OneKE: Open-Source Bilingual Knowledge Extraction Framework for Large Language Models

Ant Group and Zhejiang University have announced the open‑source release of OneKE, a bilingual (Chinese‑English) knowledge extraction framework that aims to assist researchers and developers in information extraction, text data structuring, and knowledge graph construction.

Knowledge graphs are a key technology for making large models trustworthy and controllable; OneKE helps extract risk events, person and organization entities, and more, enabling clear presentation of event timelines, trends, and inter‑entity relationships. The framework supports OpenSPG and DeepKE and can be used out‑of‑the‑box.

Despite the advances of large language models, real‑world information remains fragmented and unstructured, leading to challenges such as ambiguity, polysemy, and metaphor in extraction tasks, which in turn cause hallucinations and unstable generation. A unified knowledge extraction framework can dramatically lower the cost of building domain‑specific knowledge graphs, providing explainable reasoning, enhancing model stability, and accelerating vertical applications.

Use cases include:

Medical domain: extracting clinical knowledge to manage doctors’ expertise, enabling controllable assistance and medical Q&A.

Financial domain: extracting financial indicators, risk events, causal logic, and industry chains to automate report generation, risk prediction, and chain analysis.

Governmental sector: structuring regulations and policies to improve service efficiency and decision accuracy.

To promote generative AI adoption, Ant Group and Zhejiang University have established a joint Knowledge Graph Laboratory, focusing on model‑enhanced graph construction, controllable generation, and domain world graphs.

OneKE, built on a fully fine‑tuned LLaMA2 model, demonstrates strong performance on supervised and zero‑shot entity, relation, and event extraction tasks, especially in bilingual settings.

The project’s official homepage and related OpenSPG repository are provided for further exploration.