Reasoning Techniques in Knowledge Graphs and Their Application to a High‑School Exam Robot

In this presentation, Professor Qi Guilin from Southeast University outlines the evolution of knowledge graphs from early semantic nets to modern applications such as Google’s Knowledge Graph, emphasizing their role as structured repositories of common‑sense and domain knowledge for artificial intelligence.

What is a Knowledge Graph? A knowledge graph is a graph‑based data model where nodes represent entities or concepts and edges encode various relationships, providing a flexible schema for AI reasoning.

Logical Reasoning in Knowledge Graphs The speaker describes ontology‑based reasoning, hierarchical (is‑a, part‑of) relations, disjointness constraints, and rule‑based inference such as production‑rule systems, illustrating how formalized concepts enable machines to understand and infer new facts.

Statistical and Representation‑Learning Reasoning By integrating statistical inference, contradictions in encyclopedic data can be detected (e.g., an organism classified both as animal and plant). Representation learning maps entities and relations into low‑dimensional vectors, allowing similarity‑based reasoning, tensor representations, and path‑ranking algorithms to predict missing links.

Application to a High‑School Exam Robot Using spatial reasoning as an example, the talk shows that existing single‑model approaches (RCC, CSD, ICD) are insufficient for complex exam images; a hybrid model combining topological, directional, and distance representations is needed. The robot would also leverage geological knowledge (e.g., sedimentary rock layering) to answer domain‑specific questions.

In summary, ontologies, rule‑based and statistical reasoning are essential components of knowledge‑graph‑driven AI systems, and their integration can empower intelligent tutoring robots for high‑school examinations.