Taxonomy Inference for Tabular Data Using Large Language Models
Zhenyu Wu, Jiaoyan Chen, Norman W. Paton
TL;DR
This paper tackles taxonomy inference for heterogeneous tabular data by proposing two LLM-based approaches: EmTT, an embedding-based method that fine-tunes column encoders with SwAV and performs hierarchical clustering to build a DAG taxonomy $\mathcal{H}= (\mathcal{T},\mathcal{E})$, and GeTT, a generative method that generates per-table entity types and constructs hierarchies via Chain-of-Layer prompting. EmTT proceeds in three steps—Identify Top-level Types, Identify Attributes, and Infer Taxonomy—using subject-column embeddings and dendrogram-based pruning to derive a coherent taxonomy; GeTT uses a decoder-style LLM to generate types and then assembles a hierarchy with prompting and a ranking filter. The authors evaluate both methods on three real-world datasets with GT taxonomies against six metrics, including Rand Index, Purity, and a new Tree Consistency Score, demonstrating strong consistency for both approaches while highlighting trade-offs in stability, granularity, and computation. Overall, the work advances open-domain taxonomy construction for tabular data and offers practical pathways for improved data management, ontology learning, and downstream table-centric tasks, while identifying stability and granularity as key areas for future refinement.
Abstract
Taxonomy inference for tabular data is a critical task of schema inference, aiming at discovering entity types (i.e., concepts) of the tables and building their hierarchy. It can play an important role in data management, data exploration, ontology learning, and many data-centric applications. Existing schema inference systems focus more on XML, JSON or RDF data, and often rely on lexical formats and structures of the data for calculating similarities, with limited exploitation of the semantics of the text across a table. Motivated by recent works on taxonomy completion and construction using Large Language Models (LLMs), this paper presents two LLM-based methods for taxonomy inference for tables: (i) EmTT which embeds columns by fine-tuning with contrastive learning encoder-alone LLMs like BERT and utilises clustering for hierarchy construction, and (ii) GeTT which generates table entity types and their hierarchy by iterative prompting using a decoder-alone LLM like GPT-4. Extensive evaluation on three real-world datasets with six metrics covering different aspects of the output taxonomies has demonstrated that EmTT and GeTT can both produce taxonomies with strong consistency relative to the Ground Truth.