Reputation: 1204
This is a question that was originally asked as a comment to another question (LUIS adding patterns to intents is not taking any effect) related to LUIS Patterns. Migrating it to its own post to make the information more readily searchable to the community.
"How do you get [LUIS] phrase list to work properly? I have never been able to get it to recognize the values in the actual entity. I've always had to create a list entity and add synonyms to the normalized values."
Upvotes: 1
Views: 45
Reputation: 1204
Summary
Phrase lists are used to "boost" the signal of another entity, but relies on the entity that you are trying to boost a signal for to have enough data to identify it in at least some cases.
So let's say you properly have a simple Apple
entity defined, and LUIS detects values properly like Golden Delicious
, Granny Smith
, Cosmic
, etc. It could maybe even properly detect a brand new (made-up) apple crop, Flamingo
, even if you did not explicitly label Flamingo
value as the Apple
entity in an utterance.
LUIS uses machine learning based off the context of utterances to identify, "Oh yeah, normally an Apple
entity can occur here, so Flamingo
must be an Apple
." Flamingo
is also an English word, so LUIS has "background knowledge" of this word already (for example, could know based off of the English language that it's a noun).
But what if a university created a new apple crop name that did not exist in English--like Jirio
, for example. LUIS has no background knowledge of that word, as it did with Flamingo
. So you could create a phrase list to help identify Jirio
properly as an Apple
.
I can create a new Apples
phrase list, and include LOTS values like Golden Delicious
, Granny Smith
, Jirio
, etc. And when LUIS looks: "I don't know what Jirio
is, but I do know what a Golden Delicious
is. A Jirio
looks like it's the same thing as a Golden Delicious
according to the Apples
phrase list, and thus is more likely to label Jirio
as an Apple
entity.
Related StackOverflow Posts that are also helpful (with screen shots!):
Upvotes: 2