Co-Guiding for Multi-Intent Spoken Language Understanding.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Graph-based models show promise for multi-intent spoken language understanding (SLU) by modeling intent-to-slot guidance.
• Existing methods suffer from unidirectional guidance modeling and homogeneous graph structures, limiting performance.
• Bidirectional inter-correlations between intent and slot are crucial but underexplored.

Purpose of the Study:

• To propose a novel model, Co-guiding Net, that captures bidirectional inter-correlations between intent prediction and slot filling in multi-intent SLU.
• To introduce heterogeneous graph attention networks for improved modeling of semantics-label interactions.
• To enhance performance through supervised contrastive learning strategies.

Main Methods:

• A two-stage framework implementing mutual guidance between intent and slot tasks.
• Two heterogeneous graph attention networks operating on novel semantics-label graphs.
• Co-guiding-SCL Net incorporating single-task and dual-task supervised contrastive learning.

Main Results:

• Co-guiding Net achieves significant improvements on multi-intent SLU tasks, with a 21.3% relative accuracy gain on the MixATIS dataset.
• The model demonstrates strong performance in zero-shot cross-lingual scenarios, improving state-of-the-art by 33.5% on average across 9 languages.
• Heterogeneous graphs and mutual guidance effectively model task interdependencies.

Conclusions:

• The proposed Co-guiding Net effectively models bidirectional task correlations for enhanced multi-intent SLU.
• Heterogeneous graph attention networks and contrastive learning contribute to superior performance.
• The model shows significant potential for cross-lingual SLU applications.