Microsoft’s natural language understanding framework LUIS is essentially a semantic parser. Like Facebook’s Wit.ai it’s goal is to take the user’s input in the form of text and transform it to a machine readable form, in this case, an intent that describes what the user wants.
Unlike Wit.ai, Microsoft have taken the shrewd move of restricting which intents can be chosen for your bot. Before you can try out the API, you have to create a LUIS app for your bot and choose exactly which intents your bot will interpret. Intents can either be from the existing list of 150, or more likely, you will need to create your own. Here’s a random selection of things you can do with existing intents:
After you have selected your intents, you can train your LUIS app, and test it out with natural language queries.
I tried creating a LUIS app for my (imaginary) restaurant, The Eatalot. I wanted to add an intent for ordering food. There didn’t seem to be an appropriate one available so I created a new one, called item-cost, and added some examples of what a user might say, like “How much is the pizza?” At this point you can train the app and test it out. If I ask “How much are the burgers?” at this point, it won’t detect that I’m asking about burgers; in order for this to work you have to add “entities”. In this case, there is a set of entities for food that I can add. I can then identify “pizza” in my example sentence as being food. After retraining my model, I can ask it “How much for fries?” and it correctly identifies the intent and entity. Unfortunately, it is not very robust as asking “How much is a burger?” identifies the correct intent, but not the entity, presumably because “burger” has not been entered as a food entity. This is a pity; it’s easy to automatically identify all types of food using distributional similarity, and it’d be nice for an app to be able to say, “I’m sorry we don’t do burgers” without having to add every possible type of food as something that people may ask for.
What Microsoft has done actually makes a lot of sense. They know they can’t excel at detecting all intents, so only allow apps to detect intents they care about - they won’t be able to handle other ones anyway, so it doesn’t make sense to consider them. It’s an approach that requires designers to engineer their bots much more than relying on artificial intelligence. This is probably sensible, as any bot with a decent range of behaviour will necessarily require a lot of engineering, with or without artificial intelligence.
Unfortunately Microsoft’s offering falls short in several respects:
These concerns are not hard to address, so I hope Microsoft is taking note, and we’ll let you know if things change!
Pricing of LUIS is simple and sensible: at $1.50 per 1,000 queries, it is slightly cheaper than Dialogflow, which I would consider the main competitor. However you get a much lower quality service with fewer features.
In summary, unless you are somehow tied to Microsoft’s ecosystem, I would suggest going with Dialogflow instead of this for now, at least until Microsoft improve their offering.