Hybrid Generative-Tagging Decoding for Efficient Product Attribute-Value Extraction

Hybrid generative-tagging decoding optimizes product attribute extraction by combining span-based tagging for frequent, copyable values with auto-regressive generation for rare attributes. This architecture bypasses the speed bottlenecks of pure generation while overcoming the fixed label constraints of traditional tagging models.