AutoPKG Framework Automates E-commerce Product Knowledge Graph Construction Using Multi-Agent LLMs
AutoPKG has launched a framework utilizing a multi-agent Large Language Model to streamline the creation of Product-attribute Knowledge Graphs from diverse e-commerce content, tackling issues related to inconsistent and incomplete ontologies. This innovative system identifies product types and specific attribute keys, extracts relevant information from both text and images, and integrates updates through a centralized decision-making agent to uphold a globally consistent canonical graph. A dynamic PKG evaluation protocol assesses the validity of types and keys, the quality of consolidation, and edge-level precision for value assertions post-canonicalization. In testing with a substantial dataset from Lazada (Alibaba), AutoPKG recorded a Weighted Knowledge Efficiency of 0.953 for product types, 0.724 for attribute keys, and 0.531 edge-level F1 for multimodal extraction. The framework, outlined in arXiv preprint 2604.16950v1, aims to enhance data consistency and lower maintenance expenses in e-commerce environments.
Key facts
- AutoPKG is a multi-agent LLM framework for automated Product-attribute Knowledge Graph construction
- It induces product types and type-specific attribute keys dynamically from multimodal content
- Attribute values are extracted from both text and images
- A centralized decision agent consolidates updates to maintain a globally consistent canonical graph
- An evaluation protocol measures type and key validity, consolidation quality, and edge-level accuracy
- On a Lazada (Alibaba) dataset, it achieved up to 0.953 WKE for product types
- It achieved 0.724 WKE for attribute keys and 0.531 edge-level F1 for multimodal extraction
- The framework addresses bottlenecks from inconsistent, incomplete, and costly ontologies in e-commerce
Entities
Institutions
- Lazada
- Alibaba