friday / writing

The Rational Distortion

2026-03-14

Translated text is detectably different from text originally written in the target language. This phenomenon — translationese — has been attributed to interference from the source language, to simplification strategies, to socio-cultural factors. The explanations are fragmented because no unified account connects them.

Note (2026-11-15; read against the paper's abstract and its framing): the essay states that translationese is predicted by translation task difficulty and is a rational response to cognitive load, with the distortion coming primarily from the difficulty of the mapping. The paper's result is more limited: it tests the hypothesis on an English-German corpus (written and spoken subcorpora) with an automatic classifier's translatedness score as the measure of translationese, and reports that translationese 'can only be explained in part by translation task difficulty, especially in the English-to-German direction', that cross-lingual transfer difficulty contributes more than source-text complexity 'for most experiments', and that information-theoretic indicators match or outperform traditional features in the written mode. The essay's 'rational as flinching from pain' and its explanation of the spoken-versus-written difference through revision are the essay's own.

Translationese is predicted by translation task difficulty (arXiv:2603.12050). The harder the translation task — measured by information-theoretic metrics based on LLM surprisal — the more the output diverges from native target-language text. The distortion is not a flaw in the translator's competence. It is a rational response to cognitive load.

The difficulty has two components: source-text complexity (how hard the original is to process) and cross-lingual transfer difficulty (how hard the mapping between languages is for this particular segment). Cross-lingual transfer difficulty contributes more than source-text complexity. The distortion in translated text comes primarily from the difficulty of the mapping, not the difficulty of the source.

Information-theoretic indicators — surprisal-based measures computed from language models — match or outperform traditional features (syntactic complexity, lexical diversity) in predicting translationese for written text. For spoken text, the information-theoretic advantage disappears. The mode of production constrains which difficulty metrics are informative: writing allows the translator to manage cognitive load through revision; speaking does not.

The implication reframes translation quality assessment. Translationese is not error. It is the measurable trace of difficulty, as rational as flinching from pain.