Alexander, Natalie and Byamugisha, Joan and Makondo, Ndivhuwo and Buys, Jan (2027) Lexical Data Augmentation for Medical Text-to-SQL Question Answering, Proceedings of South African Institute for Computer Scientists and Information Technologists Conference (SAICSIT), 13-15 July 2026, Cape Town, South Africa, outh African Computer Science and Information Systems Research Trends, Springer.
Full text not available from this repository. (Use alternate locations listed below)Abstract
Existing text-to-SQL datasets do not reflect the rich clinical vocabulary used in real clinical queries. As a result, existing text-to-SQL models expect questions with medical concepts that are exact lexical matches to values in electronic health records (EHRs). Models often fail to translate questions into valid SQL queries when questions contain medical concepts synonymous with base medical concepts in the EHR. These question perturbations are referred to as value-synonym substitution, and these questions are classified as unanswerable. In this paper, we aim to improve the coverage of answerable questions (with answer abstention in place for edge cases), create a clinical value-synonym substitution dataset for question-to-SQL answering, and improve model robustness to value-synonym substitution using synonym knowledge. We use medical vocabularies in the Unified Medical Language System (UMLS) to augment input questions with additional synonym context, while also investigating inference-time augmentation. We investigate the viability of synonym knowledge from different UMLS medical vocabularies. Our results show that when augmenting these questions with synonym knowledge from SNOMED CT and RxNorm in the test set, it results in a greater than 9% improvement in exact matching and execution accuracies relative to our baseline. This work emphasizes the potential of language models to empower non-database experts in effectively querying EHRs, thereby supporting clinical decision-making.
| Item Type: | Conference paper |
|---|---|
| Subjects: | Computing methodologies > Artificial intelligence > Natural language processing |
| Date Deposited: | 09 Oct 2026 06:35 |
| Last Modified: | 09 Oct 2026 06:35 |
| URI: | https://pubs.cs.uct.ac.za/id/eprint/1811 |
Actions (login required)
 |
View Item |