Detecting Sufi Terminology in Uzbek Literary Prose Using NLP and Large Language Models: A Transformer-Based Approach

Abstract

Automated Sufi term extraction in Uzbek literary prose is constrained by dataset scarcity and high contextual polysemy. Addressing these bottlenecks, this study introduces a 52,347-token annotated corpus containing 1,843 instances across 134 term types (Cohen's kappa = 0.847), structured around four categories: Mystical States, Spiritual Stations, Relational Roles, and Doctrinal Concepts. Benchmarking three extraction frameworks (a TF-IDF lexicon baseline, fine-tuned XLM-RoBERTa NER, and contrastive span prediction) on identical splits establishes contrastive span modeling as the top approach at F1 = 0.806. This configuration outperforms XLM-RoBERTa NER by 4.5 points and the TF-IDF baseline by 16.6 points. Joint Cyrillic-to-Latin normalization and morpheme-aware tokenization supply further gains of up to 4.1 F1 points. Error diagnosis identifies contextual polysemy (29.5%) and out-of-vocabulary entries (18.5%) as primary failure modes. These resources establish a baseline for Uzbek literary NLP and Islamic digital humanities.

More by Gulnoza Salimova

• Gold-set and Probe for LLM Interpretation of Contested Images in Classical Uzbek PoetryFigshare · 2026
• A Pilot Annotation Dataset for Five-Line Takhmis Stanzas in Classical Uzbek Poetry (Ogahiy)Figshare · 2026
• A Lexicon and Detector for Folklorism in Uzbek Poetry (Iqbol Mirzo)Figshare · 2026
• Bioethical-Communicative Judgments Across Training Years: Cross-Year Evidence from Uzbek Medical EducationJournal on Efficiency and Responsibility in Education and Science · 2026 · Citations: 1