LAGIM: A Label-Aware Graph Interaction Model for Joint Multiple Intent Detection and Slot Filling

Multi-intent spoken language understanding joint model can handle multiple intents in an utterance and is closer to complicated real-world scenarios, attracting increasing attention. However, existing research (1) usually focuses on identifying implicit correlations between utterances and one-hot encoding while ignoring intuitive and explicit original label characteristics; (2) only considers the token-level intent-slot interaction, which results in the limitation of the performance. In this paper, we propose a Label-Aware Graph Interaction Model (LAGIM), which captures the correlation between utterances and explicit labels’ semantics to deliver enriched priors. Then, a global graph interaction module is constructed to model the sentence-level interaction between intents and slots. Specifically, we propose a novel framework to model the global interactive graph based on the injection of the original label semantics, which can fuse explicit original label features and provide global optimization. Experimental results show that our model outperforms existing approaches, achieving a relative improvement of 11.9% and 2.1% overall accuracy over the previous state-of-the-art model on the MixATIS and MixSnips datasets, respectively.