Hi All, I think this can be accomplished with LUIS but I don't know if this is the best way or there are other tools for it. I have a list of 1,000 terms like: Cash Paid For Income Tax Cash Paid For Interest And Income Taxes I have another list with millions of terms entered by users, and I need to find the best match to that master list of 1,000 records. So for example these two terms: Cash Paid For Tax and Interest Dollars Paid For Interest Or Income Taxes would resolve to "Cash Paid For Interest And Income Taxes". If I was doing this in LUIS I think I would make each term an Intent, and the millions of items to match would be Utterances. Since a single LUIS instance can only have 500 intents I'd have to set up a Dispatch approach with 500 intents each. Then to test it I'd use the batch testing to process 1,000 at a time. Would that work well? Is there a better tool I can use for this type of analysis?

Is LUIS the right tool for this scenario?

Aiden Kaskela 1 MVP

Hi All,

I think this can be accomplished with LUIS but I don't know if this is the best way or there are other tools for it.

I have a list of 1,000 terms like:

Cash Paid For Income Tax
Cash Paid For Interest And Income Taxes

I have another list with millions of terms entered by users, and I need to find the best match to that master list of 1,000 records. So for example these two terms:

Cash Paid For Tax and Interest
Dollars Paid For Interest Or Income Taxes
would resolve to "Cash Paid For Interest And Income Taxes".

If I was doing this in LUIS I think I would make each term an Intent, and the millions of items to match would be Utterances. Since a single LUIS instance can only have 500 intents I'd have to set up a Dispatch approach with 500 intents each. Then to test it I'd use the batch testing to process 1,000 at a time.

Would that work well? Is there a better tool I can use for this type of analysis?

romungi-MSFT 43,696 Reputation points Microsoft Employee

2021-12-17T09:06:19.577+00:00

@Aiden Kaskela Yes, LUIS would work for your scenario but I don't think you need to add every term that you have as an intent or label them as an intent. Based on the sample sentences provided you are basically having terms like cash or money paid for taxes or a service. This means you need to add these terms as utterances in your application and then map them to an intent. In this case the intent can be "paymentDone" or "duesSettled". Then you can basically train your application and test it with the terms that your customers use, all these should resolve to the "paymentDone" intent with a confidence score that should allow your application to process further based on a certain confidence threshold.

Infact, all the terms used by your users can be added as utterances to better train your application and LUIS is also intelligent to provide feedback on utterances used by end users which improves the scores of future terms or input. Also, you should identify entities in all these utterances during training and ensure they are mapped correctly to extract the same from results.
Aiden Kaskela 1 Reputation point MVP

2021-12-17T14:17:14.57+00:00
@romungi-MSFT - thanks for the response.

I may have given bad example. In my case the 1,000 terms are financial classifications which can't be combined, so a better example would be:

Income Loss From Continuing Operations Before Income Taxes Extraordinary Item Noncontrolling Interest

Income Loss From Continuing Operations Before X Y Z

Income Loss From Continuing Operations After X Y Z

If I made them a single "Income Loss From Continuing Operations" intent I'd lose the classification level I need.

The docs here "https://learn.microsoft.com/en-us/azure/cognitive-services/luis/luis-concept-intent" says that "Intents that are too similar can make it more difficult for LUIS to distinguish between them. Intents should be varied enough to capture the main tasks that the user is asking for"

So my follow up questions are 1) Do these 1,000 terms need to be intents to get the level of matching I'm looking for? 2) If they are intents, are they too similar for LUIS to be effective?

Thanks again,
Aiden
romungi-MSFT 43,696 Reputation points Microsoft Employee

2021-12-21T09:40:04.253+00:00

@Aiden Kaskela Yes the utterances do look to similar to group them. You can create more utterances and entities to distinguish the second part of the utterance like "Before", "After", "Income Tax" etc. These can be classified as entities like "Timeline" for before, after and "Events" for Income tax related terms.

The final JSON should list these under the classified intent along with the entities. The client application should be able to use this result and process the same in your application. I would recommend to try with sample data and check if the response from LUIS works for your application.

Share via

Is LUIS the right tool for this scenario?