English Español Français Deutsch
← Back to the list
Foundational figures Foundational figures Directly upstream of ai100's methodology

Juanzi Li

Combining structured knowledge from knowledge graphs with the statistical patterns language models learn from text — and what that combination buys you for evaluation.

Li's KEG Lab at Tsinghua has been the methodological home for one of the longest-running lines on knowledge-graph-augmented language models — OpenKE for KG embeddings, KEPLER for combining KG and language pretraining, and a research thread that led, via Zhipu, to the ChatGLM open LLM line. The throughline is a commitment to language models that have an external structured-knowledge anchor as well as their parametric memory, which makes the resulting systems easier to audit — you can ask the model and you can ask the KG separately, then compare. For ai100, which audits brand-mention behavior, the KG-augmented evaluation tradition is the closest precedent for separating "what the model says" from "what its sources support".

Worth following when
you want LLM evaluation methodology that takes external structured-knowledge anchors seriously — both in model design and in audit protocol.
Topics
knowledge-graph-augmented language models (KEPLER, OpenKE); the lineage from KG research to open Chinese LLMs (ChatGLM); auditability through external knowledge anchors.
Key works
OpenKE knowledge-graph embedding library (2018 onward, key contributor); KEPLER: A Unified Model for Knowledge Embedding and Pre-trained Language Representation (2021, co-author); ChatGLM open LLM line via Zhipu (2023 onward, co-founder).