Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Translationese as a Rational Response to Translation Task Difficulty

Maria Kunilovskaya

Abstract

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.

Translationese as a Rational Response to Translation Task Difficulty

Abstract

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.
Paper Structure (14 sections, 2 equations, 5 figures, 12 tables)

This paper contains 14 sections, 2 equations, 5 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Translatedness Score
  4. Translation Task Difficulty
  5. Experimental Setup
  6. Data
  7. Feature Preprocessing
  8. Machine Learning Models
  9. Results
  10. Translationese Classification
  11. Regression Results and Discussion
  12. Conclusion
  13. Appendix. Additional Details on Segment-Level Translationese
  14. Appendix. Regression Details

Figures (5)

  • Figure 1: Variance in translatedness explained by linear regression models across language pairs and mediation modes. The two panels compare difficulty components and indicator types on the x-axis.
  • Figure 2: Univariate analysis: Spearman correlation between translation difficulty indicators and translatedness score. Source text difficulty indicators (src_) appear first. Features expected to correlate negatively with translatedness are shown in bold, and IT features are in red on the y-axis. The bars for non-significant $\rho$ values are removed.
  • Figure 3: Segment vs. document feature importances: Average absolute SHAP values for translationese predictors by mode and target language on a shared scale. Darker shades indicate higher feature importance; blank cells indicate that a feature was not selected for the given task. Features are grouped by selection overlap: the top block shows predictors selected by all four classifiers, followed by predictors shared by mode (written, spoken) and by target language (German, English). Segment-level analysis dilutes feature importance and obscures patterns that are clearer at the document level. The grouping sizes further suggest that mode yields more consistent deviation patterns than translation direction.
  • Figure 4: SHAP beeswarm plots for document-level translationese classifications, showing selected features contributions to class predictions, sorted by importance for each mode-language combination. Rows correspond to mediation mode (written vs. spoken), columns to target language (German vs. English). Each point corresponds to an individual segment; colour scale from red (high) to blue (low) encodes the feature value, while horizontal position reflects the SHAP value (in the log-odds scale), i.e. the impact of that feature on the model output.
  • Figure 5: SHAP beeswarm plots for regression, showing the impact of selected features on the predicted translatedness scores. Features are sorted by mean absolute SHAP value, indicating their relative contribution to the deviation from the model's base value for each mode-language combination. Rows correspond to mediation mode (written vs. spoken), columns to target language (German vs. English). Each point corresponds to an individual segment; colour scale from red (high) to blue (low) encodes the feature value, while horizontal position reflects the SHAP value (in the log-odds scale), i.e. the impact of that feature on the model output.