| Title |
New Intent Discovery through Multi-view Analysis |
| Authors |
여은기(Eunkee Yeo) ; 이원명(Yuanming Li) ; 구본화(Bonhwa Ku) ; 고한석(Hanseok Ko) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.4.96 |
| Keywords |
New intent discovery; Large language model; Multi-view; Key word |
| Abstract |
With the recent advancements in large language models (LLMs), AI chatbot services such as ChatGPT are being utilized across various domains. These chatbot services rely on understanding the intent of user utterances to provide contextually appropriate responses, making intent classifiers an essential component. However, if existing classifiers fail to recognize new intents that emerge due to the introduction of new services or societal changes, the chatbot's performance may deteriorate, posing the risk of generating responses that do not meet user expectations. This study explores an efficient method for discovering new intents by integrating existing intent data with data containing unknown intents. To achieve this, we propose a "multi-view analysis" technique that analyzes sentences from multiple angles, leveraging not only sentence embeddings but also keywords as critical phrases to better understand sentence intent. Various models suitable for natural language processing were selected as sentence embedding models and comparative experiments were conducted. Experimental results demonstrated superior performance compared to previous studies across three publicly available datasets. |