Cross-Lingual Equivalence and Translation Asymmetries in Parallel Corpora: A Computational Analysis
Year: 2026
Type: conference-paper
Abstract: Parallel corpora serve as a critical resource for bilingual lexicography, machine translation, and cross-lingual natural language processing. However, the assumption of one-to-one translational equivalence between language pairs is frequently violated in practice, particularly across typologically distant languages. This paper investigates cross-lingual equivalence asymmetries in Uzbek-English and Turkish-English parallel corpora, focusing on three categories of translation divergence: lexical gaps, structural mismatches, and pragmatic shifts. Using a dataset of 50,000 sentence pairs extracted from publicly available parallel resources including OPUS and locally compiled educational texts, we apply alignment-based and distributional methods to identify and classify non-equivalent translation segments. Our analysis reveals that approximately 34% of sentence pairs exhibit at least one form of translational asymmetry, with lexical gaps being the most prevalent category (18.7%), followed by structural mismatches (10.2%) and pragmatic shifts (5.1%). We further demonstrate that agglutinative morphology in Uzbek and Turkish introduces systematic alignment errors in conventional word-alignment models. These findings have direct implications for the development of linguistically informed machine translation systems and parallel corpus annotation standards. The study contributes a taxonomy of cross-lingual divergence types applicable to Turkic-Indo-European language pairs and provides baseline metrics for future research in low-resource multilingual NLP.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224937196