Masterarbeit – ISWM01

This thesis examines the context-faithfulness of contemporary neural translation systems when applied to Old Testament texts, with context understood not only as lexical and syntactic adequacy but also as hermeneutical and theological interpretability. The study addresses how translation quality differs between classical neural machine translation (NMT) systems and large language model (LLM) approaches, and which model characteristics systematically affect structural fidelity, semantic precision, and the preservation of interpretive openness that is often constitutive for biblical texts. Methodologically, the thesis employs a triangulated evaluation design that combines quantitative and qualitative approaches. Quantitative assessment uses reference-based similarity measures and additional structural indicators to capture expansion tendencies, omissions, and non-source-grounded additions, thereby describing how closely outputs track the form and information structure of the source. Qualitative analysis complements this by focusing on genre-sensitive phenomena and the handling of theologically salient expressions, assessing whether model outputs remain compatible with philological and hermeneutical standards. To avoid restricting evaluation to modern translation norms alone, the framework includes diachronic reference points as interpretive checkpoints alongside established contemporary translations. Results show that translation behaviour is strongly genre-dependent: narrative texts tend to yield more stable and formally aligned outputs, whereas prophetic and especially poetic texts increase divergence, prompting more frequent restructuring and interpretive explicitation. Across models, quantitative scores prove to be informative for identifying relative tendencies and error profiles, but insufficient as stand-alone quality indicators, since high formal fluency can coincide with hermeneutically consequential shifts, and certain deviations reflect necessary target-language explicitation rather than straightforward error. At the level of model classes, NMT systems generally exhibit greater structural consistency, whereas LLM-based systems display higher variability and a stronger propensity toward explicative, interpretive renderings. Overall, the thesis concludes that context-faithful translation in a theological sense cannot be guaranteed by any single system type without methodological safeguards. However, when deployed transparently and under hermeneutical supervision, AI-based translation can support theological research – particularly for orientation, comparative analysis, and hypothesis generation – if reproducibility, documentation, and critical source-text control remain central.