Siamese-based metric joint learning for intent detection and slot filling using triplet loss optimization

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Spoken Language Understanding (SLU) systems require accurate intent detection and slot filling.
• Current joint learning models face challenges with minority intents and paraphrased queries.
• Existing methods often rely on token embeddings and cross-entropy loss, neglecting semantic similarity.

Purpose of the Study:

• To propose a Siamese-Based Metric Joint Learning model (SBJLIS) for improved SLU.
• To enhance semantic distance learning and model generalization across diverse user utterances.
• To address limitations in handling class imbalance and linguistic variations in intent detection and slot filling.

Main Methods:

• Developed a unified two-stage SLU framework utilizing a Siamese network for metric-based similarity learning.
• Employed triplet loss optimization to enforce separation between dissimilar classes and group similar utterances.
• Integrated an attention-based joint decoder for simultaneous intent detection and slot filling within the framework.

Main Results:

• SBJLIS achieved 98.87% accuracy and 98.60% F1-score on the ATIS dataset.
• The model attained 99.61% accuracy and 98.68% F1-score on the SNIPS dataset.
• Outperformed all existing baseline models on both benchmark datasets, demonstrating superior performance.

Conclusions:

• Metric-based similarity learning provides a robust foundation for advanced conversational AI.
• The SBJLIS model demonstrates significant improvements in discrimination and generalization for SLU tasks.
• The proposed approach offers an interpretable and generalizable solution for real-world spoken language understanding applications.