GPT-3.5 for ELE Leveling in a University Environment: A Comparative Study of Zero-shot Learning and Fine-tuning

Resumen

The process of leveling foreign students in Spanish as a Foreign Language (ELE in Spanish) courses is essential to ensure instruction that matches their linguistic competencies. Currently, this leveling is carried out through the manual evaluation of initial tests assessing vocabulary, writing, oral comprehension, and reading comprehension, which entails a high workload and delays in classification, especially with a large number of students. This study explores the use of Natural Language Processing (NLP) to automate the assessment of written expression, investigating the application of GPT-3.5, a large language model with extensive knowledge of Spanish, to classify texts according to their ELE level. Two approaches were implemented: Zero-shot Learning (ZSL), where the model receives explicit instructions to identify and justify the classification of texts, and a supervised method, where the model is trained with the CAES corpus, which contains leveled texts from A1 to C1. The evaluation was conducted using texts from a university environment (anonymized for review), and the results show that supervised training significantly improves the model's accuracy, enabling it to capture subtle differences between levels. These findings highlight the need for continued research to optimize automatic classification systems in ELE and their role in leveling non-native students.