Confidential — Stefan Michaelcheck Only

AutoPKG: An Automated Framework for Dynamic E-commerce Product-Attribute Knowledge Graph Construction

2026graph constructionnovelsystem

Pollawat Hongwimol, Cheng Lin Yu, Haoning Shang, Zhichao Wan, Chutong Wang, Lin Gui, Y. Li, Wenhao Sun, Yi Gao

arXiv (Cornell University)

https://doi.org/10.48550/arxiv.2604.16950OpenAlex: W7155029617arXiv: 2604.16950
1
URLs Found
0
Internal Citations
9
Authors
usable
Abstract Quality
GPT-5.5 Abstract Analysis

Problems Identified (4)

E-commerce Attribute Extraction Bottlenecks: Product attribute extraction in e-commerce is limited by inconsistent, incomplete, and costly-to-maintain ontologies.

Dynamic PKG Evaluation: Dynamic product-attribute knowledge graphs require evaluation of type/key validity, consolidation quality, and edge-level value assertion accuracy.

E-commerce Attribute Extraction Bottlenecks: Product attribute extraction in e-commerce is limited by inconsistent, incomplete, and costly-to-maintain ontologies.

Dynamic PKG Evaluation: Dynamic product-attribute knowledge graphs require evaluation of type/key validity, consolidation quality, and edge-level value assertion accuracy.

Proposed Solutions (5)

AutoPKG Multi-Agent LLM Framework: AutoPKG is a multi-agent LLM framework that automatically constructs product-attribute knowledge graphs from multimodal product content.

On-Demand Type and Attribute Induction: The framework induces product types and type-specific attribute keys on demand.

Multimodal Attribute Value Extraction: The framework extracts attribute values from both text and images.

Centralized Canonical Graph Consolidation: A centralized decision agent consolidates updates to maintain a globally consistent canonical graph.

Dynamic PKG Evaluation Protocol: The authors propose an evaluation protocol for dynamic product-attribute knowledge graphs.

Results (3)

High Product Type WKE:

Attribute Key WKE:

Multimodal Value Extraction F1:

Research Domain

E-commerce product-attribute knowledge graph construction

← Back to all papers