Similarity-Based Intent Detection Using an Enhanced Siamese Network

NLDL 2025 Conference Submission12 Authors

Keywords: Intent detection, Siamese network, Dialogue system, Similarity metrics

TL;DR: This paper presents a similarity-based intent detection model using an enhanced Siamese network that combines multiple distance metrics to improve classification accuracy, particularly in low-resource and zero-shot learning scenarios.

Abstract: In Natural Language Understanding (NLU), intent detection is crucial for improving human-computer interaction. However, traditional supervised learning models rely heavily on large annotated datasets, limiting their effectiveness in low-resource scenarios with limited labeled data. Siamese networks, which are effective at learning similarity-based representations, provide a promising alternative by enabling few-shot learning. However, Siamese networks typically rely on contrastive loss or triplet loss, both of which introduce challenges. This study introduces a similarity-based intent detection model using an enhanced Siamese network to address these limitations. Our model employs Manhattan, Euclidean, and Cosine similarity metrics combined with a fusion layer to improve intent classification accuracy. We evaluated the model on the Airline Travel Information System (ATIS) and SNIPS datasets and demonstrated its superiority over state-of-the-art methods, particularly in low-resource and few-shot learning scenarios. The results highlight significant accuracy gains while maintaining computational efficiency, making it a robust solution for real-world dialog systems.

Submission Number: 12