Modelo QA basado en DistilBERT para responder a preguntas sobre el contenido extraído de tareas académicas de la carrera de Computación de la UNL.

The adaptation of pre-trained question-answering (QA) models is an essential task so that they can be implemented in different scenarios. The objective of this research is to obtain the value of the rough metric by applying the Fine-Tuning technique to the DistilBERT model to answer questions about the content extracted from academic tasks of the Computer Science Department of the National University of Loja. To develop this work, the CRISP-ML (Q) methodology was used as a reference framework, making use of its first four phases, in which the following was done: a compilation of 30 academic tasks obtained from 6 different subjects, from which 80 questions about their content were generated through crowdsourcing, which served as a basis for creating a dataset in SQuAD1.0 format with 1410 data, of which 800 were generated through paraphrasing and the Few-shot learning approach, and the remaining 610 with the direct contribution of the author. This dataset was divided into 90% for training (train) and 10% for evaluation (test), with an additional subdivision of the train set (75% for train and 25% for validation). Having the data prepared, DistilBERT hyperparameters were adjusted to train four different models using TensorFlow on the Google Colab platform with the GPU T4 runtime environment, selecting the best model based on its level of response extraction and F1-score. Once the QA model was chosen, an evaluation was performed using the ROUGE metric, including A/B testing. The QA model was deployed in Hugging Face and achieved an accuracy of 86.93% during its training with 51 epochs, a learning rate of 1 -5, and a batch size of 32, which through evaluation achieved a maximum F-measure in ROUGE-L of 60.96. These values demonstrate the importance of applying Fine-Tuning in the development of QA models for specific contexts. Keywords: QA model, DistilBERT, SQuAD 1.0 dataset, CRISP-ML(Q), ROUGE

• Hovedforfatter: Jiménez Merino, Edy Francisco (author)
• Format: bachelorThesis
• Sprog: spa
• Udgivet: 2024
• Fag: MODELO QA; DISTILBERT; DATASET SQUAD1.0; CRISP-ML(Q); ROUGE
• Online adgang: https://dspace.unl.edu.ec/jspui/handle/123456789/30620