Formulaicity of German Controlled Language as a Translation Challenge for Neural Machine Translation

This paper examines the formulaicity of the German, considering both its perceived and actual challenges for neural machine translation (NMT) into Russian. The study aims to conceptualize German as a controlled language for formulaic structures within the DeepL neural translator, investigating its criteria, restrictions, and output evaluation. The analysis is based on a corpus of 120 examples of German formulaic constructions, comprising two categories: verb-nominal phrases and binomial expressions. The data were sourced from the Digital Dictionary of the German Language (DWDS). We present a novel two-stage experimental design for analyzing translation choices applied to a random sample of these formulas. The methodological approach involves observing lexical-level rules for German as a controlled language. The analysis of lexical restrictions in German controlled natural language for formulaic constructions is combined with a contextual method for translation quality assessment. Our findings reveal a distinct strategy employed by the state-of-the-art DeepL translator for rendering stable formulaic expressions of varying structures. It is established that the neural approach to machine translation aims to replicate cognitive models of human thinking; however, the utilization of established translation solutions is identified here as an asset. The author emphasizes that this study does not seek to evaluate the commercial product DeepL itself, as it is a proprietary tool developed by a German corporation and tailored for texts with a specific stylistic profile.

German language, controlled natural language, neural machine translation, artificial intelligence, formulaicity, formulaic sequences, DeepL