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

Spoken language understanding (SLU) relies on intent detection and slot filling to interpret user utterances accurately. However, existing joint learning frameworks struggle to generalize across minority intent classes and paraphrase queries. They depend heavily on token-level embeddings and classification losses such as cross-entropy, which do not explicitly model semantic similarity. To address this limitation, this study proposes a Siamese-Based Metric Joint Learning model for Intent Detection and Slot Filling (SBJLIS). The model uses triplet loss optimization to enhance semantic distance learning between utterances. Unlike standard cross-entropy training, triplet loss enforces separation between dissimilar classes and brings semantically related sentences closer in the embedding space. This approach improves both discrimination and generalization. SBJLIS employs a unified two-stage SLU framework. The first stage uses a Siamese network for metric-based similarity learning. The second stage integrates an attention-based joint decoder for simultaneous intent detection and slot filling. By aligning embedding geometry with multi-task objectives, the model improves semantic discrimination and robustness to class imbalance and linguistic variation. Experimental results show that SBJLIS achieves 98.87% accuracy and 98.60% F1-score on the ATIS dataset, and 99.61% accuracy and 98.68% F1-score on SNIPS, outperforming all existing baselines. These findings confirm that metric-based similarity learning offers an interpretable and generalizable foundation for advanced conversational AI systems.