AI Translation: How to Preserve Meaning in Your Translations

👁 117 reads

We would use AI for translation to prepare a draft version that we compare against the source. First, we define the audience, purpose, and meaning of key terms. We don’t just ask for fluent text. We want to preserve conditions, negations, names, and the author’s level of certainty. Natural-sounding language only matters when the translation says the same thing.

What to establish before translating

We separate translation from adaptation. With translation, we guard the information from the original. With adaptation, we can change examples or the method of explanation, but we do it consciously. We don’t want the model to decide on its own that an inconvenient fragment is unnecessary and remove it under the guise of improving style.

We specify the target language and its variant if it matters to the audience. We add the document’s subject and necessary terminology. If a word has several possible meanings, we provide context or ask for a question. We don’t expect the model to guess the industry based on a single sentence.

What to check in DeepL

DeepL describes translation of entire files while preserving formatting elements. Available formats and material size depend on the plan. Before starting, we would check support for our file and the expected result. Source: DeepL, file translation.

Separate documentation describes creating a glossary, which is a set of terms and preferred equivalents. We would consider it for recurring vocabulary. We would check availability for the needed language pair and working method. Source: DeepL, creating a glossary.

A glossary doesn’t exempt us from checking sentences. A correct term equivalent can end up in a misunderstood construction. That’s why we also check who performs the action, what the condition applies to, and whether negation was preserved. We don’t evaluate translation only by comparing word lists.

How we would manage the work

1. We prepare a complete fragment with context, instead of pulling sentences out of the document.
2. We specify the audience and purpose of the text.
3. We establish terms, names, and fragments that should remain untranslated.
4. We create a working translation and note ambiguous places.
5. We compare the original with the translation, sentence by sentence.
6. We check the entire file after export, including headers and captions.

With longer material, we maintain consistent terminology between fragments. We don’t allow the same interface element to be a button once and a tab another time just because sentences were processed separately. We record approved equivalents and return to them during review.

Ready-to-use prompts

We direct the following instructions to a text assistant. In a classic translator, we place only the material to be translated, and select settings in the interface. We don’t paste instructions as part of the document.

Translation with context establishment

Help me translate text. First, ask for the material, target language, audience, and purpose of use. Wait for a response. Preserve all facts, conditions, and the author's level of certainty. Don't add explanations to the translation itself. List ambiguous places separately and ask questions. Don't replace proper names with similar-sounding names.

Term list before translation

Ask me for the document and target language. Before translating, list industry terms, product names, and recurring concepts. Suggest their meaning in context, but don't present uncertain equivalents as approved. Ask if I have my own terminology. After my response, prepare a consistent glossary to use when translating the entire material.

Meaning consistency check

Help me check a translation. Ask for the original and translation, then compare the meaning of sentences. Pay special attention to negations, conditions, obligations, names, and omitted fragments. Separate meaning errors from stylistic suggestions. For each comment, show the relevant fragments of both versions. Don't consider a more fluent sentence better if it changes the content.

What not to expect

• We don’t expect error-free specialist translation without checking terminology.
• We don’t treat machine translation as automatic confirmation of a document’s suitability for formal purposes.
• We don’t assume that correct file appearance means all information was preserved.
• We don’t consider back-translation as independent proof of consistency with the source.
• We don’t expect recognition of the author’s intent that cannot be determined from context.

How we would check a document after translation

When the goal is to understand text while learning, we would also ask for an explanation of difficult constructions. We separate such commentary from the translation itself. We describe examples of language work in AI for learning English. We don’t add explanations to a document intended for the audience without clear need.

First, we check places where a mistake changes the recipient’s action. In instructions, this will be the sequence of actions and conditions for their execution. In a service description: scope and exclusions. We don’t start by smoothing adjectives if we still aren’t certain whether obligations were preserved.

Next, we view the document as a whole. We check headers, captions, tables, and broken sentences. If the text lengthened after translation, we check whether it fits in the designated fields. We don’t shorten it automatically at the cost of information needed for understanding.

For content intended to be spoken, we perform additional editing according to principles from the guide on AI for voice-over. If we need article adaptation, we use the approach described in AI for writing texts. We clearly separate translation from changing the method of presenting the topic.

If a document has significant consequences for the recipient, we would pass it to someone who knows the language and field for review. We don’t expect that another response from the same assistant will replace such evaluation. We choose the level of control appropriate to the application.

We would choose AI for translation to prepare a working draft and organize terminology. We accept the finished text only after comparing meaning and checking the file. Most important to us is preserving what the author actually wrote.

// read moreBest AI to Translate Contracts: ChatGPT vs DeepL and PrivacyGuidesDeepL vs Google Translate vs ChatGPT: Translation TestBenchmarksHow Much Does DeepL Pro Cost? Plans, Limits & DictionaryGuides

Kinetyra