Bio-RoBERTime

This model is a fine-tuned version of PlanTL-GOB-ES/roberta-base-biomedical-clinical-es on the E3C and Timebank datasets.

It achieves the following results on the E3C corpus test set following the TempEval-3 evaluation metrics:

E3C	Strict	Relaxed	type
RoBERTime	0.7606	0.9108	0.8357
Heideltime	0.5945	0.7558	0.6083
Annotador	0.6006	0.7347	0.5598

RoBERTime is a token classification model, it labels each token into one of the 9 posible labels. We follow the BIO label schema, so each class has two posible values: Begining or Interior. For more Details on the implementation and evaluation refer to the paper: "RoBERTime: A novel model for the detection of temporal expressions in Spanish"

Model description

• Developed by: Alejandro Sánchez de Castro, Juan Martínez Romo, Lourdes Araujo

This model is the result of the paper "RoBERTime: A novel model for the detection of temporal expressions in Spanish"

• Cite as:

  @article{sanchez2023robertime,
    title={RoBERTime: A novel model for the detection of temporal expressions in Spanish},
    author={Sánchez-de-Castro-Fernández, Alejandro and Araujo Serna, Lourdes and Martínez Romo, Juan},
    year={2023},
    publisher={Sociedad Española para el Procesamiento del Lenguaje Natural}
  }
  

Intended uses & limitations

This model is prepared for the detection of temporal expressions extension in Spanish. It may work in other languages due to RoBERTa multilingual capabilities. This model does not normalize the expression value. This is considered to be a separate task.

Training and evaluation data

This model has been trained on the Spanish Timebank corpus and E3C corpus

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 72
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 24

Framework versions

• Transformers 4.24.0
• Pytorch 1.12.1+cu113
• Datasets 2.7.0
• Tokenizers 0.13.2