Intent Classification Methods for Human Resources Chatbots

Resumen

With the rapid development of artificial intelligence, conversational agents have become prevalent in most mainstream platforms. To accomplish the user’s goal, the system needs to determine their intention, detect emotions and extract key entities from the conversational utterances. This paper presents a comprehensive analysis of intent classification techniques applied to Human Resources chatbots. First, unsupervised learning methods are explored, using pre-trained encoders and Zero-Shot Classification pipelines. Then, we investigate supervised approaches and Large Language Models using Retrieval Augmented Generation. Finally, we propose a two-stage retrieval pipeline that trains a coarse-grained classifier and uses its prediction to retrieve the fine-grained intent with Approximate Nearest Neighbor search. Results reveal that while fully supervised methods yield superior models, unsupervised approaches demonstrate competitive performance, but have the advantage of allowing new intents to be added without requiring model retraining. Our proposed two-stage approach combines the benefits of better performance with the added flexibility.