Intent Recognition for ENT-Related Questions in a Closed-Domain Medical Chatbot using TinyBERT-BiLSTM

Abstract

The field of Otolaryngology (ENT) is critical for addressing common health issues, yet access to specialized medical advice is often hindered by the limited number of specialists, leading to potential delays in patient care. Medical chatbots offer a scalable solution, but their effectiveness is highly dependent on accurately understanding user queries. Conventional chatbots often struggle with the nuances and specific terminology of medical questions. This research proposes a robust intent recognition model for a closed-domain medical chatbot focused on ENT diseases by integrating a lightweight transformer, TinyBERT, with a Bidirectional Long Short-Term Memory (BiLSTM) network. This hybrid model is designed to accurately classify the intent behind user questions, forming the core component of a chatbot that utilizes OpenAI's GPT for response generation. The proposed TinyBERT-BiLSTM model was evaluated on a specialized dataset of ENT-related queries, demonstrating a significant improvement over baseline models. The model achieved a high accuracy of 88% in the intent recognition task, substantially outperforming a general-purpose Large Language Model used for the same task. While the overall accuracy of the chatbot's end-to-end response evaluation was 67.6%, the high performance of the intent recognition module confirms its effectiveness as a critical component for specialized medical chatbots, highlighting the importance of domain-specific models for understanding user intent.

Ähnliche Arbeiten