Source-linked AI summary
Scaling E-Commerce Attribute Extraction with Parallel Decoding
Nikhita Vedula, Dushyanta Dhyani, Bryan Wang, Shervin Malmasi
TL;DR
E-commerce AVE must identify useful attributes despite messy catalogs and extract them efficiently at scale. The paper introduces a two-stage pipeline that discovers compact category schemas and extracts values with a fine-tuned compact model using HPD. It reports 85% extraction accuracy, matching the foundational teacher, with a 92% inference-cost reduction.
Problem
Manual schemas and large, inconsistent attribute sets make it difficult to identify purchase-discriminative attributes and extract values across messy e-commerce catalogs at scale.
Method
A two-stage LLM pipeline discovers category-level purchase-discriminative schemas with a foundational LLM, then extracts their values using fine-tuned Qwen3-4B with HPD.
Results
85% extraction accuracy matches the foundational teacher LLM while inference costs decrease by 92%.
Takeaways & Limitations
The pipeline automatically constructs comparable category-level product representations and makes repeated catalog-scale processing practical without manual schema engineering.
Takeaways & Limitations
The system uses textual catalog content and has been evaluated only on English-language catalogs; image-only attributes and non-English marketplaces remain unexplored.
Abstract
from arXiv · showhide
Customers rely on specific product attributes to compare products and make purchasing decisions, but e-commerce catalogs are messy and unstructured, making it difficult to identify which attributes matter most and extract them at scale. Standard Attribute Value Extraction (AVE) systems treat all attributes equally, producing large, inconsistent attribute sets that do not reflect the factors consumers use to differentiate products. We introduce a two-stage LLM pipeline that first discovers a compact, ranked schema of purchase-discriminative attributes for each product category, then extracts their values from catalog text using a fine-tuned compact LLM (Qwen3-4B) with Hyper-Parallel Decoding (HPD). This pipeline achieves 85% extraction accuracy, on par with the foundational LLM it was distilled from, while reducing inference costs by 92% over foundational LLMs, enabling production-scale use for product discovery and catalog enrichment. The resulting category-level structured representations effectively constitute automatically constructed product knowledge bases, providing consistent, comparable attributes across varied product categories that can ground downstream knowledge-intensive applications.
1 Introduction
The paper addresses the difficulty of defining useful schemas and extracting attributes efficiently from messy, large-scale e-commerce catalogs. It proposes a two-stage LLM pipeline that discovers purchase-discriminative schemas and performs compact-model extraction with parallel decoding.
- Motivation: Attribute Value Extraction supports search, filtering, comparison, recommendation, and product quality workflows, but catalog information is noisy and inconsistently expressed.Product information is distributed across multiple fields, with missing, implicit, duplicated, and non-standard values.
- Challenges: Manual schema definition across thousands of categories is expensive, while catalog-scale extraction is computationally costly as products and schemas evolve.These are the two deployment challenges motivating the proposed pipeline.
- Contribution: The pipeline targets compact, purchase-discriminative category schemas so products share comparable representations without manual schema engineering.It automatically induces category-level ontologies and populates them at scale.
- Method: The two stages use a foundational LLM for schema discovery and a fine-tuned compact model with Hyper-Parallel Decoding for simultaneous value extraction.HPD exploits conditional independence among attribute values to improve throughput.
- Results: 85% extraction accuracy matches the foundational LLM teacher, while inference cost falls by 92% and the system processes hundreds of millions of products across marketplaces.Schema discovery attributes were judged correct or relevant in 89.6% of cases, and the compact model uses more than 100× fewer parameters.
- Evaluation: The evaluation covers schema quality, extraction accuracy, quantity understanding, determinism, and efficiency, including cost and throughput analysis.These dimensions assess both output quality and production scalability.
2 Related Work
Prior AVE methods support structured extraction but commonly rely on predefined schemas or costly repeated processing. Related work addresses schema discovery, knowledge-base construction, and efficient decoding, while this paper combines automatic schema discovery with low-cost generative extraction.
- Product attribute-value extraction: Sequence-tagging and question-answering AVE methods improve extraction for predefined attributes but generally assume available schemas and can be costly across categories.These approaches recover structured attributes from noisy catalog content for downstream e-commerce applications.
- Generative AVE: Generative AVE models handle heterogeneous inputs and flexible output formats, but production deployment still requires scalable schema construction and efficient extraction.The paper targets both requirements.
- Schema discovery and knowledge bases: Schema discovery and product schema modeling reduce reliance on human supervision, connecting e-commerce attribute modeling with automated knowledge-base construction.LLMs are described as reshaping ontology engineering, knowledge extraction, and knowledge fusion.
- Efficient and scalable extraction: Efficient extraction research reduces repeated context processing through prompt-level parallel decoding, caching, lightweight interactions, and adaptation to evolving catalogs.This paper complements those approaches by specializing a small model and decoding conditionally independent attribute values in parallel with HPD.
3 Scalable Schema-Guided Attribute Value Generation
The proposed generation framework first builds shared, purchase-discriminative schemas for each category and then extracts standardized attribute values from product text. Hyper-Parallel Decoding addresses the cost of generating many attribute values at catalog scale by decoding them simultaneously.
- Overview: The pipeline converts products with unique attribute sets into compact, category-level representations using shared purchase-discriminative schemas.Category granularity matters: broad categories produce generic attributes, while overly granular categories create massive attribute sets.
- Problem Statement: For each product, the target representation pairs every category-schema attribute with its value extracted from the product context when available.The schema contains N attributes defined at the category level.
- Problem Statement: Schema discovery identifies attributes most informative for purchase decisions, and shared schemas make representations comparable among products in the same category.The two subproblems are attribute schema discovery and attribute value generation.
- Stage 1: Schema Discovery: Representative products guide category-specific attribute discovery, while global attributes add pack size, weight, and volume fields across categories.The discovery prompts also specify descriptions, data types, and standardized categorical values.
- Stage 1: Schema Discovery: LLM-based clustering standardizes semantically redundant attribute names produced across categories.This reduces variation such as multiple names for water resistance attributes.
- Stage 2: Motivation: At an optimistic throughput of 100K products per hour, foundational-LLM processing of a complete catalog would take multiple weeks and does not scale for real-world service.The cost challenge motivates a specialized extraction stage.
- Stage 2: Hyper-Parallel Decoding: HPD reduces decoding steps from P × Σ_i K_i to approximately max_i K_i by generating conditionally independent attribute values in parallel.It uses a JSON skeleton, BOV markers, position-ID gaps, and causal attention masks to maintain parallel generation.
- Stage 2: Standardization: Structured prompts and constrained HPD enforce standardized units, categorical options, numerical formats, and valid data types for extracted values.Token-level masks constrain categorical outputs, while regular expressions constrain numerical number-unit formats.
4 Experiments and Results
The pipeline is evaluated on a large catalog for schema quality, extraction quality, reliability, and efficiency. Results indicate that compact HPD models preserve extraction quality while substantially reducing processing cost.
- Experimental setup: The evaluation uses a proprietary catalog with more than 30M products spanning a few thousand categories, alongside automated and human assessments of both pipeline stages.Stage 2 is evaluated on 20K products, while human evaluation covers 500 products.
- Schema discovery quality: 89.6% of discovered key attribute names are judged correct or relevant across a few thousand categories, with 4.0% irrelevant and 6.4% too vague or too specific.The assessment covers 20 key attributes per category.
- Attribute value generation quality: Approximately 85% extraction accuracy is achieved by both fine-tuned Qwen3 models, matching or slightly exceeding the foundational teacher despite a >100× reduction in parameters.The evaluation reports no quality degradation from HPD while providing speed gains.
- Determinism and coverage: Approximately 93% of outputs are identical or semantically equivalent across repeated runs, while overall pipeline coverage reaches 99.997%.Variations primarily concern subjective attributes, and the 0.003% uncovered products lack catalog information or have incorrect category assignments.
- Efficiency and throughput: 92% cost reduction is achieved for processing 30M products with HPD compared with the foundational LLM baseline.HPD reaches approximately 159K products per hour per instance, versus approximately 18K for standard autoregressive decoding on Qwen3-4B; the foundational baseline costs approximately $63,000 for 30M products.
- Effect of attribute schema size on parallelism: Attribute-values extracted per hour increase approximately linearly with schema size while product throughput remains approximately constant.HPD decodes all attribute values simultaneously, with fixed decoding steps across the tested schema sizes.
5 Practical Use Case Evaluation
The practical use case applies the pipeline to large-scale catalog processing and publishes shared category-level attribute representations. A COFFEE_MAKER example shows how these schemas support direct product comparison and downstream catalog applications.
- Large-scale deployment: The full pipeline processes hundreds of millions of products across two major marketplaces using distributed A100 GPU instances.Products are processed independently across 4–20 instances to achieve linear throughput scaling.
- Refresh strategy: New products are processed daily or weekly, while full catalogs are reprocessed monthly and category schemas are refreshed less frequently.Schemas are versioned per refresh cycle and validated against the previous version before deployment.
- Downstream applications: Extracted attribute-value pairs are published to an internal knowledge base for product comparison, catalog enrichment, quality improvement, and personalization use cases.The shared category-level representations provide structured data for downstream querying.
- Practical validation: Seller interviews across two marketplaces report that the extracted attributes reflect factors used to position products and inform customer purchase decisions.Sellers evaluated the attributes for quality and usefulness in differentiating products.
- Sample system output: Shared category schemas make representations directly comparable across products, including budget pod machines with different brew, heating, grinder, and build-material values.The example contrasts a semi-automatic machine with a pod machine while retaining common attribute slots.
6 Conclusion
The paper presents a scalable two-stage pipeline for extracting purchase-discriminative attributes from large e-commerce catalogs. It combines automatic schema discovery with parallel value extraction, achieving 85% accuracy while reducing inference costs by 92%.
- The pipeline automatically discovers compact, category-level schemas using a foundational LLM, eliminating manual schema engineering across thousands of categories.
- The fine-tuned Qwen3-4B model uses Hyper-Parallel Decoding to extract multiple attribute values simultaneously.HPD exploits the conditional independence of attribute values.
- 85% extraction accuracy matches the foundational teacher LLM while reducing inference costs by 92%.
Limitations
The current system extracts attributes only from English textual catalog content and does not capture information conveyed exclusively through images. Its parallel decoding assumption and model-family evaluation also constrain the scope of the reported results.
- The system cannot capture product information conveyed only through images, including visual design details, color variations, and packaging quantities.Multimodal inputs are left for future work.
- The pipeline has been evaluated only on English-language catalogs, leaving non-English marketplaces unexplored.Multilingual deployment may require multilingual fine-tuning and different schema discovery prompts.
- HPD assumes that extracting one attribute value does not depend on previously extracted values, although some attributes may be correlated.Examples include product weight and volume, or material and durability.
- Evaluation used only Qwen3 models in 4B and 8B parameter sizes, so other architectures or model scales may yield different quality-efficiency tradeoffs.