Reproduce “A survey of joint intent detection and slot-filling models in natural language understanding” on OpenResearch

About this paper

H. Weld, X. Huang, S. Long, J. Poon, S. C. Han

This survey meticulously reviews the evolution of joint intent detection and slot-filling models in Natural Language Understanding, categorizing approaches from classical methods to deep learning and Transformer-based architectures. It synthesizes findings on model architectures, feature utilization, and generalizability, while also noting the high performance achieved on standard benchmarks like ATIS and SNIPS.

Key insights

• Joint deep learning models consistently outperform separate or pipelined approaches for intent detection and slot filling, effectively leveraging the interdependency between the two tasks.
• While semantic embeddings (e.g., BERT) are the primary drivers of performance, syntactic features serve as useful, albeit supplementary, enhancements.
• Existing benchmark datasets like ATIS and SNIPS show performance saturation, suggesting they are largely "solved" in supervised settings and that new, more challenging benchmarks are needed.

Summary by alphaXiv.