One Word, Three Names
Last month I published an article in four languages and then reread it. Something felt off. The Korean word for 「prompt」 appeared as 「提示詞」 in the Traditional Chinese version, 「提示语」 in the Simplified version, and a few paragraphs later in the same piece as 「指令」. None of these is a wrong translation. But when the same object changes its name three times inside one article, readers assume they are different objects.
Counting: 7 per Post
I picked ten articles written by the translation bot and counted how much 12 key terms drifted. On average, 7 places per post used a different rendering for the same term. The bot chose the most natural phrase sentence by sentence, and it did not remember what it had written in the earlier paragraphs.
The Fix Was One Table
I built a glossary. Four columns (Korean, Traditional Chinese, Simplified Chinese, English) and 12 rows of key terms. I pasted it at the very top of the translation bot's instructions and added two lines:
Always use the table's wording for any term listed in it.
List any new term not in the table separately at the end under 「New terms」.
The second line mattered more than expected. When the bot met a word that was not in the table, it stopped quietly deciding on its own and raised its hand instead.
Result
I reran the same ten articles, and mismatches dropped from 7 per post to 1. The remaining one was usually a genuinely new term, and it got added to the table the next day. The glossary has since grown from 12 rows to 19.
Notes for Today
• Translation quality depends more on term consistency than on how polished each sentence is.
• The shorter the table, the better the bot sticks to it.
• Do not let the bot silently pick a word that is missing from the table. Make it flag it.