Translationese as a Rational Response to Translation Task Difficulty

Translations systematically diverge from texts originally produced in the target language, a phenomenon widely referred to as translationese. Translationese has been attributed to production tendencies (e.g. interference, simplification), socio-cultural variables, and language-pair effects, yet a unified explanatory account is still lacking. We propose that translationese reflects cognitive load inherent in the translation task itself. We test whether observable translationese can be predicted from quantifiable measures of translation task difficulty. Translationese is operationalised as a segment-level translatedness score produced by an automatic classifier. Translation task difficulty is conceptualised as comprising source-text and cross-lingual transfer components, operationalised mainly through information-theoretic metrics based on LLM surprisal, complemented by established syntactic and semantic alternatives. We use a bidirectional English-German corpus comprising written and spoken subcorpora. Results indicate that translationese can be partly explained by translation task difficulty, especially in English-to-German. For most experiments, cross-lingual transfer difficulty contributes more than source-text complexity. Information-theoretic indicators match or outperform traditional features in written mode, but offer no advantage in spoken mode. Source-text syntactic complexity and translation-solution entropy emerged as the strongest predictors of translationese across language pairs and modes.

💡 Research Summary

The paper “Translationese as a Rational Response to Translation Task Difficulty” proposes that the systematic linguistic differences observed between translated texts and original texts—commonly called translationese—are not merely artifacts of interference, simplification, or stylistic choices, but rather reflect the cognitive load imposed by the translation task itself. The authors operationalise translationese using a segment‑level “translatedness score,” which is the probability output of a binary classifier trained to distinguish translations from original texts in the target language. This classifier relies on a rich set of 58 linguistically motivated features (POS tags, dependency relations, discourse markers, etc.) that capture structural rather than topical cues, ensuring construct validity.

Translation task difficulty is modelled as two components: (1) source‑text comprehension effort and (2) source‑to‑target transfer effort. Both components are quantified using information‑theoretic (IT) measures derived from large language models (LLMs) and traditional structural metrics. For source‑side difficulty, the authors compute average surprisal (AvS) of source tokens using monolingual GPT‑2 models (src_gpt_AvS). For transfer difficulty, they use (a) target‑side surprisal from neural machine translation (NMT) models conditioned on the source (mt_AvS), (b) entropy of translation solutions (tot_entropy) estimated from a massive parallel corpus, and (c) alignment strength (mean_align) based on multilingual BERT attention. Complementary structural features include tree depth, branching factor, mean dependency distance, clause count, lexical density, word length, multi‑word expression frequency, numeral and proper‑name counts.

The experimental corpus consists of English‑German bidirectional data, split into written (documents) and spoken (speech) sub‑corpora. Each aligned segment is annotated with the translatedness score and all difficulty features. Regression analyses are performed to assess how much variance in translationese can be explained by the difficulty predictors, both separately and jointly, and to compare the explanatory power of IT‑based versus traditional features.

Key findings: (1) Translation difficulty accounts for a modest but consistent portion of translationese variance—approximately 20 % of the total variance in the English‑to‑German direction, with weaker effects in the opposite direction. (2) Cross‑lingual transfer difficulty (mt_AvS, tot_entropy) contributes more than source‑text complexity (src_gpt_AvS) across both language pairs, supporting the hypothesis that the act of mapping between languages drives many translationese phenomena. (3) In the written modality, IT‑based measures (especially src_gpt_AvS and mt_AvS) match or outperform traditional syntactic/lexical features in predicting translatedness. In the spoken modality, however, the advantage disappears, likely due to the noisier nature of oral data and the limited ability of current LLMs to capture prosodic or disfluency cues. (4) The strongest individual predictors across all settings are source‑side syntactic depth (src_tree_depth) and translation‑solution entropy (tot_entropy), indicating that complex source structures and high ambiguity in target‑language realizations both promote translationese.

The authors interpret these results as evidence that translators adopt rational strategies to reduce cognitive effort: they may simplify complex source constructions, rely on more literal renderings when transfer uncertainty is high, or increase explicitation to compensate for processing difficulty. This reframes translationese from a “defect” to a cost‑minimising adaptation. The paper also acknowledges limitations: LLM‑derived surprisal values can be model‑dependent, the spoken data may suffer from alignment errors, and the current approach does not directly measure real‑time cognitive load (e.g., eye‑tracking or EEG). Future work is suggested to incorporate diverse language pairs, real‑time interpreting data, and multimodal cognitive measurements, as well as to explore ensemble surprisal estimates to mitigate model bias.

In sum, the study offers a unified, cognitively grounded explanation for translationese, demonstrates the practical utility of information‑theoretic difficulty metrics, and opens avenues for more nuanced quality‑estimation and translator‑support tools that account for the inherent difficulty of the translation task.