From Examples to Patterns: LLM-Generated Reg-ular Expressions for Entity Extraction in Czech Clinical Texts
| Authors | |
|---|---|
| Year of publication | 2024 |
| Type | Article in Proceedings |
| Conference | Proceedings of the Eighteenth Workshop on Recent Advances in Slavonic Natural Languages Processing, RASLAN 2024 |
| MU Faculty or unit | |
| Citation | |
| web | https://nlp.fi.muni.cz/raslan/2024/paper6.pdf |
| Keywords | NLP; LLM; regex; text mining; clinical notes |
| Description | Entity extraction in clinical texts is essential for converting unstructured data in clinical notes into structured formats, facilitating large-scale analysis and clinical decision support. Traditional methods often rely on handcrafted regular expressions (regexes), which, while effective, demand significant time and specialized knowledge to create -- resources that healthcare professionals may lack. We introduce a novel approach leveraging large language models (LLMs) to automate regex generation for clinical entity extraction. Our method involves prompting LLMs to generate regex patterns from examples, followed by iterative refinement using a feedback loop. Despite regex limitations, this approach is practical for extracting frequently patterned information common in clinical texts, such as dates, specific data about medical procedures or event detection. Our experiments on Czech clinical notes show this method outperforms current SOTA genetic-programming-based methods for generating regular expression patterns from examples, especially when there are few of them. |
| Related projects: |