Knowledge graph - NLP recognition errors

XO Platform
1

I’m having a lot of trouble getting the Knowledge Graph to understand basic utterances and give a correct response.

For example, for the basic query: “How do I set up voicemail?”, it seems that I need to enter over a dozen similar alternative questions (eg “setup voicemail”, “set-up voice-mail”, etc. etc) for the bot to give the correct response.

The bot can’t seem to handle the difference between “can” and “do” (eg “how can i setup voicemail” vs “how do i setup voicemail”…the bot cannot understand the second utterance).

I’ve included multiple synonyms for voicemail (“voice mail”, voice-mail") and setup (set-up, set up, activate, etc.etc), in the Synonyms&Concepts section… but the NLP doesn’t seem to pick them up unless I’ve included the word in a separate question in the KG input for the FAQ.

Often, the bot sometimes doesn’t even recognise exact matches to FAQ questions (eg “how do I setup voicemail”)… see picture.

I’ve saved the utterances and click “train” button, every time, but the bot still has errors.

It will take me a verry long time to setup a simple FAQ bot if i need to enter every single variation of words for every simple question. I thought the NLP would be more “intelligent”?

knowledge%20graph%20errors%201
knowledge%20graph%20errors%2011898×915 267 KB
knowledge%20graph%20errors%202
knowledge%20graph%20errors%2021887×921 282 KB
knowledge%20graph%20errors%203
knowledge%20graph%20errors%2031159×549 94.1 KB
knowledge%20graph%20errors%204
knowledge%20graph%20errors%2041196×583 84.6 KB
knowledge%20graph%20errors%205
knowledge%20graph%20errors%2051927×935 210 KB

Can you please help?
thanks

2

Check out the response in this thread.
It should help to explain how the KG works and how you need to structure the training.

https://community.kore.ai/t/most-of-the-faqs-added-give-fallback-response-until-we-add-few-tags-but-tags-added-for-some-faq-does-not-work-too/1150/4

3

Thanks for the very fast response Andy, much appreciated!

I found that the problem wasn’t because my KG was structured incorrectly, but instead it was a config issue in the Natural Language/Training section.

Under Thresholds/Configurations, I changed both Auto Correction and Bot Synonyms to “enabled”. That made the KG work as advertised :slight_smile:

I’m not sure if these are both set to disabled from the start (seems odd to disable the KG out of the box, maybe I disabled accidentally somehow), but changing these setting to enabled made the KG work properly.

4

Both of these items were exposed as options after the initial release of Kore.ai. When such things happen we generally leave the options as disabled to preserve backwards compatibility and not force a bot to account for unexpected behavior.

5