Fine-tuning de un Modelo de Lenguaje Largo para la clasificación de Curriculums Vitae

Resumo

The main objective of this work was to classify resumes according to their professional area, an important task in human resources management and personnel recruitment. This research seeks to explore the classification capabilities of Large Language Models (LLM) by performing a comparative analysis versus traditional Machine Learning methods. To achieve this objective, a fine-tuning technique was used on the Long Language Model pre-trained by Google in English called BERT BASE UNCASED using a dataset of more than 3,000 resumes from 25 professional areas and 3 training epochs against traditional Random Forest, SVM, Logistic Regression, and Naive Bayes Multinomial models. The methodology consists of 7 essential stages to adapt a pre-trained model to a specific task, ensuring optimal performance. The research provides a comparative analysis focusing on the metrics Accuracy, F1-score, Precision, and Recall. The most significant results obtained were 83,0% Accuracy and 82.3% Precision for the base Google model, and 82.8% F1-score and 86,2% Recall for the Naive Bayes Multinomial model, revealing that the base Google model performs well in predicting resume classification, while Naive Bayes Multinomial is better at detecting the majority of positive cases. This research not only contributes by showing the performance of MLLs for the classification task in contrast to their traditional Machine Learning peers, but also offers an innovative approach to human resource management and staff recruitment practices.