A TYPOLOGY-INFORMED METHODOLOGICAL MODEL FOR BUILDING AN UZBEK–CHINESE PARALLEL CORPUS

Abstract

Uzbek and Mandarin Chinese lie at opposite ends of the morphological spectrum: Uzbek is an agglutinative Turkic language with SOV order and a rich case system, whereas Chinese is an isolating Sino-Tibetan language with SVO order, function words and numeral classifiers, written without word boundaries. Our search did not identify a systematic, openly available Uzbek–Chinese parallel corpus; the only directly relevant study found builds a filtered, partly LLM-generated pseudo-parallel dataset from news crawls. This paper proposes a typology-informed methodological model for building such a corpus. It has five components: (1) a layered design separating a human-verified gold core, an automatic extension and a labelled synthetic layer, with provenance metadata; (2) asymmetric linguistic processing linked through Universal Dependencies; (3) multi-level alignment down to the morpheme, with correspondence-type labels; (4) a "unit–set of equivalents" lexical database; and (5) retrieval-augmented integration with large language models. A correspondence table and two author-composed sentence pairs show that most grammatical correspondences are not one-to-one. An evaluation protocol is specified for future empirical work.