Hybrid method speeds and improves product attribute extraction
Hybrid Generative-Tagging Decoding for Efficient Product Attribute-Value Extraction
Computation and Language
Summary
Extracting product details like color or size is important for online stores but can be tricky. The authors developed a new computer method that mixes two ways of finding these details: one that quickly picks out common facts from text and another that writes out rarer or more complex ones. This mix helps the system find many more correct details faster without losing accuracy. Their approach balances speed and coverage better than previous methods.
What this means in practice
- •For e-commerce platform engineers: Improve automated product catalog building by extracting more product details faster and with no loss in precision.
- •For retail data integration teams: Enhance integration of diverse product data sources by reliably capturing uncommon or normalized product attributes.
Authors
Keshav Kolluru, Manas Gupta, Maroof Ahmad, Suchitra Sathyanarayana
Abstract
Attribute-value extraction is important for building high-quality product catalogs in e-commerce. Existing systems either rely on generative or tagging strategies for extracting the attribute names and their corresponding values. Tagging models directly label spans in the source, making them fast and grounded, but they cannot produce values that require reformatting or generate rare attribute names outside their label space. Generative models avoid these drawbacks by generating both attribute names and values token by token, but this auto-regressive nature makes them slow and limits attribute coverage. We introduce hybrid generative-tagging decoding, a novel model architecture that combines tagging source spans and open-ended generation. For frequent attributes whose values are copyable, a tagging head extracts values directly from the source. This lets the generative head focus on the remainder - rare attributes outside the tagger's label space and non-copyable values, including implicit or normalized values. Integrated into vLLM inference, the hybrid decoder raises attribute coverage by 10% and extracts 63% more correct attribute-value pairs at no precision cost, while reducing latency by 44%.