Shopee's E‑commerce Knowledge Graph Construction and Integration with Large Models

Shopee shares its experience in constructing an e‑commerce knowledge graph and combining it with large AI models, outlining the motivations, technical challenges, and solutions.

The knowledge graph aims to bridge buyer intent and seller information, unify product data across markets, and improve user experience, while addressing issues such as heterogeneous data sources, inconsistent language, and massive scale.

Key challenges include multi‑source information diversity, varying data quality, language differences, and the sheer volume of billions of items across eight markets and six languages.

Compared with pure deep‑learning approaches, knowledge graphs offer stronger interpretability by structuring entities, relations, and attributes, though they require careful construction and domain expertise.

The construction pipeline consists of defining an ontology, extracting information from product pages, images, and reviews, assessing text and image quality, performing multimodal extraction, and handling entity disambiguation and attribute alignment.

Extraction methods combine rule‑based filters, multi‑task BERT models, NER, and prompt‑tuned T5, while multimodal alignment leverages models such as ALBEF, BLIP, Labse‑DinoV2‑ViT, and multilingual encoders; model compression techniques like MiniLM are also explored.

Knowledge fusion tackles entity linking, spelling correction, synonym handling, and unit conversion, using similarity measures, translation, and learned standardization models.

Knowledge processing includes reasoning and inconsistency detection through rule‑based association mining, graph embedding inference, and analogical reasoning to complete and validate the graph.

Practical applications span search (enhancing query understanding), recommendation (fine‑grained category recall), operations (seller data quality checks, product selection), AI assistants, and multimodal product generation (titles, descriptions, images, videos).

The integration with large models follows three patterns: feeding the graph into models, using models to enrich the graph, and co‑training; challenges include graph completion, hallucination mitigation, and real‑time updates via RAG and selective layer execution.

The article concludes with a Q&A session covering quality scoring, entity disambiguation, low‑resource training, tool recommendations, and the relationship between knowledge graphs and product catalogs.