translation agent garbled the reply: language detection fix

TL;DR

Garbled translations usually mean the source language was misdetected: the agent thought English was Spanish, or a mixed-language ticket got one wrong label for the whole thing. Translating from the wrong source produces confident nonsense. The fix is explicit language detection per message, and never translating text that's already in the target language.

The query

translation agent garbled the reply: language detection fix

Use this when

• Translated replies come out garbled
• The agent translates already-English text into broken English
• Mixed-language tickets confuse the translator

Not for

• Translation quality or style debates
• Human translator workflows
• Terminology glossary management

Steps

1. Detect language per message, not per ticket

Run language detection on each message independently. Tickets mix languages constantly: the customer writes Spanish, the agent's macro is English, the customer replies in Spanglish. One label per ticket is always wrong for someone.

Expected output: a per-message language label on every translated message.

2. Skip translation when source equals target

If the detected source language is already the reply language, don't translate. Round-tripping English through a translator is how 'already-English' becomes garbled English. Detection first, translation only on mismatch.

Expected output: no translation applied when source matches target.

3. Set a confidence floor for detection

Short messages ('ok', 'gracias') detect unreliably. Below a confidence threshold, keep the original and flag for human review instead of guessing. A wrong guess garbles; a flag just waits.

Expected output: low-confidence messages flagged, not translated.

4. Log source, target, and confidence

Record the detected source language, target, and confidence with every translation. When garbling is reported, the log shows whether detection or translation was at fault, and you fix the right stage.

Expected output: translation logs carrying language pair and confidence.

Variant phrasings

ai translation wrong language

Steps 1 and 2: per-message detection, skip on match.

chatbot garbled translated reply

Step 4's logs distinguish detection faults from translation faults.

Why it happens

Translation pipelines assume one source language per input, but support conversations are multilingual by nature. The detector picks the majority language, the minority-language messages get translated from the wrong source, and the output is fluent garbage. Per-message detection costs little and fixes the whole class.

Edge cases

• Proper nouns and product names shouldn't be translated. Keep an exclusion list.
• Right-to-left languages need UI handling beyond the translation itself.
• Machine translation of legal or compliance text needs human review. Never auto-send those.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_zCGWqXh8nWzKVL-Zlil9Yw