Do Sensory Details Survive Translation? A Hypothesis and a Test Plan

The question

In a Turkish scene nobody is “sad” or “afraid”; there is only a cup of tea gone cold, a refrigerator motor that falls silent, a folded napkin. When this scene is translated into English, does the translator or translation system name the emotion to help the reader? Add a simile? Or replace the physical detail itself, for example the nuance of a Turkish onomatopoeic word, with a more general word?

The question matters for Objective Projection, because the method's two constitutional rules, the Emotion Embargo and the Exclusion of Similes, require something to be absent from the surface of the text. Even if the source text follows these rules, translation can break that condition.

What translation theory says

In his essay “Translation and the Trials of the Foreign”, published in 1985, Antoine Berman lists twelve deforming tendencies at work in literary translation. Among them are rationalization, clarification, expansion, ennoblement and qualitative impoverishment. According to Berman, these tendencies often operate unconsciously and shape the text even in a translation that seems relatively “good”. The English translation of the essay appears in The Translation Studies Reader, edited by Lawrence Venuti.

Shoshana Blum-Kulka (1986) put forward the explicitation hypothesis, which proposes that translated texts show more explicit cohesion than their source texts; the translator's interpretation of the source can make the target text more explanatory than the source. The hypothesis was debated in later research; it was shown that explicitation does not appear in the same way in every translation and every language pair, and that it is also affected by differences between language systems.

Lawrence Venuti, in The Translator's Invisibility (1995), discusses the tension in the Anglo-American translation tradition between strategies that favour fluency and domestication and foreignizing strategies that preserve the foreignness of the source text.

Hypotheses

H1 (measurable). In some share of scenes whose source contains no emotion name and no simile, translation adds at least one emotion name or simile. This is Berman's clarification tendency and Blum-Kulka's explicitation hypothesis applied to Objective Projection's two prohibitions.

H2 (not measurable today). Translation may weaken the fineness of physical detail: sensory details such as Turkish reduplications and onomatopoeic words may turn into more general equivalents in English. This corresponds to Berman's qualitative impoverishment. However, the features needed to measure it, like Objective Projection's four techniques, cannot be labelled reliably across human and machine annotators. H2 therefore stands, for now, only as a question.

H3 (exploratory). Machine translation adds emotion names more often than human translation. We discussed the observation that language models lean toward naming emotion rather than building it as delivery-mode drift; we also discussed how a similar drift may spread between agents during summarization in What Gets Lost When AI Agents Summarize for Each Other.

This distinction between what can and cannot be measured is what the Doctrine contributes here: prohibitions can be counted on the surface, techniques cannot. We discussed what readers build in their minds from details in Situation Models.

Why the existing data are not enough

The Objective Projection dataset is bilingual and contains English counterparts of the 300 original Turkish main scenes. But these counterparts are not literal translations: according to the dataset's source paper, both language versions were independently rebuilt to comply with the rules. Looking for rule violations in a text written to comply with those rules would be circular. For this reason the existing English scenes cannot be used to test H1.

Another trap of cross-language comparison can be seen in the dataset itself. According to the detector's marks, micro-focus is present in 79.0% of the Turkish scenes and in 40.5% of the English scenes. Interpretation: This difference is not a difference of language or quality; according to the source paper, the Turkish scenes contain a structural field that the detector uses for this feature, while the English scenes do not. What looks like a language difference may be a data-pipeline difference; a translation test needs the same care.

Test plan

The plan below is a draft; it is not registered on OSF and no data have been collected. Registration must happen before data collection, and the result must be published whatever it turns out to be.

Source texts. Original Turkish scenes from the dataset that two independent humans have confirmed contain no emotion name and no simile.

Conditions. Each scene is translated in two ways: (A) by a professional human translator, (B) by uninstructed machine translation. An optional third condition (C) is machine translation with an instruction stating that the source contains no emotion name and no simile. The translator does not know the purpose of the study.

Primary outcome. The share of scenes in which the translation adds at least one emotion name or simile. The decision is made by two independent bilingual annotators who do not know the translation condition; the detector is used only as an aid. Agreement between annotators is reported separately.

Analysis. For each condition, the addition rate is reported with a confidence interval. Because the same scenes are translated in both conditions, A and B are compared with an exact McNemar test.

Precision. The effect of the number of scenes when estimating a single condition's addition rate:

Power. With no prior data, the rates below are assumptions; the table shows power according to the shares of scenes in which only B adds and only A adds:

Interpretation: If the difference is large (a net difference of about eight points), 200 scenes give sufficient power; if the difference is small (about four points), power stays below 0.80 even with 300 scenes. The target effect must be stated explicitly in the registration.

What the Doctrine cannot say here

It cannot say that translation adds emotion names; this has not yet been measured. It cannot say that adding emotion names makes a translation worse; a translator may be making a deliberate choice for target-language readers, and Berman's critique is also a position on translation ethics, not a measured quality judgment. It cannot measure today whether the fineness of physical detail is lost in translation. Even if findings emerge, it cannot say that they hold for language pairs other than Turkish and English.

Limitations

This is a hypothesis article; it contains no measurements. The source scenes were written with a single methodology, to comply with these rules; results may differ for literary texts written naturally. Emotion name and simile are the most reliably labelled of the Doctrine's features, but this reliability was measured largely on Turkish scenes and against a single human reference; the same reliability should not be assumed for English translations and must be measured again with two independent annotators. For use of the source texts, see the Prompt and SFT Guide; for carrying the showing principle into another medium, see Screenplay Action Lines.

Frequently Asked Questions

Does this article report a measured result?

No. The article presents a hypothesis and a test plan. Whether translation adds emotion names or similes has not yet been measured; until it is, this remains a hypothesis.

Why can't the English scenes in the dataset answer this question?

Because the English counterparts of the Turkish scenes are not literal translations; both language versions were independently rebuilt to comply with the rules. Looking for rule violations in texts written to comply with those rules would be circular. Testing requires the same Turkish scenes to be translated again, for this purpose.

Why are only emotion names and similes counted?

Because these two features are the ones that can be labelled most reliably among Objective Projection's rules. The loss of fine physical detail is also an important possibility in translation, but there is no reliable label to measure it today.

Does machine translation behave differently from human translation?

Unknown. The observation that language models lean toward naming emotion rather than building it suggests that a similar drift may occur in machine translation; but in this article it stands as an exploratory question, not a tested finding.

References

• Berman, A. (1985). La traduction comme épreuve de l'étranger. English translation: Translation and the trials of the foreign (trans. L. Venuti). In L. Venuti (Ed.), The Translation Studies Reader. Routledge.
• Blum-Kulka, S. (1986). Shifts of cohesion and coherence in translation. In J. House and S. Blum-Kulka (Eds.), Interlingual and Intercultural Communication (pp. 17–35). Gunter Narr.
• Bulut, L. (2026). Objective Projection Corpus: dataset source paper. doi:10.5281/zenodo.22841625
• Venuti, L. (1995). The Translator's Invisibility: A History of Translation. Routledge.

How Levent Bulut's works connect is shown on the knowledge graph, and the full list of Levent Bulut's registered works is on the corpus page.

Cite this article

@misc{bulut2026translation,
  author       = {Bulut, Levent},
  title        = {Do Sensory Details Survive Translation? A Hypothesis and a Test Plan},
  year         = {2026},
  howpublished = {\url{https://leventbulut.com/do-sensory-details-survive-translation/}},
  note         = {leventbulut.com}
}