Multilayer-DIET: Multilayer Dual Entity and intent Transformer Classifier

Intent recognition and entity extraction are two classic natural language processing tasks that form the foundation for building intelligent conversational AI frameworks. The development of related technologies has evolved from rule-based templates and statistical learning to deep learning and the current stage of pre-training models. The joint model of intent recognition and entity extraction is an advanced technology aimed at integrating both tasks into a single model for simultaneous execution. In this technological direction, we propose a novel joint model called Multilayer-DIET, which is capable of simultaneously performing multi-layer intent recognition and multi-layer entity extraction. Through hard parameter sharing in multi-task learning and sequential conditional random fields, the model achieves the joint classification and extraction of multiple layers of intents and entities, thus expanding its functionality. Practical evidence demonstrates that this approach not only equips the model with the capability to handle multi-layer intents and entities but also simplifies the complex processing flow of traditional models for multi-layer tasks, resulting in higher accuracy compared to traditional methods.