sprakbanken-se/portbert-portuguese_CEFR-classification

Model Card for Model ID

This model was developed as part of the Computational SLA working group at Språkbanken Text. It takes essays written in Portuguese by second language learners and assigns them one of the CEFR levels.

This model was trained following the process described in this paper using the UniversalCEFR version of the COPLE2 dataset. Feel free to check it out for more in-depth information.

Model Details

Model Description

• Developed by: Språkbanken Text, as part of the Computational SLA group
• Shared by: Ricardo Muñoz Sánchez (rimusa)
• Model type: BERT for text classification
• Language(s): Portuguese
• License: GPL-3.0
• Finetuned from model: PortBERT/PortBERT_base

Uses

Direct Use

This model is meant for building demos and tools that display an approximate CEFR level of language learner texts. It is important to note that the predictions from this model should be taken as illustrative rather than as autorithative.

Out-of-Scope Use

This model should not be deployed on high-stakes situations, such as actual language assessment or decision-making regarding migration, work, education, etc. It has relatively low performance compared to what would be needed for such situations, not to speak of potential issues regarding accountability.

Bias, Risks, and Limitations

The model has been trained in a heterogeneous dataset of Portuguese language learner essays. While this exposed the model to a variety of contexts, it also means that there might be biases in terms of topics and format.

We are currently studying the impact that the essay authors' first language(s) has on these models. This model card will be updated once we have more results on this regard.

Model Card Authors

Ricardo Muñoz Sánchez (rimusa)

Model Card Contact

For more information about the model or the present Model Card, you can reach out to:

• Ricardo Muñoz Sánchez ([mailto:ricardo.munoz.sanchez@gu.se])
• Elena Volodina ([mailto:elena.volodina@gu.se])