Taxonomy Inference for Tabular Data
Contrastive representation learning and generative hierarchy construction.
This work studies how to discover entity types in collections of tables and organise them into useful hierarchies without relying on a pre-existing ontology.
EmTT fine-tunes encoder language models with contrastive learning before hierarchical clustering. GeTT uses decoder-only LLMs with iterative, top-down Chain-of-Layer prompting. Evaluation across three real-world datasets showed strong consistency with ground-truth taxonomies; contrastive fine-tuning improved Rand Index by 10% and Purity by 18%.
Stack: Python, PyTorch, Transformers, contrastive learning, hierarchical clustering
Code: PierreWoL/TwoMethods
Paper: ESWC 2025