Presentation Information
[5L3-OS-6b-06]An Evaluation-Based Approach to Automating Schema Design for Generative Knowledge Graph Construction
〇Younghun Lim1, Tatsunori Mori1 (1. Yokohama National University)
Keywords:
Generative Knowledge Graph Construction,Schema Optimization,Evaluation-based Search
Knowledge graphs have played an important role as a structured representation for effectively utilizing knowledge from large document collections. Recently, generative knowledge graph construction using large language models (LLMs) has attracted increasing attention. However, many existing studies assume a schema as a fixed input, and schema design still largely relies on manual effort.
This study reformulates schema design not as a fixed component but as an optimization target. In LLM-based generative environments, a schema functions not only as a semantic design artifact but also as a condition that guides the model’s output. As the quality of a schema cannot be determined without evaluating the generated results, schema design can be regarded as a black-box optimization problem that can only be explored through evaluation signals.
To address this issue, we propose an evaluation-based schema optimization method inspired by evolutionary algorithms. Experimental results show that naive schema design does not guarantee stable performance, while evaluation-driven exploration is effective for improving schema quality.
This study reformulates schema design not as a fixed component but as an optimization target. In LLM-based generative environments, a schema functions not only as a semantic design artifact but also as a condition that guides the model’s output. As the quality of a schema cannot be determined without evaluating the generated results, schema design can be regarded as a black-box optimization problem that can only be explored through evaluation signals.
To address this issue, we propose an evaluation-based schema optimization method inspired by evolutionary algorithms. Experimental results show that naive schema design does not guarantee stable performance, while evaluation-driven exploration is effective for improving schema quality.