---
language: es
license: cc-by-4.0
tags:
- spanish
- roberta
- bertin
pipeline_tag: text-classification
widget:
- text: Las Líneas de Nazca son una serie de marcas trazadas en el suelo, cuya anchura oscila entre los 40 y los 110 centímetros.
- text: Hace mucho tiempo, en el gran océano que baña las costas del Perú no había peces.
---

# Readability ES Sentences for three classes

Model based on the Roberta architecture finetuned on [BERTIN](https://huggingface.co/bertin-project/bertin-roberta-base-spanish) for readability assessment of Spanish texts.

## Description and performance

This version of the model was trained on a mix of datasets, using sentence-level granularity when possible. The model performs classification among three complexity levels:
- Basic.
- Intermediate.
- Advanced.

The relationship of these categories with the Common European Framework of Reference for Languages is described in [our report](https://wandb.ai/readability-es/readability-es/reports/Texts-Readability-Analysis-for-Spanish--VmlldzoxNzU2MDUx).

This model achieves a F1 macro average score of 0.6951, measured on the validation set.

## Model variants

- [`readability-es-sentences`](https://huggingface.co/hackathon-pln-es/readability-es-sentences). Two classes, sentence-based dataset.
- [`readability-es-paragraphs`](https://huggingface.co/hackathon-pln-es/readability-es-paragraphs). Two classes, paragraph-based dataset.
- `readability-es-3class-sentences` (this model). Three classes, sentence-based dataset.
- [`readability-es-3class-paragraphs`](https://huggingface.co/hackathon-pln-es/readability-es-3class-paragraphs). Three classes, paragraph-based dataset.

## Datasets

- [`readability-es-hackathon-pln-public`](https://huggingface.co/datasets/hackathon-pln-es/readability-es-hackathon-pln-public), composed of:
  * coh-metrix-esp corpus.
  * Various text resources scraped from websites.
- Other non-public datasets: newsela-es, simplext.

## Training details

Please, refer to [this training run](https://wandb.ai/readability-es/readability-es/runs/1qe3kbqj/overview) for full details on hyperparameters and training regime.

## Biases and Limitations

- Due to the scarcity of data and the lack of a reliable gold test set, performance metrics are reported on the validation set.
- One of the datasets involved is the Spanish version of newsela, which is frequently used as a reference. However, it was created by translating previous datasets, and therefore it may contain somewhat unnatural phrases.
- Some of the datasets used cannot be publicly disseminated, making it more difficult to assess the existence of biases or mistakes.
- Language might be biased towards the Spanish dialect spoken in Spain. Other regional variants might be sub-represented.
- No effort has been performed to alleviate the shortcomings and biases described in the [original implementation of BERTIN](https://huggingface.co/bertin-project/bertin-roberta-base-spanish#bias-examples-spanish).

## Authors

- [Laura Vásquez-Rodríguez](https://lmvasque.github.io/)
- [Pedro Cuenca](https://twitter.com/pcuenq)
- [Sergio Morales](https://www.fireblend.com/)
- [Fernando Alva-Manchego](https://feralvam.github.io/)