A joint learning classification for intent detection and slot filling from classical to deep learning: a review

A joint learning classification for intent detection and slot filling from classical to deep learning: a review

연구 분야: Artificial Intelligence

학회: Neural Computing and Applications

In a dialogue system, the natural language understanding component plays a critical role in enabling effective communication. The two core tasks within this component are intent detection and slot filling. Intent detection identifies the user’s goal, while slot filling extracts relevant information to fulfill that goal. Traditionally, these tasks were approached separately or in a pipeline-like manner. However, recent studies have emphasized the benefits of solving them jointly due to their natural interconnections. This study explores the evolution of joint learning models for intent detection and slot filling from 2008 to 2024, covering both classical and deep learning approaches. It discusses the limitations of classical models, which led to the rise of deep learning techniques, and introduces a new taxonomy for joint learning classifying joint learning architectures. Key benchmark datasets, evaluation metrics, and the challenges faced by joint models are also analyzed. Finally, the review identifies open research questions and proposes directions for future exploration in this field.