Automated Robust Dynamic Generation of KGQA Benchmark Datasets

Knowledge Graph Question Answering (KGQA) systems transform natural-language questions into structured SPARQL queries to retrieve information from Knowledge Graphs (KGs). However, existing KGQA benchmarks are static, prone to obsolescence as KGs evolve, and increasingly unreliable for evaluating systems based on Large Language Models (LLMs) due to memorization effects (i.e., a model has reproduced the correct output from existing training data). This paper presents DynBench, a fully automated and robust framework for the dynamic generation of KGQA benchmark datasets. Unlike prior approaches, DynBench is enabled to automatically generate an arbitrary number of new benchmark datasets while preserving the natural-language surface forms and structural complexity of the original SPARQL queries in comparison to the given original KGQA dataset. Using entity and property substitutions within the same KG, DynBench produces semantically consistent questionquery pairs and automatically validates them without human intervention. We introduce an automatic validation mechanism that ensures semantic and syntactic integrity through backtransformation and metric-based validation. Two dynamic benchmarks were generated from the well-known datasets QALD-9-plus as well as LC-QuAD 2.0 and evaluated using both human experts and automated metrics to show DynBench’s datasetagnostic capabilities. Given our results, we can infer that the Levenshtein distance serves as the most reliable automatic validation measure, achieving a precision of up to 0.96 and demonstrating a strong correlation with human assessments. DynBench thus enables scalable, repeatable, and memorization-resistant dataset generation—providing a foundation for sustainable and fair evaluation of LLM-based KGQA systems.