TabuLLM: feature extraction and interpretation of text columns in tabular data using large language models
File(s)
Published version
Author(s)
Taghavi Azar Sharabiani, Mansour
Bottle, Alex
Mahani, Alireza S
Type
Journal Article
Abstract
TabuLLM is a Python package for extracting and interpreting LLM-based text features from datasets that combine free-text and structured columns. Motivated by a clinical study predicting postoperative acute kidney injury in paediatric cardiopulmonary bypass patients from electronic health records, the package integrates large language model (LLM) embeddings into scikit-learn-compatible pipelines for tabular datasets containing text columns alongside structured features. The package provides three core components: 1) TextColumnTransformer wraps the LangChain embedding ecosystem in a scikit-learn transformer interface with native multi-column handling; 2) GMMFeatureExtractor extends scikit-learn’s Gaussian mixture model with a transform() method that returns per-cluster log-joint features for dimensionality reduction; and 3) ClusterExplainer generates natural language cluster descriptions using LLMs with recursive summarization, optional outcome-based statistical testing, and narrative synthesis. On a fraud detection benchmark (4.8% fraud rate), incorporating text features via TabuLLM improves PR-AUC from 0.64 (structured features only) to 0.82 (text concatenation), with per-column stacking ensembles reaching 0.91. TabuLLM thus lowers the barrier to text–tabular integration: state-of-the-art embeddings become a drop-in addition to existing scikit-learn pipelines, with interpretability tools to understand what the text features capture. The package is available via pip install tabullm.
Date Issued
2026-09-01
Date Acceptance
2026-06-10
Citation
SoftwareX, 2026, 35
ISSN
2352-7110
Publisher
Elsevier
Journal / Book Title
SoftwareX
Volume
35
Copyright Statement
© 2026 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
License URL
Publication Status
Published
Article Number
102810
Date Publish Online
2026-06-17