Related Experiment Video
Updated: Jun 30, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
DataAtlas: automatic generation of data dictionaries using large language models
Raffaele Giancotti1,2, Rajna Fani1,3, Rafi Al Attrach1,3
1Laboratory for Computational Physiology, MIT Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA 02139, United States.
Objectives:
Reuse of shared datasets is often limited by incomplete or inconsistent documentation, hindering interpretation and secondary analysis. We developed DataAtlas, an open-source system for automated generation of data dictionaries from tabular datasets.
Materials And Methods:
DataAtlas combines deterministic structural profiling with large language model (LLM)-based semantic inference to generate human-readable descriptions of datasets, tables, and variables. The system integrates column metadata, statistical summaries, and representative sample values to produce concise, context-aware descriptions. We evaluated the approach on 3 heterogeneous clinical datasets using structural validation, blinded LLM-based comparison with official documentation, human expert review, and downstream task evaluation.
Results:
Generated descriptions were frequently preferred over official documentation, particularly when existing descriptions were incomplete or ambiguous. Human expert review showed strong agreement with high-performing models, with most descriptions rated as correct and low rates of hallucination. The system accurately reconstructed schema elements and flagged inconsistencies in existing documentation. In a text-to-SQL benchmark, augmenting database schemas with generated data dictionaries improved execution accuracy from 0.52 to 0.88.
Discussion:
Automated data dictionary generation improved both dataset interpretability and downstream analytical performances. Results highlight the importance of column-level metadata, particularly representative sample values, in grounding LLM-generated descriptions. Limitations include sensitivity to sample quality, potential domain-specific misinterpretation, and variability across model architectures, indicating that generated outputs should augment rather than replace expert curation.
Conclusion:
DataAtlas provides a practical approach for generating structured data dictionaries, enhancing the accessibility, reproducibility, and reuse of clinical data.
