OPPO Knowledge Graph: Algorithms, Applications, and Future Directions

OPPO's knowledge graph, built by the XiaoBu assistant team, is a self‑developed, large‑scale general graph containing hundreds of millions of entities and billions of triples, supporting millions of daily QA requests across devices such as phones, wearables, and IoT hardware.

The system architecture consists of three layers: a low‑level data processing platform with NebulaGraph as the graph database, a middle layer handling data acquisition, graph construction, and management, and a top layer offering applications like intelligent QA, search, content understanding, risk control, and health services.

Core algorithms include entity classification (rule‑based + pre‑trained language model pipeline), a two‑stage entity alignment approach (Dedupe grouping followed by a BERT‑based semantic matcher), and information extraction for both common attributes (using a CASREL pointer‑network model) and long‑tail attributes (using MRC models).

In the XiaoBu assistant, knowledge‑based QA is divided into structured KBQA (handling chain, multi‑variable, entity‑relation, and comparative queries) and unstructured QA (leveraging large‑scale web data and MRC models). Query processing involves domain identification, intent classification, entity recognition, template‑based parsing, and entity linking via a BiLSTM‑CRF recognizer, candidate retrieval, and a disambiguation model.

To address challenges such as voice‑input aliasing, recognition errors, and long‑tail queries, the team employs alias mapping, phonetic features, template pruning, and dual‑tower vector retrieval with BERT embeddings.

The summary concludes with future plans: integrating commonsense reasoning, multimodal graphs, personalized recommendation, large‑scale pre‑trained models, and low‑resource information extraction to further enhance the knowledge graph and its applications.