DocMidFuse-MN: Mongolian long-text output safety detection via middle semantic awareness and document-level risk fusion
Abstract With the expansion of large language models into cross-linguistic applications, safety detection for generated content has become important for trustworthy deployment. Experimental observations indicate that, under semantically equivalent instructions, Cyrillic Mongolian outputs are often longer than Chinese and English outputs, causing potential risk information to be dispersed across long texts. Existing safety detection methods are mainly designed for high-resource languages and have limited ability to model Mongolian word order and long-text risk distribution. To address this problem, this paper formulates safety detection for Cyrillic Mongolian long-text model outputs and proposes DocMidFuse-MN. The method first applies sentence-level segmentation to reduce truncation and risk signal dilution. It then introduces an SOV-inspired middle semantic enhancement mechanism to strengthen the representation of risk objects and key semantic information. Finally, it integrates local chunk-level risk, middle semantic risk, and document-level risk for overall safety prediction. A manually reviewed Mongolian output safety detection dataset is also constructed. Under the dataset and experimental protocol of this study, DocMidFuse-MN achieves strong long-text detection performance with the Mongolian-specific encoder relative to the evaluated no-segmentation, hierarchical modeling, and general safety detection baselines, while maintaining competitive performance on short texts.
Authors
- Hao Wang (ORCID: https://orcid.org/0009-0001-8969-9772)
- Gang Wang
- Yan Li
Institutions
- Inner Mongolia University of Technology (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-13
- DOI
- https://doi.org/10.21203/rs.3.rs-10689691/v1
- Primary Topic
- Hate Speech and Cyberbullying Detection
- Type
- preprint