UzMorphSim v5.1 GOLD: A Document- and Concept-Disjoint Uzbek Semantic Textual Similarity Dataset
Year: 2026
Type: dataset
Abstract: UzMorphSim is a research dataset and reproducibility package developed for semantic textual similarity (STS) experiments in the Uzbek language. The resource contains train, development, and test pair datasets, a document collection, automatically generated morphological features, source code for the UzMorphSim architecture and experimental pipeline, and a Jupyter/Google Colab notebook. The dataset comprises 1,000 Uzbek documents and 5,950 document pairs distributed across train, development, and test subsets. The accompanying morphological resource contains token-level linguistic information, including lemma, UPOS, morphological features, and morphemic representations. The package is intended to support research on semantic textual similarity, low-resource natural language processing, morphology-aware text representation, contextual chunking, transformer-based encoders, and bidirectional late interaction for Uzbek. This release corresponds to version 1.0.0 of the UzMorphSim reproducibility package. The package includes the dataset, source code, configuration files, tests, documentation, and an executable notebook. Important methodological note: the currently released train/dev/test configuration uses a pair-level split and is not document-disjoint or concept-disjoint. Therefore, results obtained with these splits should be interpreted as evaluation of new document pairs within an overlapping document/concept collection rather than direct evidence of generalization to completely unseen concepts.
Authors: Dauletov, Adilbek, Matyakubova, Noila
Institution: Alfraganus University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213640598
