Xiaomi Knowledge Graph: Architecture, Key Technologies, and Business Applications

Since its inception in 2017, Xiaomi's knowledge graph has grown to support billions of daily accesses across products such as XiaoAi, Xiaomi Youpin, intelligent Q&A, user profiling, and virtual assistants, significantly improving content understanding, user comprehension, and entity recommendation.

The platform is positioned within Xiaomi's AI department, forming a middle‑layer infrastructure that supplies foundational capabilities (vision, NLP, knowledge graph, speech, deep learning) to downstream services like smart Q&A, intelligent customer service, XiaoAi, and e‑commerce.

Key application scenarios include:

Smart Q&A: serving phones, speakers, wearables, cars, TVs, and children devices with billions of daily queries, supporting both general and rule‑based reasoning, multi‑condition and multi‑hop inference, and list‑style disambiguation.

Intelligent Customer Service: deployed in Xiaomi.com and Xiaomi Finance, using NL2SQL and AI engines for data standardization, entity extraction, knowledge graph inference, and data services.

E‑commerce & Game Center: product and game graphs enhance search discovery, recommendation, sentiment extraction, and have increased conversion rates by over 30%.

AI Virtual Assistant: multimodal graph integration enables features like plant identification.

Core technical pillars are:

Entity Linking : combines Chinese word segmentation, NER (BERT+CRF), candidate selection, and disambiguation using popularity, semantic similarity (DeepType, DeepMatch) and other contextual features, achieving F1 = 0.8954 in the 2020 CCKS competition.

Knowledge Fusion : merges equivalent entities from heterogeneous sources using attribute importance, popularity, authority, richness, and co‑occurrence, followed by attribute conflict resolution.

Concept Graph : builds taxonomy via ontology, autophrase mining, and sequence labeling (BERT+BiLSTM+CRF), extracting high‑quality concepts for downstream tasks.

Automated Construction : provides end‑to‑end pipelines for entity and relation generation, supporting custom builds.

Industry‑specific explorations include:

Product Graph: constructs classification, main‑term extraction, synonym mining, hierarchical relations, and scenario concepts, improving search and recommendation in Xiaomi Mall and Youpin.

Hypernym Mining: uses BERT classification, hierarchical multi‑classifiers (HMC), and text‑pair classification to associate products with higher‑level concepts.

Synonym Extraction: treats synonym discovery as a sequence labeling problem, filtered by hierarchical conflicts, word‑vector similarity, and BERT similarity, reaching 94% precision.

Financial Graph: supports identity verification and related‑party detection in loan and insurance services, with a dedicated architecture diagram.

In summary, Xiaomi's knowledge graph now contains over a hundred billion facts, powers more than ten business scenarios, offers twenty‑plus technical capabilities, and provides a mature automated construction workflow, with ongoing expansion into additional industry domains.