Multi-language surveys

Why keying on question IDs is the whole design

The obvious way to run a bilingual consultation is to build the survey twice. It is also the way that ruins the analysis: two response sets, two sets of column headings, and a merge done by hand in a spreadsheet on the day the report is due.

Here, a translation is a layer of display text sitting over a question that keeps its own identity. The question ID never changes with language, and neither does the option value behind a choice. So a conditional logic rule written against "answered Bus" fires for a Welsh respondent who saw Bws, a score attached to that option is awarded either way, and the exported column is one column, not two.

The consequence is the reason to care about any of this: without canonical storage, analysis fragments by language. Minority-language responses become a small separate file that gets summarised in a sentence, if at all. With it, Welsh responses sit in the same counts as English ones, and the language each person read is still recorded if you want to look at them separately.

A language per link

The language a respondent reads is normally decided by the link they open. That makes distribution the place you choose it, which is usually where the decision already sits — the Welsh link on the Welsh-language poster, the Urdu link through the community organisation that asked for it.

The /pk alias exists because links get read out at meetings, printed on leaflets and typed in from memory, and a country code is easier to carry than a language code.

Once a survey carries more than one language, a switcher appears at the top of it as well, listing each available language in its own script. So the link decides what someone sees first, and the switcher covers the person who was handed the wrong one. What does not happen is guessing: the survey never picks a language from a respondent's browser settings, because a silent guess is worse than a visible choice.

Right-to-left, and what falls back

Urdu and Arabic render right-to-left: the reading direction, the alignment and the flow of the question through the page all mirror. That applies to the survey a respondent fills in and to the builder an author works in.

Fallback is the other half of making this survivable. If a question, option or heading has no translation for the language being viewed, the source wording is shown in its place, so a respondent never meets an empty question or a raw key on screen. That matters in the ordinary case: someone adds a question a fortnight into a live consultation and the translation catches up on Thursday.

The builder is translated too, which is a different thing

Per-survey translation is for an organisation that works in one language and needs to reach people in another. That is not the only situation.

The admin interface — the builder, the settings, the results screens — is available in ten languages. Open survey.noissime.com/pk and the whole product is Urdu, right-to-left, including the question type palette and the results views. A Pakistani organisation running the builder in Urdu writes Urdu questions because Urdu is the language it works in. The survey comes out in Urdu with no translation layer involved at all, and English becomes the thing you add later if you need it.

So if your staff and your respondents share a language, you do not need per-survey translation — you need the builder in that language, which exists. Per-survey translation is for the gap between the two.

Welsh: the argument is legal duty, not comprehension

The Welsh Language (Wales) Measure 2011 gives Welsh official status in Wales and sets out the principle that it should be treated no less favourably than English. The duties that actually bind a particular public body are the Welsh language standards imposed on it, listed in its own compliance notice — so read that, and take advice on it, rather than a vendor's summary. Consultation and engagement generally sit inside the service delivery standards. The general shape is that a consultation issued in English is expected to exist in Welsh at the same time, not later and not only on request.

Be precise about why, because the usual justification is the wrong one. Essentially all Welsh speakers also read English, so this is not about whether someone could understand the survey. It is about status and choice: the Welsh version has to be equally available, equally prominent and equally easy to complete. A Welsh link three clicks deeper than the English one, or published a week later, satisfies nobody reading the standards properly. One survey behind a pair of links is easier to keep genuinely equal than two builds that drift apart the first time a question is edited. Sector notes for councils are on the local government page.

The wording is yours to supply

NumoForms does not translate anything for you. It holds the translated wording, keeps it attached to the right question and option, serves it to the right link and falls back cleanly when a string is missing. Producing the Welsh or the Urdu is your job, whether that means a translator, a bilingual colleague or a machine translation tool.

If it is the last of those, treat the output as a first draft. It gets the structure and most of the meaning in place quickly, and for an internal pulse survey or a quick community poll that is usually fine. For a statutory consultation it is a starting point and nothing more.

Survey wording is unusually unforgiving. Response options have to stay distinct from one another after translation, scale labels have to keep their spacing between points, and a legal or planning term often has a settled Welsh equivalent a general-purpose model will not reach for. Have a fluent speaker read the translated survey end to end before you publish it. That review costs an hour and protects the whole dataset, because a mistranslated option does not produce a wrong answer you can spot. It produces a plausible one you cannot.

What stays in the respondent's own words

Canonical storage applies to structured answers: choices, scales, ratings, matrix rows. Free text cannot work that way. Someone answering an open question in Welsh writes Welsh, and that is what lands in your results and in the CSV. Budget for reading it in the same way you budget for translating the questions. Column layout and the "Other (please specify)" handling are covered on the exports and PDF page.