AliCoCo: Alibaba’s E‑commerce Cognitive Concept Net – Architecture, Construction, and Applications

AliCoCo (Alibaba E‑commerce Cognitive Concept Net) is a knowledge graph designed to represent user shopping demands as explicit concepts, enabling more intelligent e‑commerce experiences.

The system consists of four layers: E‑commerce Concepts (short user‑need phrases), Primitive Concepts (fine‑grained word tokens), a Taxonomy (hierarchical classification of primitive concepts), and the Item layer (billions of products linked to concepts).

To build the taxonomy, experts defined about 20 top‑level classes (e.g., Category, Function, Material) and refined them into millions of leaf nodes, providing a rich ontology comparable to Freebase or ConceptNet but with extensive concept instances.

Primitive concepts are extracted using a BiLSTM‑CRF model trained on large e‑commerce corpora, followed by crowdsourced verification. Hypernym (up‑down) relations are discovered via pattern‑based unsupervised methods and projection‑learning supervised models, with active learning to reduce labeling cost.

E‑commerce concepts are generated in two stages: massive candidate generation from text mining (AutoPhrase) and combinatorial generation from primitive concepts, then filtered by a knowledge‑enhanced Wide&Deep model that incorporates BiLSTM features, POS/NER tags, Wikipedia gloss embeddings, and statistical BERT perplexity scores.

Linking concepts to items is treated as a semantic matching problem; the model enriches item representations with associated primitive concepts and external Wikipedia knowledge to improve relevance, especially for short concept queries.

AliCoCo has been deployed in Alibaba’s core e‑commerce services. Over 98% of Taobao/Tmall items are covered, each linked to an average of 14 primitive concepts and 135 e‑commerce concepts, raising query coverage from 35% to 75% and powering knowledge cards in search as well as theme‑based recommendations.

Future work includes expanding commonsense relations, probabilistic modeling of concept‑item links, and multilingual extensions to support Alibaba’s global strategy.