Does anyone on this forum know enough about Microsoft LUIS to defend it?  LUIS seems quite marvelous but also quite inefficient. Without becoming an expert, it is not clear if the capability exists in LUIS and, if it exists, if it is easy to use. So, bearing in mind my lack of LUIS expertise, I want to pose critical questions.

1. How much training and defining is needed to handle a sentence like “I want a loaf of bread”, when “I want” has 15 variations, “loaf of bread” has 4 variations, and the same mechanisms are needed for 10 varieties of cans of soup? It should be something like 15+4+10+1 definitions. What is the estimate for LUIS: O(N*M) or O(N+M) ?

2. Does LUIS support regular expression matching? EG can I define a pattern like [A-Z]* (meaning anything starting with a capital letter) and train or define that?

3. Does LUIS allow definition of implicit entities - entities that can be optional in a match, without decreasing the ‘score’? If so, are additional definitions needed?

4. Does LUIS support an organized architecture of topic relationships? Or is it just “one lump” of code, divided as smartly into files and directories as can be arranged by the developer.

5. Does LUIS support sentiment? Is there a good/bad value field, filled into LUIS’s ‘intent’ structure?

6. Does LUIS handle negations? (eg “I do not want a loaf of bread”).

7. Is there any support for nesting of definitions inside of definitions? Can “I want a loaf of bread” appear as part of some other statement like “I need to go shopping and order some bread”. 

8. How is LUIS on context? If we just talked about something, can I refer to that using “it” or “both”, or similar phrases with indeterminate words? Like “I want to buy it”. What about the more difficult context handling, like using a category name after a particular name was mentioned, like “yes that is the loaf I want?” or “who won the game?”

I briefly tried LUIS but was disappointed to find it was just a language classifier, like a middle man. You can’t create responses for a chatbot. Unless I missed something in my brief visit.

Maybe you are expected to create a response on the web page using LUIS, as part of back end processing. But this is a good point, that some kind of default response could be a useful part of LUIS.

“Does anyone on this forum know enough about ... to defend it?”, Peter asked.

Commercial APIs generally have large tech companies to defend them.  For the sake of friendly conversation with you Peter, I will discuss my opinion and experience with building and training neural networks. But, I certainly don’t want to put anyone at a large tech company out of a job.  Especially a large tech company that has always supported me as an ultra advanced computer programmer.

“How much training and defining is needed to handle a sentence”, Peter asked in part.  In general, training sessions are not a guarantee, but neither are they for people.  A neural network is transparent with either a low score, or high score of certainty. This response may not make a neural network the best simulated conversationalist, but it may simulate a more convincing sense of being alive (in my opinion).  Not to put down of other A.I. approaches, but in its own unique way a neural network is rewarding to experience. For study or just because it is so cool, I recommend building a neural network of your own.

Reference: The sample output of LUIS at https://www.luis.ai/home may help identify valuable information for your critical questions.