ArchiveGene–UzMorphSim MultiDomain v1.1: An Uzbek Multi-Domain Resource for Semantic Similarity and Document Understanding
Abstract
ArchiveGene–UzMorphSim MultiDomain v1.1 is a multi-domain Uzbek research dataset developed for semantic similarity, adaptive model routing, document understanding, layout-aware processing, and document-orientation analysis. The resource comprises a primary corpus of 2,000 text documents for semantic-similarity experiments, complemented by 60 auxiliary document-understanding items: 24 form-understanding documents, 16 layout-annotated contract pages, and 20 rotation-labeled document images. The semantic-similarity component contains 6,876 document pairs with leakage-controlled document-level train, development, and test partitions. Of these, 4,060 pairs are UzMorphSim gold semantic-similarity pairs and 2,816 are ArchiveGene-derived silver pairs. Gold and silver labels are explicitly distinguished and should not be interpreted as having the same annotation reliability. The package additionally provides normalized tables, morphological features, structured ArchiveGene annotations, dataset manifests, schemas, statistics, train/development/test splits, source provenance information, and SHA-256 checksums to support reproducibility. The auxiliary components include form-understanding JSON data, layout-annotated document pages with CVAT annotations, and document images with verified rotation-angle labels. This dataset is intended for research on Uzbek natural language processing, semantic textual similarity, document-level similarity, low-resource and agglutinative language processing, reinforcement-learning-based model routing, document understanding, and computational efficiency. Licensing note: CC BY 4.0 applies to the original dataset contributions, annotations, derived labels, metadata, splits, schemas, and documentation created by the dataset authors. Any bundled third-party or source-specific materials remain subject to their original terms and conditions. See LICENSE and THIRD_PARTY_NOTICES.md included in the archive.