Data Augmentation Based on Selective Masking of Language Models for One Health Context

Read full text Summarise with VIRAT
Authors: Youssef Mahdoubi, Najlae Idrissi, Mathieu Roche A, Sarah Valentin
Cited by: 0
References: 26

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This study focuses on improving the performance of language models for two critical applications within the One Health context, specifically in epidemiological monitoring using textual data: (i) thematic classification across syndromic surveillance, biomedical and plant health domains, and (ii) detection of epidemic misinformation.A key challenge in these tasks is the limited availability of labeled textual data, which constrains the effectiveness of supervised learning methods.To overcome this limitation, we introduce two families of selective masking-based data augmentation strategies: lexical and non-lexical.Each family is implemented in a standard variant (Aug-SM-Lex and Aug-SM-NonLex), and a TF-IDFweighted variant (Aug-SM-Lex-TFIDF and Aug-SM-NonLex-TFIDF).We perform two complementary experiments: the first determines the optimal masking rate, while the second evaluates the proposed strategies against LLM-based text reformulation.Experimental results indicate that selective masking-based augmentation outperformed both LLM-based reformulation (Mistral-7B and GPT-Neo-1.3B)and baseline models trained on original data alone across three of the five evaluated datasets, with the best performance achieved at a masking rate of 20%.This suggests that selective masking is a promising approach, potentially more effective than computationally expensive LLM-based reformulation.

Fields of Study & Subjects

Topic Modeling Data-Driven Disease Surveillance Text and Document Classification Technologies

Publication details

Journal / source
Workshop on Linguistic Analysis for Health (HeaLing)
Pages
266–276
Publication date
1 January 2026
Language
EN