Table-Guided Hyperspherical Diffusion for Preserving Semantic Dependencies in Column Type Annotation

This repository contains the source code for the CIKM 2026 paper “Table-Guided Hyperspherical Diffusion for Preserving Semantic Dependencies in Column Type Annotation”.

Overview

We propose a table-guided hyperspherical diffusion framework for Column Type Annotation (CTA) within Semantic Table Interpretation (STI). Unlike conventional discriminative approaches, our framework generates column type representations through iterative denoising in hyperspherical space, guided by table-level context and column representations.

✨ Key Components

Component	Description
Table Context Encoder	Learns contextualized column representations for both categorical and numerical columns, while extracting table-level contextual signals
Hyperspherical Diffusion Denoiser	Progressively refines noisy semantic representations through iterative denoising conditioned on table-guided signals
Semantic Label Retrieval	Retrieves final semantic types via cosine similarity over the hyperspherical label space

🚀 Installation

Requirements

Python  >= 3.10
PyTorch >= 2.0

📂 Datasets

We evaluate on four public CTA benchmark datasets:

Dataset	# Tables	# Types	Download
GitTables-DBpedia	3,737	101	SemTab 2022
GitTables-Schema	2,853	53	SemTab 2022
SOTAB-CTA	24,275	91	WDC SOTAB
WikiTables-CTA	406,705	150	TabEL

Place the downloaded datasets under the following structure:

data/
├── GitTables-DB/
├── GitTables-Schema/
├── SOTAB-CTA/
└── WikiTables-CTA/

🏋️ Training

python train.py --data GitTables-DB

📊 Test

python test.py \
  --data GitTables-DB \
  --checkpoint checkpoints/model.pt

📈 Results

Our framework consistently outperforms discriminative and LLM-based baselines across all benchmarks.

Dataset	Micro-F1	Macro-F1
GitTables-DBpedia	56.18	31.91
GitTables-Schema	66.42	40.23
SOTAB-CTA	87.50	86.46
WikiTables-CTA	93.37	73.19

References

If you use our dataset or useful in your research, please kindly cite our papers.