Improving Machine Understanding of Czech Medical Text Using Self-supervised and Rule-Based Data Augmentation

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F26%3AALR8QNGJ" target="_blank" >RIV/00216208:11320/26:ALR8QNGJ - isvavai.cz</a>

Result on the web

<a href="https://link.springer.com/10.1007/978-3-032-00891-6_25" target="_blank" >https://link.springer.com/10.1007/978-3-032-00891-6_25</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-3-032-00891-6_25" target="_blank" >10.1007/978-3-032-00891-6_25</a>

Result language

angličtina

Original language name

Improving Machine Understanding of Czech Medical Text Using Self-supervised and Rule-Based Data Augmentation

Original language description

Medical doctor decision-making benefits from the development of effective support software. But for software to accurately interpret meaning and assist in clinical contexts, high-quality annotated health record data must be available for training and evaluation. This paper addresses this issue in the Czech language context, detailing a stage in a unique electronic health record (EHR) bootstrapping project. Using over 42 million words of Czech oncology records, we curated the creation of the CSEHR dataset: over 62,000 words of text with manually annotated medical concepts, out of which over 12,000 have been developed through multiple stages of review to serve as ground truth. We are leveraging this seed data to bootstrap larger annotated corpora, enabling scalable development of Czech healthcare NLP applications. This paper focuses on combining two data augmentation approaches. Approach 1, semi-supervised, consists in automated dataset augmentation using self-annotation to increase annotation density. Approach 2, based on distant supervision, consists in manual development of rules for improving annotations in training data. Results show that combining these two approaches on training data and fine-tuning an XLM-RoBERTa model for entity recognition increases the token classification F1 score by more than 5 points. This demonstrates the promise of this technique in further bootstrapping steps.

Czech name

—

Czech description

—

Type

D - Article in proceedings

CEP classification

—

OECD FORD branch

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Project

—

Continuities

—

Publication year

2026

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Article name in the collection

Modeling Decisions for Artificial Intelligence

ISBN

978-3-032-00890-9

ISSN

—

e-ISSN

—

Number of pages

13

Pages from-to

315-327

Publisher name

Springer Nature Switzerland

Place of publication

—

Event location

Valencia

Event date

Jan 1, 2026

Type of event by nationality

WRD - Celosvětová akce

UT code for WoS article

001585667300025