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

The paper presents a hybrid generative-tagging decoding approach for extracting product attribute-value pairs in e-commerce. It combines a tagging head that directly copies frequent attribute values from the source text with a generative head that handles rare attributes and non-copyable values. Experiments integrated with vLLM inference show a 10% increase in attribute coverage, a 63% rise in correct extractions, and a 44% reduction in latency without sacrificing precision.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Computation and Language.