenarios where the model was both pre-trained and fine-tuned on datasets utilizing the method, but improvements are also noticeable when only used during fine-tuning. The best-performing Galician SRL model achieved an f1 score of 0.74, introducing a baseline for future Galician SRL systems. We also tested our method on Spanish where we achieved an f1 score of 0.83, outperforming the baseline set by the 2009 CoNLL Shared Task by 0.025 showing the merits of our verb indexing method for pre-processing.",
}