The research introduces a two-stage approach to Attribute Value Extraction (AVE) for e-commerce catalogs. The first stage utilizes a compact LLM, specifically Qwen3-4B, with Hyper-Parallel Decoding (HPD) to generate a ranked schema of purchase-discriminative attributes for each product category. The second stage then extracts the values for these attributes from the catalog text. The pipeline demonstrated 85% extraction accuracy, matching the performance of the foundational LLM from which it was distilled. Furthermore, inference costs were reduced by 92% compared to foundational LLMs. This reduction enables the use of the system at scale for product discovery and catalog enrichment. The resulting category-level structured representations provide consistent attributes across product categories, suitable for grounding downstream applications.
Source: https://arxiv.org/abs/2609.09716