Microsoft Copilot | Windows Copilot | 기능
Windows Copilot에서 제공하는 특정 기능 또는 향상된 기능
Understood. Thank you for your assistance. Please let me know if you need any additional information from me.
Classic orchestration fails to trigger on a near-exact Korean utterance (particle difference)
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(281.6, 73%, 37%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EKN%3C/text%3E%3C/svg%3E)
KIM NURI
0
평판 포인트
Environment: Copilot Studio, standard harness, classic orchestration (generative AI orchestration = No). Agent primary language: Korean (ko-KR).
Setup: A custom topic has 8 trigger phrases, including the exact phrase 부적합 나왔어요 ("a nonconformity came up").
Issue: The utterance 부적합이 나왔어요 — identical except for the Korean subject particle 이 — does not trigger the topic. Meanwhile, shorter and looser utterances containing the same keyword do trigger it.
| Utterance | Triggers? |
|---|---|
부적합 나왔어요 (registered phrase) |
(to be filled in) |
| -------- | -------- |
부적합 나왔어요 (registered phrase) |
(to be filled in) |
부적합이 나왔어요 |
No |
부적합이요? |
Yes |
부적합은요? |
Yes |
부적합은 어떻게 처리하죠? |
Yes |
Since 부적합이요? triggers, the particle 이 alone doesn't break keyword recognition. The failing utterance is the one closest to a registered phrase, which is unexpected.
Already ruled out: Tested as the first utterance in a fresh test session (not mid-conversation topic switching). Topic is enabled and saved, no duplicate/overlapping topics.
Questions: Is this a known limitation of classic NLU tokenization for Korean? Is there guidance for Korean trigger phrases regarding particles (조사)?
Microsoft Copilot | Windows Copilot | 기능
-
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(246.4, 15%, 33%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EAL%3C/text%3E%3C/svg%3E)
Alex-L 12,985 평판 포인트 Microsoft 외부 직원 중재자2026-09-01T11:53:55.1666667+00:00 Please note that this is KO-KR forum, all threads should be posted in Koreans, however, I will still use English for your case today.
Hi KIM NURI
I’m not aware of any documented limitation that specifically states Korean particles (조사) are normalized in classic orchestration topic matching. Classic orchestration relies on NLU similarity matching against trigger phrases, and near-identical variations can sometimes score differently than expected.
Since
부적합이요?and other particle variations are matching, this does not appear to be a simple issue with the particle이itself. It may be a classic NLU classification edge case for that specific utterance.As a workaround, consider adding common Korean variants as trigger phrases, for example:
-
부적합 나왔어요 -
부적합이 나왔어요 -
부적합이 발생했어요 -
부적합 발생
If the behavior is reproducible in a minimal agent with no competing topics, it would be worth raising a support case or product feedback item, as most Korean speakers would reasonably expect both phrases to match the same topic.
-