NEUROSYMTAX: NEURO-SYMBOLIC AUTOMATED TAXONOMY DEVELOPMENT

Taxonomies organize complex domains, but their development is labor-intensive, difficult to reproduce, and dependent on designer judgment. This study investigates how neuro-symbolic artificial intelligence can automate the extended taxonomy design process while preserving methodological control and evidence traceability. The proposed approach, NeuroSymTax, assigns interpretation and proposal tasks to a large language model and process control to deterministic code. The approach was evaluated through exploratory comparisons with a reference benchmark taxonomy in the dynamic domain of decentralized finance, using quantitative metrics and a qualitative analysis. Three NeuroSymTax executions generated taxonomies in under 40 minutes at language-model API costs of US$0.31-0.36 per run. These taxonomies partially matched the benchmark and introduced additional distinctions relevant to the stated purposes. Compared to baseline and ablation configurations, NeuroSymTax full runs provided stronger methodological assurance through evidence traceability, symbolic verification, and evaluation on new objects, and their taxonomies achieved greater dimensional similarity with the benchmark taxonomy. This study advances the automation of taxonomy development through an efficient and methodologically controlled approach and illustrates its potential to organize knowledge in rapidly evolving domains.