800 agents. We recently deployed a new Bot Flow to handle basic customer inquiries. The Natural Language Understanding (NLU) model is configured for Japanese (ja-JP). We are finding that the bot accurately detects the intent when we test it with simple phrases, but it fails completely when real customers use polite honorifics (Keigo), such as starting a sentence with “Osoreirimasu ga” (Excuse me, but…). The NLU engine seems to get confused by the polite filler words and drops the actual intent. Do we need to manually add every variation of Keigo to our utterance lists, or is there a way to train the bot to ignore standard Japanese polite prefixes?
I am transitioning from PureConnect to Genesys Cloud. In PureConnect we used strict grammar files, but the GC NLU relies on statistical models. You should not try to build a “Stop Word” list for honorifics. Instead, you need to provide more training utterances that include the Keigo prefixes. The AI needs to learn that “Osoreirimasu ga password wo wasuremashita” maps to the “Password Reset” intent just as strongly as “Password wasureta”.
Hey! I just inherited a GC org and I am learning the bot stuff too. To add to what As noted above, check out the “Confusion Matrix” in the bot analytics! It shows exactly what the bot thought the customer was saying when it got confused by the polite words. If it is mapping those polite phrases to a “Greeting” intent by mistake, you can use the optimization tool to explicitly tell it “No, this specific phrase belongs to the Password intent”. It takes some manual tuning at first but it gets way smarter over time!