How Large-Scale Knowledge Graphs Are Shaping AI and Natural Language Understanding

On December 20, Alibaba and the Chinese Society of Computational Linguistics' Knowledge Graph Committee held a Knowledge Graph symposium in Hangzhou, chaired by Alibaba Group Vice President Qiang Hui, with participation from many of China’s top experts.

Sun Le

Researcher, doctoral supervisor, and head of the Chinese Information Processing Lab at the Institute of Software, Chinese Academy of Sciences; Vice‑President and Secretary‑General of the Chinese Society of Computational Linguistics.

Large‑scale knowledge graphs are crucial for natural language understanding.

In the era of big data, language comprehension requires not only data models and computation but also insights from cognitive neuroscience and the use of massive knowledge graphs. Sun’s research focuses on two aspects: extracting knowledge from text to build large‑scale Chinese knowledge bases, and leveraging existing knowledge to understand text, such as entity linking and semantic analysis.

Li Juanzi

Professor, doctoral supervisor, and head of the Knowledge Engineering Lab at Tsinghua University; Director of the Language and Knowledge Computing Committee of the Chinese Society of Computational Linguistics. She leads the development of the large‑scale cross‑language knowledge graph XLORE.

Knowledge engineering in the big‑data era accelerates machine intelligence.

Transforming massive data into knowledge adds semantic information, enabling insight extraction and intelligent services. Knowledge graphs express Internet information in a form closer to human cognition, bridging the gap between low‑value features of machine learning and human understanding, thus acting as an accelerator for AI.

Zhao Jun

Researcher and doctoral supervisor at the Institute of Automation, Chinese Academy of Sciences, with extensive work on information extraction and QA systems.

Question‑answering and dialogue systems need knowledge graphs as foundational support.

Knowledge graphs provide essential infrastructure for QA and dialogue, enabling systems to retrieve factual information and learn interaction patterns from large‑scale conversational data. Combining KG infrastructure with deep learning offers promising avenues for natural, user‑friendly interactions.

Chen Huajun

Professor and doctoral supervisor at Zhejiang University, initiator of OpenKG, and Vice‑Director of the Zhejiang Provincial Big‑Data Intelligent Computing Key Laboratory.

Alibaba’s KG faces challenges that demand integrated knowledge representation, reasoning, NLP, and deep learning.

Knowledge graphs enable AI to acquire a “knowledgeable” mind, supporting search, QA, intelligent recommendation, platform governance, and consumption‑trend insights. The most difficult technical hurdle is deeply embedding reasoning capabilities.

Qi Guilin

Professor and doctoral supervisor at Southeast University, co‑founder of the OpenKG alliance, and editorial board member of several semantic web journals.

The connotation of knowledge must be combined with reasoning to be fully expressed.

Knowledge representation can take many forms—graphs, logical formulas, vectors, tensors—but these are merely manifestations. Logical reasoning adds the necessary depth, enabling inconsistency detection during KG construction and query rewriting during application, thereby delivering more precise answers.

Chen Wenliang

Professor and doctoral supervisor at Suzhou University, Deputy Director of the Human Language Technology Institute.

Generating training data directly from knowledge graphs achieves “more, faster, better, cheaper”.

By combining a small number of expert annotations with large‑scale crowd contributions, or by using remote supervision to generate noisy training data from KGs, high‑quality datasets for tasks such as segmentation, POS tagging, and NER can be produced efficiently, though further work is needed to close the performance gap.

Liu Zhiyuan

Assistant Professor and doctoral supervisor at Tsinghua University, with over 20 papers in top AI conferences and more than 1,900 citations.

Knowledge‑representation learning such as TransE will play a major role in knowledge acquisition and application.

Embedding KG triples into low‑dimensional vectors overcomes sparsity, enables multi‑source fusion and knowledge transfer, and has been widely applied to KG completion, relation extraction, KG fusion, and entity classification.

Feng Yansong

Associate Professor at Peking University, leader of the PKUBase large‑scale structured Chinese encyclopedia KG.

Leveraging prior knowledge and human expertise is essential for building robust knowledge graphs.

KG construction can follow three main routes: expert manual creation, automatic extraction from existing structured resources, and crowd‑sourced annotation. Combining expert‑crafted rules or ontologies with modern deep‑learning methods, and using statistical inference to generate and refine rules, is a promising direction for future KG research.

The symposium highlighted that large‑scale knowledge graphs are indispensable for advancing AI, bridging the gap between raw data and human‑like understanding, and enabling a new generation of intelligent services.