Comparing Few-Shot Learning with LLMs for Efficient Text Classification in Road Maintenance Applications

Abstract

Efficient road maintenance is vital for long-lasting and safe transportation networks, but traditional methods that rely on manual inspection are labour-intensive and error-prone. The integration of Natural Language Processing (NLP) and Large Language Models (LLMs) presents a transformative solution for automating text-based tasks in road maintenance. This study investigates the application of LLM-based text classification models to process unstructured textual data from road maintenance logs, with a focus on resource-constrained scenarios characterised by limited labelled datasets. Two primary approaches were evaluated: traditional fine-tuning and few-shot learning. Using a public dataset from New York City roadwork inspections containing 83 distinct classes, we performed extensive model comparisons. Pre-trained transformer models Llama and BERT were fine-tuned to achieve baseline performance. Additionally, a few-shot learning method, SetFit, was employed to address data scarcity through efficient task adaptation using minimal labelled examples. Results showed SetFit outperformed fine-tuning in low-resource scenarios, achieving high accuracy and F1 scores with as few as 1-10 examples per class, reducing annotation efforts. This highlights the potential of few-shot learning for real-world deployment. Future work will address scalability, multimodal data fusion, and integration into predictive maintenance.