Building a Conversational Language Understanding model

Question 1

which 3 factors are captured by CLU model to support a conversation

Answer 1

Utterance
Intention
Entities

Question 2

What would be a reasonable mapping of utterance, intention and entities in programming terms

Answer 2

Utterance can be mapped to user or system input

Intention can be mapped to a sub routine e.g. GetDevice, TurnLightOn

Entities are extracted from utterance input as potential objects e.g. DateTime, People, Locations…

Question 3

Two categories of Azure AI Language features and 3 examples of each

Answer 3

Built in features: Sentiment, Language, Key phrases, summarisation of text, PII, PHI, orchestration workflow, named entities, text analytics for Health

Learned features:
CLU, QnA, custom named entities, custom text classifcation

Question 4

How is patterns used to improve CLU

Answer 4

patterns are used to disambiguate utterance to determine correct user intention, in case similar terms are used to mean different things

Generally, multiple version of utterance are used to train CLU to understand the same intention by using different version of phrasing. However, patterns are used for the exact opposite where similar utterance are used to determine difference intention e.g. Is light on? VS Turn light on, where difference structure and punctuation attached to ‘light on’ leads to different intent