The use of distilled large language models to determine the sentiment of a text

Abstract

This paper describes a distilled T5-large Transformer-based model for text sentiment analysis optimised for educational applications. The model, created by transferring knowledge from Large Language Models, demonstrates an accuracy of 97.22% on a test sample with a significant reduction in computational requirements. The distillation process, training methodology and experimental results are described. Application scenarios in educational contexts including adaptive learning and personalisation of feedback are proposed. The study shows promise for the use of distilled Large Language Models in resource-limited educational technologies.