A MULTI-METHOD STUDY ON INTENT DETECTION IN HUMAN–MACHINE INTERACTION

Abstract

Intent classification is a core component of human–machine and human–robot interaction systems, playing a decisive role in accurately understanding and responding to human requests. This study presents a comprehensive comparative analysis of major methodological paradigms for intent detection, including traditional machine learning approaches, deep learning models, Transformer-based models with fine-tuning, and large language models (LLMs) leveraging prompting techniques. By evaluating all methods under a unified experimental setup on the same dataset, the study examines their respective strengths and limitations in terms of performance, computational cost, deployment complexity, and generalization capability, with particular emphasis on the zero-shot potential of LLMs in HRI scenarios. Qualitative findings indicate that each class of methods exhibits distinct advantages and is suited to different deployment contexts, ranging from low-cost embedded systems to intelligent cloud-based applications.