Modelo de Subtitulación en Español de Imágenes con Gráficos Estadísticos de Tipo Pastel y Barras ajustando el modelo preentrenado UniChart
The purpose of Chart Captioning is to facilitate the understanding of the information present in graphics, using descriptions that explain or describe the data in a textual manner, thus allowing access to such information both for people with visual disabilities and for those who prefer text. The ob...
Guardado en:
| Autor principal: | |
|---|---|
| Formato: | bachelorThesis |
| Lenguaje: | spa |
| Publicado: |
2024
|
| Materias: | |
| Acceso en línea: | https://dspace.unl.edu.ec/jspui/handle/123456789/30600 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
Agregar Etiqueta | Sumario: | The purpose of Chart Captioning is to facilitate the understanding of the information present in graphics, using descriptions that explain or describe the data in a textual manner, thus allowing access to such information both for people with visual disabilities and for those who prefer text. The objective of this Curricular Integration Work (TIC) was to adjust the pre-trained UniChart model to generate Spanish subtitles of images with statistical graphs, only pie and bar charts, taking as a reference the phases of the CRISP-ML (Q) process model, which included: data collection, data engineering, model engineering, and model evaluation. In the first phase, data were collected from two sources, Google and the Digital Repository of the National University of Loja, to create a diverse and representative dataset. In the second phase, data engineering was carried out, creating a customized dataset with image and text pairs. The third phase involved making adjustments to the UniChart model. The final phase involved using the ROUGE metric to assess the final adjusted model's ability to generate subtitles; results showed scores of 0.8709 in ROUGE-1, 0.7847 in ROUGE-2, 0.8465 in ROUGE-L, and 0.8506 in ROUGE-Lsum. In addition, an adapted A/B test was performed to test its capability in subtitle generation, identifying important aspects when generalizing new data, such as limitations. The results showed that the creation of a customized dataset in Spanish based on a predefined structure allowed for obtaining suitable captions for images with statistical graphics. Likewise, it was demonstrated that pre-trained models in English can be adjusted for Spanish subtitling, providing a useful tool for the inclusion and accessibility of statistical data. Keywords: Graphics Captioning, CRISP-ML (Q), UniChar, ROUGE. |
|---|