Graphlogic.ai documentation

Natural language recognition in Agents

Natural language recognition

Natural language recognition by

@Agent⁠

s is performed using the

@NLU⁠

@Slot⁠

. The

@NLU⁠

@Slot⁠

is trained on a

@Training Dataset⁠

of the

@Intent⁠

s included in it.

Learn more about creating Intents:

Intents tab⁠

Learn more about the operation of the

@NLU⁠

@Slot⁠

Intent Recognition | NLU slot + Intent and Fallback subslots⁠

The training sample of the Intent

⁠

@Training Phrase⁠

s are phrases with the same or similar meaning, with the help of which a person can express a specific intention (

@Intent⁠

), and on which the

@NLU⁠

model is trained in order to recognize these and all other variants of expressing the same intention (

@Intent⁠

). The quality of the

@Training Phrase⁠

directly depends on the quality of recognition in the

@Agent⁠

The goal to strive for when making a sample is to give the

@Agent⁠

as many different options for the formulation of the intention as possible (preferably within 30 phrases).

@NLU⁠

@Slot⁠

trained on a specific

@Training Phrase⁠

will recognize not only a message of the

@Bot User⁠

that completely repeats the

@Training Phrase⁠

from the sample, but also a message that is close in meaning.

Recommendations for filling the training sample of Intents

For successful recognition, it is recommended:

Choose from 10 to 30 training phrases-examples for each

@Intent⁠

⁠

@Training Phrase⁠

s of the same

@Intent⁠

should be synonymous with each other – denote the same intention.

The

@Training Dataset⁠

should be a set of phrases, meaningful sentences, and not a set of keywords or topics.

⁠

@Training Phrase⁠

s should be diverse. To do this, you should use different synonyms for words and different formulations of intentions.

⁠

@Training Phrase⁠

s should sound realistic. In order to understand how the

@Bot User⁠

s formulate questions, you can, for example, view the conversation history of the

@Bot User⁠

s with a consultant or a support operator in a chat.

Recommendations for the list of Intents

It is very important not only to correctly fill in the

@Training Dataset⁠

@Intent⁠

s, but also to correctly compile a list of

@Intent⁠

s themselves:

Similar in meaning

@Intent⁠

s should be combined. If the samples of different

@Intent⁠

s are very close in meaning, it is highly likely that the NLU will be "confused" between them.

⁠

@Intent⁠

s containing different intentions in meaning should be separated. An

@Intent⁠

that contains a lot of different intentions, although related to the same topic, is better divided into several separate ones. Thus, the

@Agent⁠

@NLU⁠

will be able to more accurately form an idea of the meaning of each

@Intent⁠

and more accurately recognize them in the future.

The key to high—quality recognition in an

@Agent⁠

is not only a competently compiled list of

@Intent⁠

s and

@Training Dataset⁠

, but also testing. Read more:

Testing and improving Agent recognition: selection of Confidence threshold⁠

Important: it is necessary to avoid a random coincidence of the
@Training Phrase⁠
in different
@Intent⁠
s — if the same
@Training Phrase⁠
is in different
@Intent⁠
s, then the
@Bot user’s Message⁠
that matches it will be assigned by
@NLU⁠
to one of these
@Intent⁠
s randomly.

The training sample of the Intent

Recommendations for filling the training sample of Intents

Recommendations for the list of Intents

Want to print your doc?
This is not the way.

Try clicking the ⋯ next to your doc name or using a keyboard shortcut (

CtrlP

) instead.