DIGITAL PHILOLOGY IN TRANSLATION RESEARCH: A COMPARATIVE ANALYSIS OF TRANSLATION DECISIONS BY LARGE LANGUAGE MODELS

Abstract

The article presents a comparative study of translation solutions produced by large language models within the methodological framework of digital philology. The aim of the study is to identify the specific features of translating English-language texts into Ukrainian using ChatGPT, Gemini, and Claude, to compare shared and model-specific translation solutions, and to determine the degree of their variability depending on the thematic and terminological specificity of the source text.

The empirical dataset comprised 30 English-language text fragments divided into three thematic and discourse-based groups: academic texts on digital philology, computational text analysis, and language technologies; official and analytical materials addressing artificial intelligence, digital transformation, and the regulation of digital technologies; and academic texts focusing on large language models and machine translation. Each fragment was translated by the three large language models using a standardized translation task, resulting in a corpus of 90 translations. The research methodology was based on a digital-philological approach combining quantitative and qualitative analysis. The translations were evaluated according to six criteria: lexical equivalence, terminological accuracy, semantic correspondence, stylistic adequacy, contextual interpretation, and textual coherence.

The results of the study revealed the interaction of two major tendencies – inter-model convergence and variability in translation solutions. The highest degree of inter-model convergence was observed in the translation of established terms and semantically unambiguous linguistic units with stable Ukrainian equivalents. Semantically and contextually sensitive units allowed for a broader range of terminological, lexico-semantic, and structural realizations. The nature of translation solutions was found to depend on the thematic, terminological, and discourse-specific characteristics of the source material. The quantitative comparison demonstrated generally high performance across ChatGPT, Gemini, and Claude according to the established criteria; however, the observed differences characterize the models’ performance within the investigated corpus and the specified experimental conditions and do not provide grounds for establishing a universal hierarchy among them.

It is concluded that translations generated by large language models can be conceptualized as a system of shared and model-specific translation solutions whose nature and degree of variability depend on the thematic, terminological, and discourse-specific characteristics of the source text. The proposed digital-philological approach integrates quantitative comparison with qualitative interpretation of the ways in which the linguistic and textual characteristics of the source text are reproduced and shifts the research focus from the overall assessment of LLM translation quality to the systematic analysis of translation solutions.