AI Translator - SDK.ac

Neural MT vs LLM Translation: What Actually Changed

Classic neural machine translation systems — Google Translate, DeepL, Papago — are encoder-decoder models trained specifically on parallel corpora. They translate one sentence at a time, with little or no memory of the surrounding document. That design makes them fast, cheap, and deterministic: the same input almost always produces the same output. But it also explains their most familiar failures. A pronoun in sentence twelve gets the wrong gender because the antecedent lived in sentence three. A product name is translated in the title and transliterated in the body. Formal address flips to casual halfway through a manual. Large language models translate differently. They read the entire passage inside one context window, so cross-sentence references, running jokes, and document-level terminology stay coherent over thousands of words. They also follow instructions: you can hand an LLM a glossary, a target register, a character limit, or a "do not translate brand names" rule, and it will usually comply. The trade-off is that LLMs are slower, cost more per word, and are non-deterministic — the same prompt can yield different phrasings, and a model may quietly paraphrase or smooth over an ambiguity instead of flagging it. A practical rule: use dedicated NMT for high-volume, low-stakes text where speed matters, and use LLM translation when context, tone, and terminology consistency matter more than latency. This tool builds the instruction-rich prompt that makes the LLM path work.

Formality Is Not Optional: 반말/존댓말, Keigo, and T-V

English mostly hides formality in word choice, so English speakers routinely underestimate how much of a translation is register, not vocabulary. Korean grammaticalizes politeness in the verb ending of nearly every sentence: 해체 (banmal, intimate), 해요체 (polite everyday), and 합쇼체 (formal, used in news and business documents) are all correct translations of the same English sentence, and picking the wrong one can read as rude or absurdly stiff. Japanese layers this further with keigo — plain form, です/ます polite form, and separate honorific (尊敬語) and humble (謙譲語) verb sets, where the verb changes depending on whose action is being elevated. European languages carry the T-V distinction: tu/vous in French, du/Sie in German, tú/usted in Spanish — and business German still defaults to Sie with people you would happily first-name in English. Machine translators must guess a register, and they guess from surface cues that are often missing in short input. This is the single most fixable translation failure: state the register explicitly. Tell the model who is speaking, to whom, and in what setting — "a customer support reply from a company to an adult customer, polite but not stiff (해요체)" removes the guesswork entirely. When translating into a language you cannot read, ask the model to state which register it chose, so a native reviewer can verify the choice rather than reverse-engineer it.

Weak prompt:
"Translate to Korean: We're sorry, but your refund could not be processed."

Strong prompt:
"Translate to Korean. Context: customer-support email from an
e-commerce company to an adult customer. Register: polite 해요체,
warm but professional. Keep the brand name 'PayNow' untranslated.
Text: We're sorry, but your refund could not be processed."

Result (weak):  환불을 처리할 수 없다.        <- abrupt, near-반말
Result (strong): 죄송하지만 PayNow에서 환불을
                 처리하지 못했어요.            <- correct register

Why Japanese and Korean Break Translators: は/が, Dropped Subjects, Counters

The hardest translation errors are the ones that produce fluent, grammatical, wrong output. Japanese は and が are the classic case. は marks the topic ("as for X"), が marks the grammatical subject, and swapping them changes emphasis or meaning: 私は行きます is a neutral "I will go," while 私が行きます means "I am the one who will go" — an answer to "who is going?". Translators flatten both to "I will go" and lose the contrast; going the other way, they must invent a distinction the English never expressed. Both Japanese and Korean are pro-drop languages: subjects and objects vanish when context supplies them. A Korean sentence like 어제 봤어요 is literally "yesterday saw" — who saw what must be recovered from prior sentences. Sentence-by-sentence NMT guesses, and its favorite guess is "I," which is why so much translated Korean dialogue has phantom first-person subjects. LLMs with the full passage in context guess far better, but only if you actually give them the passage instead of one isolated line. Then there are counters and classifiers. Japanese counts flat objects with 枚, machines with 台, small animals with 匹; Chinese requires measure words (一本书, 三张桌子); Korean pairs native and Sino-Korean numerals with different counters. Getting a counter wrong is instantly visible to natives and invisible to you. Always translate with surrounding context, and have a native speaker check anything customer-facing.

Context-free input (1 line):
  "봤어요?"
NMT output:
  "Did you see it?"   <- "you"/"it" are pure guesses

Same line with context supplied:
  "A: 어제 새로 나온 예고편 봤어요?
   B: 아직요. 오늘 밤에 보려고요."
LLM output:
  "A: Did you watch the new trailer that came out yesterday?
   B: Not yet. I'm planning to watch it tonight."

Rule: never translate dialogue line-by-line.
Paste the whole exchange and translate it in one pass.

A Working Translation Workflow: Context, Glossary, Back-Translation

Professional translators do not start typing immediately; they build context first. You can borrow that discipline in three steps. First, front-load context. Before the text itself, state the domain (legal, medical, marketing, casual chat), the audience, the purpose, and any constraints such as maximum length for UI strings. Two sentences of context routinely improve a translation more than any amount of post-editing. Second, maintain a glossary. Every product, feature, and role name should have exactly one approved translation, and you should paste that list into every translation request. Terminology drift — the same feature called three different things across your help center — is the most common defect in AI-translated documentation, and a pasted glossary eliminates most of it. Include the terms you do NOT want translated: brand names, code identifiers, legal terms of art. Third, verify with back-translation, carefully. Translate the output back to the source language with a fresh request (not in the same conversation, where the model can see the original and simply echo it). Back-translation is a smoke test: it reliably catches omissions, negation flips, and wrong numbers, but it cannot certify nuance, register, or naturalness — a stilted translation often back-translates perfectly. For anything with legal, medical, or financial consequences, back-translation supplements a human native review; it never replaces one.

Glossary block to prepend to every request:

GLOSSARY (use exactly these translations):
- "Workspace"      -> 워크스페이스 (never 작업 공간)
- "billing cycle"  -> 결제 주기
- "seat"           -> 시트 (a license, not 좌석)
DO NOT TRANSLATE: SDK.ac, OAuth, API key, webhook

CONTEXT: SaaS admin console UI. Max 20 characters
per string. Register: 합쇼체 for buttons, 해요체 for
descriptions. Return a table: source | translation.

What Still Goes Wrong — and When to Call a Human

Even with perfect prompting, certain categories fail often enough that you should check them by hand every time. Idioms and wordplay translate literally when the model does not recognize them: "break a leg" has become 다리를 부러뜨리세요 in real subtitle tracks. Personal and place names get inconsistent transliteration — 김서연 may appear as Kim Seo-yeon, Seoyeon Kim, and Kim Seoyeon in one document unless you pin a spelling. Dates and numbers silently change format: 03/04/2026 is March 4 in New York and April 3 in London, and Korean 억 (100 million) has no clean English unit, which invites off-by-a-factor errors in financial text. Units, currencies, and legal terms of art (consideration, force majeure, 전세) frequently have no true equivalent and need adaptation, not translation. Treat AI translation as a highly capable first draft. It is genuinely good enough to publish for low-stakes content — internal notes, community posts, support macros — after a quick read-through. It is not good enough, unreviewed, for contracts, medical instructions, safety warnings, or marketing copy whose tone is the product. The model will not tell you which sentence it was unsure about; it renders confident output either way. A native speaker reading the target text cold, without the source, remains the single highest-value quality check available.