Don't translate, work in the original: translating into English consistently hurts accuracy

Key result

The "I'll translate it into English so the model understands better" strategy isn't a life hack, it's a source of errors. This was proven on Dutch medical documents: every time the text was run through translation before analysis, accuracy dropped, consistently and significantly. The method lets you work with documents in any language without intermediate steps. Translation loses information before the model even starts thinking: professional vocabulary, nuances, and idioms vanish at the input. Modern large language models (LLMs) are trained on huge multilingual datasets; they understand your language directly, and that's always more accurate than going through a middleman.

Leveraging Open-SourceLargeLanguageModelsfor Clinical Information Extraction in Resource-Constrained Settings

Key Aspects of the Research:

The researchers tested how well open (non-commercial) LLMs handle extracting information from medical documents written in Dutch. They compared different models and found that mid-sized models (14B parameters) already deliver excellent results. They also identified which kinds of tasks LLMs are better suited for (e.g., extracting numbers) and which they handle worse (named entity recognition).

Key Result: First translating the text from Dutch into English (supposedly to "help" the model) consistently and significantly hurt extraction quality.

The Core of the Method, Explained:

The method you can pull from this study is dead simple and practical: always work with the LLM in the original language of your document or task, as long as the model supports that language.

Many people who don't work in English follow a flawed strategy: "I'll take my text, say, in Spanish, run it through Google Translate into English, and then send the English version to ChatGPT, because it knows English best." This study convincingly shows that's a bad idea.

Every act of translation is lossy. Subtle nuances, cultural details, specialized terminology, and idioms get lost. The model receives a signal that's already "damaged." Modern LLMs are trained on massive multilingual datasets, and they can pick up these nuances directly from text in the original language.

The practical method: 1. Identify the language of your task/text. 2. Write the prompt and provide the context in that same language. Don't use external translators as an intermediate step. 3. Trust the model's native abilities. If you need to analyze customer reviews in Polish, write the prompt in Polish and feed the reviews in "as is." The result will be more accurate than if you'd translated everything into English.

Practical Applicability:

Direct applicability: Maximum. Anyone working with an LLM in a language other than English can apply this finding right away. The simple rule: don't translate the source text into English before sending it to the chat. That alone improves the accuracy and relevance of the answer with zero extra effort.
Conceptual value: Very high. The study gives users the right "mental model" of an LLM. Instead of seeing it as an "English-speaking brain with a translator bolted on," you start to understand it as a truly multilingual system, for which translation isn't a help but a hindrance that adds noise and distortion.
Adaptation potential: The findings adapt easily to any field. Instead of medical terms in Dutch, it could be legal vocabulary in German, marketing slogans in Japanese, or technical documentation in Portuguese. The adaptation logic is simple: realize that domain-specific vocabulary is the most vulnerable to translation, so for any expert task, working in the original language is critical.

A Practical Example:

Say you're a marketer and want to analyze reviews of a new coffee drink that customers left in Polish.

**Role:** You are an experienced marketing analyst.

**Context:** Below are customer reviews of our new coffee drink "Leśna Fantazja." All reviews are written in Polish. Your job is to work **directly with the Polish text**, without translating it.

**Task:** Analyze the reviews and identify the 3 main advantages and 3 main drawbacks of the product that customers mention. For each point, give 1-2 quotes from the reviews in the original (in Polish).

**Response format:**
### Advantages:
1.  **[Advantage 1]:**
    *   Quote: "[quote in Polish]"
2.  **[Advantage 2]:**
    *   Quote: "[quote in Polish]"
3.  **[Advantage 3]:**
    *   Quote: "[quote in Polish]"

### Drawbacks:
1.  **[Drawback 1]:**
    *   Quote: "[quote in Polish]"
2.  **[Drawback 2]:**
    *   Quote: "[quote in Polish]"
3.  **[Drawback 3]:**
    *   Quote: "[quote in Polish]"

**Reviews to analyze:**
"""
1. "Kawa 'Leśna Fantazja' ma niesamowity, głęboki aromat. Czuć nuty jagód, ale cena jest trochę za wysoka jak na co dzień."
2. "Smak jest naprawdę oryginalny, nie piłem czegoś podobnego. Niestety, opakowanie jest bardzo niewygodne, ciężko się otwiera."
3. "Idealna na poranny rozruch! Bardzo intensywna i pobudzająca. Szkoda tylko, że jest tak słodka, wolałbym wersję bez cukru."
4. "Zapach powala, serio, cała kuchnia pachnie lasem. Ale jest za droga. Może kupię od czasu do czasu jako luksus."
5. "Super pomysł z tymi leśnymi nutami. Moim zdaniem trochę za słodka, cukier zabija część smaku."
"""

Why It Works:

This prompt works well because it follows the study's findings directly:

Working in the native language: The data is provided in Polish. This lets the LLM analyze the text directly, catching nuances like "poranny rozruch" (a morning kick-start) or "zapach powala" (the smell knocks you off your feet), which a machine translation into English might drain of their emotional color or render too literally.
An explicit instruction against translation: The phrase Your job is to work directly with the Polish text, without translating it further reinforces this behavior, stopping the model from "cutting corners" by quietly using English as a go-between language.
Preserving terminology: The product name "Leśna Fantazja" (Forest Fantasy) and specific flavor descriptions ("nuty jagód" - berry notes) are processed in their original context, which makes the analysis more accurate. Translation could introduce artifacts.

Another Practical Example

Say an HR specialist wants to quickly assess a résumé from a candidate in Germany for a technical role.

**Rolle:** Du bist ein erfahrener technischer Recruiter.

**Kontext:** Unten findest du den Lebenslauf eines Kandidaten für die Position "Senior Java Entwickler". Der Lebenslauf ist auf Deutsch. Deine Aufgabe ist es, **direkt mit dem deutschen Text zu arbeiten**, ohne ihn zu übersetzen.

**Aufgabe:** Fasse die wichtigsten Informationen aus dem Lebenslauf zusammen. Konzentriere dich auf die folgenden Punkte:
1.  **Erfahrung:** Wie viele Jahre Berufserfahrung hat der Kandidat insgesamt und in der Java-Entwicklung?
2.  **Schlüsseltechnologien:** Liste die 5 wichtigsten Technologien auf, mit denen der Kandidat laut Lebenslauf gearbeitet hat.
3.  **Ausbildung:** Welchen relevanten Hochschulabschluss hat der Kandidat?
4.  **Rote Flaggen:** Gibt es offensichtliche Lücken im Lebenslauf oder etwas, das im Interview geklärt werden sollte?

**Format der Antwort:**
*   **Gesamterfahrung:** [Antwort]
*   **Java-Erfahrung:** [Antwort]
*   **Top 5 Technologien:** [Liste]
*   **Ausbildung:** [Antwort]
*   **Mögliche rote Flaggen:** [Antwort]

**Lebenslauf zur Analyse:**
"""
**Name:** Klaus Schmidt
**Berufserfahrung:**
*   **Softwarearchitekt, TechSolutions GmbH, Berlin (2019 - heute)**
    *   Entwurf und Implementierung von Microservices mit Java 11+, Spring Boot, und Kafka.
    *   Leitung eines Teams von 5 Entwicklern.
    *   Einsatz von Docker und Kubernetes für die Bereitstellung.
*   **Softwareentwickler, CodeCrafters AG, München (2015 - 2019)**
    *   Entwicklung von Backend-Systemen mit Java 8 und Hibernate.
    *   Arbeit mit PostgreSQL und Oracle Datenbanken.
**Ausbildung:**
*   **Master of Science in Informatik, Technische Universität München (2015)**
"""

Why This Example Works: The Mechanism.

This prompt is effective because it applies the same principle of preserving the language context to another field that's very sensitive to terminology:

Preserving technical vocabulary: German technical terms and job titles such as Softwarearchitekt, Entwickler, Informatik are analyzed directly. A translation into English might render them correctly, but any less standard term or abbreviation risks getting mangled.
Avoiding ambiguity: A prompt in German (arbeite direkt mit dem deutschen Text) ensures the model interprets the context within the German job market and education system. That matters because education systems and career paths differ from country to country.
Direct data extraction: The model finds and extracts the data (years, technology names) from the German text, which lowers the chance of errors that could creep in along the "German text -> machine translation into English -> model analysis" pipeline. Every extra step is a potential point of failure.

Usefulness Score: 88

Main Evaluation Criteria

A. Relevance to prompting techniques: Medium. The study doesn't introduce new techniques, but it clearly demonstrates the effectiveness of standard approaches (role-play, structuring, zero-shot CoT) and, more importantly, tests a fundamental interaction strategy (which language to write the prompt in).
B. Better conversational answers: High. The main finding directly affects the accuracy and relevance of answers for all non-English-speaking users.
C. Direct practical applicability: High. The core finding, "don't translate text into English before sending it to an LLM," can be applied immediately by anyone, no technical skills required.
D. Conceptual value: High. The study offers a key insight into the "mental model" of an LLM: models trained on many languages work better with them directly, while translation is a "noisy," "lossy" channel that degrades the result. It also shows that not all tasks are solved equally well in zero-shot mode.
E. New useful practice: Yes, the work fits these clusters:
- Cluster 2 (LLM behavioral patterns): The key result is identifying the pattern that pre-translating into English consistently degrades model performance.
- Cluster 7 (Reliability and stability): Following the study's findings directly improves the reliability and accuracy of answers for tasks in languages other than English.

Practicality checklist: * Provides ready-made phrases/structures for prompts? Yes (in the paper's appendix). * Reveals non-obvious features of LLM behavior? Yes (the negative effect of translation). * Suggests ways to improve answer consistency/accuracy? Yes (work in the native language).

Numerical Usefulness Score

The study gets 88 points out of 100. That's a very high score, because the work delivers one piece of practical advice that's extremely important and universal for a huge audience of non-English-speaking LLM users.

Arguments for a high score: * Busting a myth: The work empirically debunks the popular approach of "I'll translate into English first, since the model knows it better, and then send it." For many users this is a counterintuitive finding that immediately improves the quality of their work. * Universality: Although the study used highly specialized Dutch medical texts, its main conclusion about the harm of pre-translation applies to any language and any subject area. * Conceptual clarity: The study helps users understand that an LLM isn't just "an English model with a translator." Nuances, idioms, and specialized terminology are handled better directly, without translation as a "middleman."

Counterarguments (why the score isn't 95+): * Narrow focus: The paper's main context is medical reports and comparing open-source models, which may put off an everyday user working on marketing or creative tasks. * No new techniques: The work doesn't propose revolutionary prompt structures (like Chain-of-Thought, for example); rather, it validates and compares existing approaches under specific conditions. The value isn't in "how to write" but in "which language to write in."

Research digest

Key result

Principle

Work in the original language, period. Dutch text → straight into the model. Polish → straight into the model. Spanish → straight into the model. Every intermediate step is a potential point of distortion that you can simply skip. This is especially critical for professional language: medical terms, legal wording, technical jargon are exactly the words machine translation mangles first. Add an explicit instruction to the prompt, "work with the text in its original language, don't translate it," to stop the model from internally switching to English as a go-between language.

Why it works

LLMs are trained on massive bodies of text in dozens of languages. The model doesn't "think in English"; it works with multilingual representations directly. Translating before you send adds noise at the input: the model receives an altered signal instead of the original. The irony is that you're trying to "help" the model, but you're actually feeding it broken data instead of clean data. On top of that, the study showed that mid-sized models, from 14 billion parameters up, already handle such tasks on par with much larger systems. No need to wait for access to the most powerful models.

When to apply

Any analysis of non-English text → especially for documents with professional terminology: medicine, law, engineering, finance, the fields where a single word carries a specific meaning and distorting it changes the substance. Good for: extracting data from documents, analyzing reviews, processing résumés, breaking down reports, all in the original language. NOT suitable for rare languages the model supports poorly; check those in advance.

Mini recipe

1. Check language support: make sure the model handles your language well: ask it a simple question in that language and judge the answer. 2. Write the prompt in the document's language: not "analyze this Polish text" but "przeanalizuj poniższy tekst," or at least provide the task and the data together in one language. 3. Feed the text as is: no intermediate runs through Google Translate, DeepL, or other tools before sending it to the model. 4. Add an explicit instruction: put a phrase in the prompt like work directly with the text in [language], don't translate it into another language. This locks in the model's behavior and removes internal "switching."

Examples

[BAD]: `Here are customer reviews in Polish. Translate them into English and tell me what they like about the product.` [GOOD]: `Jesteś analitykiem marketingowym. Poniżej znajdują się opinie klientów w języku polskim. Pracuj bezpośrednio z polskim tekstem, nie tłumacz go. Wyodrębnij 3 główne zalety i 3 wady produktu, cytując fragmenty oryginalnych opinii. [reviews in Polish]`: the model analyzes the nuances and emotional shading directly, with nothing lost in translation.

Apply this study to your task

Describe what you're working on. AI takes the method and the measured results from this paper and adapts them to your case, or writes a ready prompt built on them.

Based on Don't translate, work in the original: translating into English consistently hurts accuracy arXiv:2507.20859