Predicting English learning performance in vocational colleges using explainable large language models

Abstract

English proficiency prediction is an important task in vocational education because it is helpful in support teaching evaluation, personalized learning and early academic intervention. Traditional methods for predicting student performance mainly use structured educational properties such as test scores, classroom participation and attendance records. Although these methods can achieve certain prediction results, they often pay no attention to the semantic relationships contained in the students’ overall learning portfolios and have limited interpretability. To address this problem, this study proposes an interpretable large language model-based framework to predict English proficiency. Firstly, student learning records are transformed into natural language prompts, and semantic embeddings are learned using large language models. At the same time, structured educational features are standardized and encoded to build a structured representation. Then, a gated multimodal fusion module is designed to adaptively fuse semantic information and structured information. Finally, residual attention multi-layer perceptrons are used to classify the fused features to predict the students’ English proficiency. The experiment was carried out on a dataset of 200 student records from a vocational college for English learning. The results showed that the proposed model achieved an accuracy rate of 92.5% and a weighted F1 score of 91.3%, which is better than several traditional machine learning baseline models. Further analysis showed that the gated fusion mechanism could integrate semantic information and structured information. Feature importance analysis and case studies also shown that the proposed framework could provide interpretable prediction results and personalized learning recommendations.

The authors' abstract, as published at the source. Discover Artificial Intelligence, 2026 · DOI ↗