Vietnamese Content Is Rewriting, Not Translating

Where a literal translation starts to sound wrong

Literal translation carries the words across but not the job the sentence was doing. Source copy written for one market rests on things that audience already knows. Strip that footing away, swap the words into Vietnamese, and the grammar holds while the reader loses the thread of what is being asked of them.

The first thing to break is sentence architecture. Korean, and to a lesser degree English business writing, tolerates a long stack of modifiers ahead of the noun they qualify. Vietnamese generally places the qualifying material after the thing being qualified, so a front-loaded pile of clauses forces the reader to hold several unattached ideas until the sentence finally resolves. The habit that causes this is insisting that one source sentence become exactly one target sentence. Splitting it in two solves the problem, but translators often will not split, because splitting looks like editorializing.

Dropped subjects cause the same trouble from a different direction. A great deal of business writing leaves the actor implied and trusts context to supply it. Vietnamese reads more naturally when the actor is present. A literal rendering either leaves that slot empty or fills it with the wrong party, and in service terms, incident notices, or anything touching liability, that empty slot is where a dispute later forms. Who checks, who replies, within what window: if the sentence does not say, each side supplies its own answer.

Compressed noun phrases travel badly too. A phrase along the lines of realizing sustainable growth through maximized customer satisfaction slides past unnoticed in the original because it reads as an idiom. To put it into Vietnamese you have to unpack it into a subject and a verb, and once unpacked it becomes obvious that the sentence carries almost no information. The empty sentence that hid in the source stands out in the target.

Cultural framing does not carry either. Company introductions written for a Korean audience often open with modesty and commitment. A Vietnamese reader finds nothing usable there. What the company builds, who is already using it, and how to reach someone need to appear early or the reader leaves. Rearranging that order looks like it exceeds the translator's mandate, and yet the order is precisely what determines whether anyone reads the second screen.

So in practice the work does not start sentence by sentence. Someone extracts the argument skeleton from the source, and the person writing Vietnamese looks at that skeleton and writes from the beginning. The output does not line up sentence to sentence and may not even have the same number of paragraphs. That is uncomfortable for anyone who wants a side-by-side table. In exchange, the reader understands.

Review criteria have to change with it. Checking whether every item in the source made it into the target pulls the piece back toward the source and preserves exactly the passages that felt stiff. Local-language copy should instead be judged on whether the reader can tell what to do next and where the prose snags. The person doing that reading should not have seen the source.

Address forms and tone are what build trust

Vietnamese forces a decision about how to address the reader in almost every sentence. A neutral English you becomes a specific choice: a formal corporate address, a familiar one, a kinship-based one. There is no way to abstain. Decline to choose and the prose floats; choose badly and the copy reads as either presumptuous or cold.

Business-to-business documents and product interfaces need different voices. Capability decks, quotations, and anything adjacent to a contract call for formal address. The screens a user sees every day need shorter, closer language. The complication is that the same company owns both voices. Without a decision recorded in advance about which channel takes which voice, every screen drifts, and users notice fast. One page is ceremonious, the next is nearly casual, and the whole thing reads as several companies stapled together.

Whether to use age-and-relationship address forms is its own decision. Using them makes marketing copy noticeably warmer while narrowing the set of readers who feel addressed. Using them in formal notices makes the notice feel lightweight. This is not something a writer should settle line by line while drafting; the brand has to settle it first. Left unsettled, the taste of whoever wrote that day becomes the voice of the company.

Regional vocabulary differences deserve an early decision as well. Some words differ between the north and the south, and neither is incorrect. The only question is where the customers are. Without a decision, pages mix, and later, when someone tries to standardize, there is no way to tell a deliberate choice from an accident. Recording each decision at the moment it is made saves an enormous amount of archaeology later.

The short strings turn out to be the hardest part. Buttons and notifications are short enough to invite direct translation, and short enough that tone shows through completely. Confirm, cancel, apply now: pushed into an imperative they read as curt, softened they no longer fit the control. This kind of copy should be settled on the actual interface, at actual width, rather than in a spreadsheet. Error messages carry the same risk, where a single word decides whether the system sounds like it is blaming the user or helping them.

How numbers, dates, and addresses are written belongs to tone as well. Date order, the way currency figures are grouped, and the order in which an address runs from small unit to large all follow local convention. Prose can read beautifully while a date format left over from the source market announces that the page was assembled somewhere else. Trust erodes at that level before anyone has evaluated the argument.

For that reason it is worth fixing address rules and formatting rules in a document. Which situation takes which form of address, how product and feature names are written, how dates and amounts appear, all in one place. It is the mechanism that keeps the voice stable when the person doing the work changes. Without it, six months later the same site holds three or four different personalities talking to customers.

Search terms do not translate

What people type into a search box is habit, not prose. Convert a source keyword into grammatical Vietnamese and you get a respectable phrase that almost nobody actually types. Search terms are not something to carry across; they are something to go and find again inside the market.

The most common mismatch sits between written and spoken registers. A formal phrase from the source, translated carefully, produces the kind of wording that appears only in documents. Meanwhile people search with something shorter, more colloquial, often clipped or abbreviated. The paradox is that the more a translator polishes, the further the page drifts from the terms people use. Elegant copy and findable copy are not automatically the same copy.

How much borrowed English to keep is a case-by-case call. In technology, many English words are used unchanged in ordinary speech. Replace all of them with native equivalents and the prose is tidy but nobody finds it. Keep everything in English and the page reads heavy and insider-ish. The workable approach is to decide word by word based on how each one is actually used, rather than applying a single rule across the site.

Search intent itself differs between markets. On the same topic, a reader in one market may ask about cost and timeline first while a reader in another asks something else first. That changes the order of what belongs on the page. Translate the source page and publish it, and the material answering earliest is answering somebody else's question. The writing is not bad; the sequence belongs to a different audience.

Collecting the real vocabulary is straightforward work. Pull it from the wording customers already used in their inquiries, from how local colleagues describe the task out loud, from the words that recur in relevant groups and forums. Treat tool volume figures as reference only and spend the attention on the context each term appears in. A single word can mean different things in different industries, and choosing on volume alone brings in readers with no relation to the business.

Once chosen, put those exact terms in the title and the opening paragraph and mark them as not to be edited. Editors downstream have a reflex to smooth wording, and the search terms disappear quietly. This happens often and is hard to trace afterward. Separating the fixed strings from the freely editable prose at the draft stage prevents a whole round of argument.

The same issue reaches inside the product. In Job Connect VN, the recruitment service built by Yeowubie Interaction, the way a job title is written is the search term. The official title an employer posts and the words a candidate types frequently differ, so the system has to carry both forms for the two sides to meet. At that point it stops being a translation question and becomes a question about how data is stored.

Designing for the habit of typing without diacritics

Vietnamese users routinely type without diacritics on a phone. They do it when they are in a hurry, and everyone still understands each other. If search and filtering compare only fully marked strings, existing data appears not to exist. The user concludes the system is empty and leaves, and nobody files a bug, because from their side nothing looked broken.

The design principle is compact. Store exactly what the user entered, and alongside it store a separate normalized value built for searching. Display the original, match on the normalized copy. Never overwrite the original, because once the marks are gone they cannot be reconstructed. This is an additional field, not a substitution.

Normalization means applying Unicode normalization to separate the marks from the base letters and then removing the marks. The character most often missed is the d with a stroke. It does not decompose the way the others do, so common strip-accents routines skip it, and both its uppercase and lowercase forms need explicit handling. That single character is enough to make a set of very common names unfindable, and the failure usually surfaces only after a few thousand records exist.

Case and whitespace belong in the normalized value too. People space things differently, writing a term as two words or as one. Folding whitespace in the normalized copy means either form returns the same result. Name sorting and duplicate detection should run on that same value. If two records differing only in diacritics are treated as different people, duplicate accounts accumulate silently.

Slugs in the address bar and file names should be diacritic-free from the outset. A slug, once published, is better left alone, since changing it breaks links already shared. Email subjects, downloaded file names, and any string that has to cross several systems are safer without marks as well. One weak link mishandling the encoding is enough to corrupt the string, and it is usually a customer who discovers it.

Search should tolerate partial matches and mild misspellings. If only exact matching is supported, users try once or twice and conclude the record is not there. Autocomplete that surfaces values which actually exist reduces how precisely anyone has to type. Making the system accommodate how people type is consistently cheaper than requiring people to type correctly.

Worth stating plainly: this is a design problem, not a translation problem. Assigning it to whoever owns content will not solve it. It has to be settled alongside the data model, before screens are built. Retrofitting means reprocessing everything already stored, and while that runs, search results are unstable. This is why a conversation about Vietnamese content keeps turning into a conversation about engineering.

A working method that lets content keep coming

A page written well once is worth less than a method that lets the next page get written. Treat local-language content as a one-time translation project and six months later all that remains is aging copy. The product changes and the questions change while the Vietnamese pages sit still.

Separating roles comes first. Deciding what to say, writing it in Vietnamese, and reading it back for judgment are three different jobs. When one person holds all three, content stops during any week that person is busy. The person judging should read without having seen the source, because anyone who has read the source gets pulled along by its structure and stops noticing where the prose snags.

A terminology list is needed earlier than most teams expect. Product names, feature names, address conventions, date and currency formats, all in one place. It is the mechanism that gets a new contributor using the same words. It does not need to be large. Twenty entries covering the things that keep getting confused start paying off in the first week.

The content management structure has to allow for this as well. If each page in one language is hard-bound to exactly one page in another, neither side can grow independently. Local readers sometimes need an explanation the source never included, and sometimes the reverse. The structure should permit different section counts and different lengths per language. What it must expose is when each language version was last updated.

Update signals should not depend on memory. When the source changes, the other language versions should visibly show that they have fallen behind. Even a simple marker tells the next person where to start. Without one, a few months later nobody knows which page is current, and once nobody knows, nobody dares edit, and the content freezes.

Keep the writer inside the organization where that is possible. Sending work out sentence by sentence means explaining the context afresh every time, and the explaining costs more than the writing. Langtori, the tool Yeowubie Interaction built to connect learners with teachers, makes the point plainly: the two sides call the same thing by different names. Write the interface copy from one side's vocabulary and the other side feels the product is not addressed to them. That kind of gap is only caught by someone who sits with it daily.

This article exists in three languages, and none of them is a translation of the others. Only the order of the arguments matches; the examples and the sentences were written again for each audience. That approach makes a side-by-side comparison table impossible and leaves something readable in each language instead. What local-language content needs to produce is not a comparison table but sentences that change what the reader does next.