Linguistic and Computational Solutions for Translating Lexical Lacunae in Neural Machine Translation Systems

In recent years, the exponential growth of neural machine translation (NMT) systems has led to significant improvements in automated multilingual communication. However, these systems still struggle with the accurate and contextually appropriate translation of lexical lacunae—words or expressions that are deeply embedded in one culture and have no direct equivalents in another. This research proposes a hybrid methodology that combines linguistic classification, ontology-based semantic mapping, and Transformer-based NMT adaptation to enhance the translation of culturally and conceptually bound lexical gaps. A total of 182 lacunae expressions were identified across Uzbek-English and Turkish-English corpora and categorized into three main types: cultural, administrative-legal, and lexical-conceptual lacunae. These were encoded in a domain-specific bilingual ontology, which provided semantic descriptors and translation strategies. The ontology was integrated into the NLLB-200 neural translation framework via a lacunae-detection and input-enrichment module. Experimental results demonstrated a significant improvement in translation quality. The enhanced model achieved BLEU score gains of +6.3 for Uzbek-English and +4.7 for Turkish-English translation tasks, and a 43% increase in lacunae-aware adequacy score (LAAS). Qualitative analysis confirmed that the enriched translations better preserved cultural and semantic integrity. The findings indicate that combining linguistic insights with machine learning techniques can address one of the key limitations in current translation technologies and offer scalable solutions for low-resource languages.

Hali tarjima qilinmagan