Translation in Qualitative Research: Key Decisions

Why translation is a method question

In cross-language research, the words you analyse are often not the words participants spoke. A translator chooses between possible renderings of every sentence, and some of those choices carry interpretation: whether a word means "duty" or "obligation," whether a proverb is translated literally or by meaning, whether a polite form signals respect or distance. Those choices shape codes, themes and quotes.

Qualitative methodologists have long argued that translators are not neutral conduits but participants in producing the data, and that their role should be acknowledged rather than hidden. You do not need to adopt a particular philosophical position to act on that. It is enough to make translation decisions deliberately, involve people who understand the context, and document what you did.

Speech also brings specific problems that written text does not: fragments, false starts, run-on sentences and register shifts that translate awkwardly. The article on translating spoken language explains those in general terms.

When to translate

There are three common points at which translation happens, each with trade-offs.

During the interview: An interpreter translates live. Fast, but the interpreter summarizes and filters in real time, and the recording holds two languages.
After transcription, before analysis: Full transcripts are translated, then coded in the analysis language. Common and easy to share with a team, but every code rests on a translation.
After analysis: Analysis happens in the source language; only codes, themes and selected quotes are translated for writing up. Keeps closest to participants' words but needs a researcher fluent in the source language.

Mixed designs are common: a bilingual researcher codes in the source language, translates the codebook for a co-coder, and translates only the quotes used in publications.

If you work with an interpreter in the interview, transcribe both languages from the recording, not just the interpreter's version. The difference between what the participant said and what the interpreter relayed is itself worth checking on important passages.

Who translates

The translator's background affects the result. Options include:

The researcher, if fluent in both languages. Strong contextual knowledge, but risks reading the data through the researcher's own expectations without anyone noticing.
A professional translator. Linguistic precision and speed, but may lack familiarity with the research topic, regional dialect or the community's vocabulary.
A bilingual community member or research assistant. Deep knowledge of local meanings, but variable translation experience and possible confidentiality concerns if they know participants.
A machine translation draft reviewed by one of the above. Faster first drafts, but errors in idioms, names, negation and emotional nuance are common, and review is not optional.

Whoever translates, brief them properly: the research question, the participants' context, any glossary of key terms, the level of literalness you want, and how to flag uncertainty. The guidance on choosing and briefing a reviewer for AI translation adapts well to research translators. Make sure anyone handling transcripts is covered by your ethics approval and confidentiality agreements; check with your ethics board how they expect this to be arranged.

Analysing in the source language or in translation

Analysing in the source language keeps you closest to meaning. Words with no direct equivalent, culturally specific metaphors and shifts in formality remain visible. This is generally preferable when the researcher is fluent and the team is small.

Analysing translations makes collaboration possible across a team that does not share the source language, and it suits large, multi-site studies. The cost is that every code depends on a translator's choice. Mitigations include coding with the source text alongside, marking terms that resist translation, and asking a bilingual team member to review codes that rest on a single word or phrase.

Hypothetical: a migration study in two languages

A researcher interviews thirty Arabic-speaking participants about access to healthcare. She is fluent in Arabic, but her two co-coders are not. She codes the Arabic transcripts herself, produces English translations of all transcripts for the team using a machine draft that she corrects line by line, and keeps a running list of terms that carry meaning the English loses, such as words for family obligation and for shame. Co-coders work on the English, but any code that rests on one of those listed terms is checked against the Arabic before it enters the codebook. Quotes in the final paper are re-translated by her from the Arabic, with the original wording available in a supplementary file.

Handling words that do not translate

Some terms will resist translation. Rather than forcing an equivalent, many researchers:

Keep the original word in italics, with a gloss on first use.
Add a translator's note explaining the range of meaning.
Record the term in a glossary that the whole team uses.
Discuss the term in the findings if it is central to a theme.

Names, places, institutions and local acronyms deserve the same care. The article on translating names and technical terms covers general rules for consistent treatment.

Checking key passages

You cannot back-check every sentence in a large dataset, but you can check what matters most:

Every quote used in a publication or report.
Passages that support a central theme or a surprising finding.
Passages where the translator flagged uncertainty.
Any segment containing idioms, humour, irony or emotionally loaded language.

Methods for checking include a second bilingual reader comparing source and translation, discussion between researcher and translator, and back-translation into the source language by someone who has not seen the original. Back-translation is useful for spotting meaning shifts, though it has blind spots: a smooth back-translation can hide a poor forward translation if both translators make the same assumption.

Documenting translation decisions

Readers and examiners should be able to see how translation shaped your data. A short translation log, kept from the first interview, covers most of what is needed:

Who translated each transcript, their language background and relationship to the community.
When translation happened relative to coding.
Whether a machine draft was used, which tool, and how it was reviewed.
Glossary decisions and changes over time.
Passages checked, by whom and by which method, and what changed.

In the methods section, summarize this in a paragraph or two. Some researchers also present key quotes in both languages so readers can see the original wording.

Pitfalls and limits

Translating too early and discarding the original. Always keep the source-language transcript and audio.
Smoothing participants' speech into polished prose in translation, which changes how they come across.
Assuming machine translation handles dialect, slang or code-switching well. It often does not.
Using a translator who knows participants, without considering confidentiality.
Quoting translated passages without checking them against the original.
Leaving translation out of the methods section entirely.

For the practical workflow of running user research sessions in other languages, including recruiting, sessions and stakeholder reporting, see the article on translating UX research sessions; this article stays with the methodological choices.

How mydubly can support cross-language interviews

mydubly can produce a source-language transcript and a machine-translated transcript from the same recording. You choose an interview file from your device, mydubly detects the spoken language automatically, transcribes it, and translates it into one of 21 supported languages, returning the transcript in both languages, plus a timestamped version. The audio translator page shows how that works for recordings.

Treat the translation as a draft for a fluent reviewer, never as final data. The translation engine works on chunks of speech of roughly 30 seconds, so context from earlier in the interview is not always carried forward, and pronouns, formality and terms of art can shift between chunks. There are no speaker labels in either language. One job produces one target language. And whether participants' audio may be processed by an external service is a question for your consent forms and ethics approval; check with your ethics board or institution before uploading.

Cost is one credit per minute for the transcript, including the translated transcript, so a 45-minute interview costs 45 credits (4.5¢).

Next step

Before translating anything, write a half-page translation plan: when translation happens, who does it, what you will analyse, how you will check key passages, and where the log lives. Share it with your supervisor or team and keep it with your data management records. Then transcribe and translate one interview under that plan, following the steps in the interview transcription use case for the transcript itself, and adjust the plan before scaling up.

Frequently asked questions

Should I analyse interviews in the original language or in translation?

If you or a team member are fluent in the source language, analysing in that language keeps you closest to participants' meaning. Analysing translations makes team collaboration easier but every code then depends on a translator's choices. Many studies combine both: coding in the source language, translating the codebook, and translating only quotes for publication.

Can I use machine translation for research interviews?

As a first draft that a fluent person corrects, it can save time. It is not reliable on its own for idioms, dialect, irony, names or emotional nuance. Record that a machine draft was used and how it was reviewed, and check with your ethics board whether sending participant audio or text to an external service is permitted.

Do I need to report translation in my methods section?

Yes. Readers should know who translated, when translation happened relative to analysis, how translations were checked and how untranslatable terms were handled. A short paragraph drawn from your translation log is usually enough.

Is back-translation required in qualitative research?

It is one checking method among several, not a universal requirement. It works well for spotting meaning shifts in key passages. Discussion between researcher and translator, or review by a second bilingual reader, can be equally useful, and many studies combine methods.

Should published quotes appear in both languages?

It is good practice when the original wording matters to your argument or when a term resists translation. Space limits often mean the original goes in a supplementary file or footnote. Check the journal's guidance and consider whether showing the original could identify the participant.

What should a translator brief include?

The research question, the participants' context, a glossary of key terms, how literal the translation should be, how to mark uncertainty and how to handle names. Include confidentiality expectations and make sure the translator is covered by your ethics arrangements.