MAG: Improved Multi-Intent Detection and Slot Filling Model Based on Mamba

Intention detection and slot filling are two main tasks in the field of natural language understanding in natural language processing. Since the two tasks are highly correlated, the two tasks are often modeled jointly. Because of the high information density of natural spoken language, there may be multiple intentions in one utterance. The joint model of multi-intention detection and slot filling has become the most important task in natural language understanding. Because of the presence of multiple intents, there are often multiple centers of gravity in all sentences, and the task tests the performance of the model more. After learning and understanding Mamba model, the author was attracted by its clever structure and advanced performance. Mamba model has the characteristics of selective input and can adjust model parameters according to the input data, which means that Mamba has unique advantages in the task of intention detection. Taking this as inspiration, the author applied Mamba model to multi-intention detection and slot filling tasks, which improved the quality of semantic feature extraction, improved the accuracy of the model, accelerated the running speed of the model and reduced the space occupation of the model.