Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

StraTyper: Automated Semantic Type Discovery and Multi-Type Annotation for Dataset Collections

Christos Koutras, Juliana Freire

TL;DR

StraTyper tackles the challenge of open-world semantic type discovery for dataset collections by eliminating fixed taxonomies and enabling multi-type annotation. It integrates semantic clustering, length-stratified value sampling, dynamic type retrieval, and iterative cascading to generate collection-specific type vocabularies with substantially reduced reliance on proprietary LLMs. The approach achieves strong annotation quality, high multi-type recall, and clear cost benefits, while improving downstream tasks such as join discovery and schema matching. These results demonstrate practical impact for large, heterogeneous data repositories where domain-specific semantics and polysemy are prevalent.

Abstract

Understanding dataset semantics is crucial for effective search, discovery, and integration pipelines. To this end, column type annotation (CTA) methods associate columns of tabular datasets with semantic types that accurately describe their contents, using pre-trained deep learning models or Large Language Models (LLMs). However, existing approaches require users to specify a closed set of semantic types either at training or inference time, hindering their application to domain-specific datasets where pre-defined labels often lack adequate coverage and specificity. Furthermore, real-world datasets frequently contain columns with values belonging to multiple semantic types, violating the single-type assumption of existing CTA methods. While proprietary LLMs have shown effectiveness for CTA, they incur high monetary costs and produce inconsistent outputs for similar columns, leading to type redundancy that negatively affects downstream applications. To address these challenges, we introduce StraTyper, a cost-effective method for column type discovery (CTD) and multi-type annotation (CMTA) in dataset collections. StraTyper eliminates the need for pre-defined semantic labels by systematically employing LLMs to discovery types tailored to the dataset collection at hand. Through strategic column clustering, controlled type generation, and iterative cascading discovery, StraTyper balances type precision with annotation coverage while minimizing LLM costs. Our experimental evaluation-both manual and LLM-assisted-on real-world benchmarks demonstrates that StraTyper discovers accurate types for both numerical and non-numerical data, achieves substantial cost savings compared to commercial LLMs, and effectively handles multi-typed columns. We further show that StraTyper's annotations improve downstream tasks, including join discovery and schema matching, outperforming LLM-only baselines.

StraTyper: Automated Semantic Type Discovery and Multi-Type Annotation for Dataset Collections

TL;DR

StraTyper tackles the challenge of open-world semantic type discovery for dataset collections by eliminating fixed taxonomies and enabling multi-type annotation. It integrates semantic clustering, length-stratified value sampling, dynamic type retrieval, and iterative cascading to generate collection-specific type vocabularies with substantially reduced reliance on proprietary LLMs. The approach achieves strong annotation quality, high multi-type recall, and clear cost benefits, while improving downstream tasks such as join discovery and schema matching. These results demonstrate practical impact for large, heterogeneous data repositories where domain-specific semantics and polysemy are prevalent.

Abstract

Understanding dataset semantics is crucial for effective search, discovery, and integration pipelines. To this end, column type annotation (CTA) methods associate columns of tabular datasets with semantic types that accurately describe their contents, using pre-trained deep learning models or Large Language Models (LLMs). However, existing approaches require users to specify a closed set of semantic types either at training or inference time, hindering their application to domain-specific datasets where pre-defined labels often lack adequate coverage and specificity. Furthermore, real-world datasets frequently contain columns with values belonging to multiple semantic types, violating the single-type assumption of existing CTA methods. While proprietary LLMs have shown effectiveness for CTA, they incur high monetary costs and produce inconsistent outputs for similar columns, leading to type redundancy that negatively affects downstream applications. To address these challenges, we introduce StraTyper, a cost-effective method for column type discovery (CTD) and multi-type annotation (CMTA) in dataset collections. StraTyper eliminates the need for pre-defined semantic labels by systematically employing LLMs to discovery types tailored to the dataset collection at hand. Through strategic column clustering, controlled type generation, and iterative cascading discovery, StraTyper balances type precision with annotation coverage while minimizing LLM costs. Our experimental evaluation-both manual and LLM-assisted-on real-world benchmarks demonstrates that StraTyper discovers accurate types for both numerical and non-numerical data, achieves substantial cost savings compared to commercial LLMs, and effectively handles multi-typed columns. We further show that StraTyper's annotations improve downstream tasks, including join discovery and schema matching, outperforming LLM-only baselines.
Paper Structure (17 sections, 5 equations, 7 figures, 7 tables)

This paper contains 17 sections, 5 equations, 7 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Problem Setting & Our Approach
  3. Discovering Semantic Types
  4. Column Clustering
  5. LLM-based Type Discovery
  6. Type Propagation & Annotation
  7. Seed Cluster Type Propagation
  8. Targeted Closed-Set Annotation
  9. Iterative Cascading Discovery
  10. Experimental Evaluation
  11. Experiment Setup
  12. Evaluation Strategies and Metrics
  13. Discovery and Annotation Effectiveness
  14. Ablation Studies & Detailed Analysis
  15. Semantic Types for Column Join/Matches
  16. ...and 2 more sections

Figures (7)

  • Figure 1: StraTyper vs. CTA methods: CTA methods assign each column one type from a pre-defined list. For columns Location, LOC, and Name, they assign City, Abbrv, and ID—incorrect, generic, or ambiguous types that fail to capture multi-type values. In contrast, StraTyper discovers collection-specific types and uniformly annotates similar columns, correctly assigning borough and DBN to all three columns.
  • Figure 2: StraTyper receives a dataset collection in its input and proceeds to open-type discovery and multi-type annotation, either in a single-pass or an iterative cascading manner.
  • Figure 3: Type Discovery Prompt, integrating column names in seed cluster, length-stratified value samples, context based on co-occurrence profiles, dynamically retrieved possible semantic types and detailed instructions, as discussed in \ref{['sec:discovery']}.
  • Figure 4: Iterative Cascading discovery: i) StraTyper discovers semantic types from seed clusters at a high threshold ($\tau=0.99$), ii) propagates them to sibling columns in lower threshold clusters (LOC and Location Address for $\tau \in \{0.9, 0.75\}$) via closed-set annotation, iii) discovers types for residual unannotated columns (Loc for $\tau=0.75$).
  • Figure 5: Precision and Hit Rate results across all three dataset collections using automated evaluation with the LLM as a judge.
  • ...and 2 more figures

Theorems & Definitions (3)

  • definition 1: Column Type Annotation (CTA)
  • definition 2: Column Type Discovery (CDT)
  • definition 3: Column Multi-Type Annotation (CMTA)