Advances in Baidu Knowledge Graph: Technologies, Applications, and Future Directions

Knowledge graphs are the foundation that enables machines to understand the objective world like humans. The talk begins with a brief overview of Baidu's position and overall development of its knowledge graph, followed by an introduction to the two main branches: general and industry knowledge graphs, as well as two special graphs—event and video understanding.

The basic structure of a knowledge graph is illustrated: nodes represent entities or concepts, while edges capture relationships or attributes, turning rich real‑world knowledge into a machine‑processable form.

Baidu's knowledge graph development is divided into four stages: pre‑KG (before 2013), the formative period (2014‑2015) when the methodology and architecture were solidified, rapid expansion (2016‑2017) focusing on multi‑domain integration, and the recent years emphasizing heterogeneous multi‑graph integration, self‑learning, and multi‑media knowledge, reaching 5 billion entities, 5.5 trillion facts, and handling 400 billion daily requests.

Recent technical advances for the universal knowledge graph include:

Open knowledge mining using active learning and automatic template generation.

Structure self‑attention networks for rich entity‑level relation detection (AAAI 2021 paper).

Self‑learning that combines top‑down schema construction with bottom‑up automatic discovery, expanding schema scale 30× and doubling fact coverage.

Multi‑source data fusion via semantic space transformation for large‑scale entity disambiguation.

Knowledge‑graph‑based QA handling entity queries, web‑based answers, and reasoning‑heavy questions through KBQA, IR‑QA, and dynamic function computation.

Knowledge‑enhanced machine reading comprehension that merges graph knowledge with text representations, especially effective in knowledge‑dense domains such as medicine and law.

Event graphs model dynamic world changes. An event is defined by time, space, participants, and a theme. The architecture consists of an ontology layer, an event layer, and an argument layer that links to the entity graph. Extraction is performed via multi‑turn reading‑comprehension QA, enabling minute‑level hot‑event ingestion and supporting applications like hot‑event tracing, POI change detection, and intelligent writing.

Video understanding graphs extend knowledge‑graph techniques to multimodal video content. By parsing visual, audio, and textual streams and linking them to a video‑specific knowledge subgraph, Baidu achieves deep semantic understanding that powers search, information‑flow, and video‑platform products.

Industry knowledge graphs target professional domains such as medical, legal, financial, and risk control. The knowledge middle‑platform provides end‑to‑end pipelines for data ingestion, knowledge organization, and intelligent application. Notable results include over 300 k medical entities with 90% coverage, 90% interception of unreasonable medication, 95% accuracy in medical record quality control, and a 90% satisfaction rate in legal case recommendation.

Data and community openness are highlighted: Baidu releases datasets annually, hosts competitions (CCKS, CCKS‑Entity‑Linking), and launched the “Qianyan” project—an open‑source Chinese NLP data‑co‑creation initiative covering 7 tasks and 20+ datasets, with plans to expand to over 100 datasets.

Future outlook discusses trends and challenges in knowledge acquisition (dynamic, specialized, multi‑modal), representation (learning and reasoning for complex knowledge), and application (integration with deep learning, NLP, speech, vision, and explainability).

Finally, the speaker thanks the audience and encourages sharing, liking, and following the DataFunTalk community.