Azure-docs: There is no way to mentioned a sub Entity part of the pattern

[babak-kobari]

[Enter feedback here]

As explained in decomposable entity tutorial, I have created an entity named Phonecall with a sub entity as CallerRole which uses a phrase list as descriptor. the phrase list contains roles such as CEO, CIO,etc.

I have to create a pattern matching following utterance

The CEO has called me yesterday

^{I DON'T KNOW WHAT TO PUT HEAR REFERRING TO CallerROle ENTITY } [((has|’s)] ((called|phoned|buzzed)|(ringed|spoken|talked)) [to] me [on] [{datetimeV2}] [(to|regarding) {Purpose}]^

---

Document Details

⚠ Do not edit this section. It is required for docs.microsoft.com ➟ GitHub issue linking.

• ID: 18a5f287-70ea-bfb1-34eb-9236510e65f6
• Version Independent ID: 3ae73060-6cb2-5f8f-0022-a71e96d76654
• Content: Tutorial: extract structured data with machine-learned entity - LUIS - Azure Cognitive Services
• Content Source: articles/cognitive-services/LUIS/tutorial-machine-learned-entity.md
• Service: cognitive-services
• Sub-service: language-understanding
• GitHub Login: @diberry
• Microsoft Alias: diberry

[YutongTie-MSFT]

@babak-kobari
Thanks for the feedback! We are currently investigating and will update you shortly.

[YutongTie-MSFT]

@diberry Hi Dina, do you have any idea regarding to this issue?

[diberry]

@YutongTie-MSFT Asking product group...

[diberry]

My understanding is this is not supported. A pattern can only reference the top/parent component. I'll add this to the docs. #please-close

[babak-kobari]

So then I am failed to understand the main reason for using the decomposition model and creating subcomponents, as we need to create them separately anyway.

for example :

Let say we have three intents as CreateTask, CreateEmail and CreateMeeting.

These intents are correlated in the real world. I mean one task can cause certain emails and meetings or one meeting can result in several tasks and emails and so on.

Lets say that we want to analyze the tasks, meetings, and emails related to the same subject through the client app, hence I need an entity named “Subject” which is a common entity, across all of them as a ML entity.

From the other side, it is advised in V3 to create a ML top-level entity along with their sub-components

Since it is not possible to have a common sub-component across two entities, then I need to create a separate entity as “Subject” and also create a sub-component for each of top level entities, and then assigned “Subject” to CreateEmiil.subject, CreateTask.subkec and CreateMeeting.subject as a descriptor

So if sub-components can not be referred to/ used in other parts of the whole model such as patterns or other entities then why do we need them?

[diberry]

@babak-kobari - Thank you for this example - it always helps.

@hazemelh @cahann - I think this example can be handled with a machine-learned entity - can you provide any specific guidance about how you would build that out for this example.

[hazemelh]

@babak-kobari Hello, trying to better understand what the top level entity you're looking for contains and looks like, and why you'd need to share a sub component across components.

A subcomponent is by nature a detail specific to the entity you are decomposing. So having a shared Subject removes the detail, as each Subject is a Subject within a specific context, they must be trained separately.

What you could do is train a separate Subject extractor, and for each subcomponent Subject entity, you could use the separate Subject extractor as a descriptor. This may allow you to have similar results across subjects, assuming labeling is consistent.

Sub components allow you to decompose a specific detail in a larger entity built with parts, it's value comes in extracting more granular information and grouping them together.

For example, take a Pizza ordering app, and each pizza can include sizes, toppings, quantities. When extracting the order "one large pizza with extra mushrooms and two small pizzas with pepperoni" you can group that "one large pizza with extra mushrooms" has size large, quantity one, and toppings mushrooms, while the other pizzas have size small, quantity two, and toppings pepperoni. Previously, you wouldn't be able to map which quantity belonged to which pizza, and the attributes of each pizza, you would merely extract 2 quantities (one and two), 2 sizes (small and large) and 2 toppings (mushrooms and pepperonis) and would have to figure out which ones matched with which. This can get even more complicated in more difficult orders.

Let me know if this helps.

[babak-kobari]

Hello, I guess it is more clear now, Thanks for the clarification

[diberry]

please-close