Transferring from Conversation to Discovery

Let's talk for a moment about how to know when a question should be handled by 0:00 the Conversation Service versus when it should handled by Discovery. 0:04 It all boils down to how confident Watson is with its responses. 0:09 Take a look at this graphic. 0:14 Notice that the frequently asked questions, the common questions, 0:16 are those that take place within the short tail of the distribution. 0:22 This is where the Conversation Service excels. 0:28 As questions become more esoteric and uncommon, 0:32 then we rush to the aid of the Discovery Service. 0:36 >> Now that reminds of an example here at Treehouse, right? 0:41 We could think of students asking very targeted questions 0:44 needing help learning programming. 0:48 They may say, what's the key word for a four loop in Javascript? 0:49 And that would something that we could easily answer and anticipate. 0:55 But then we get some other questions that would be either more obscure and 0:59 less frequent, or more like you said, esoteric, like 1:04 what's some advice for how to get unstuck when I'm working on a problem. 1:08 And for that kind of thing, maybe we could, upload, for example, 1:13 blog data that we have where we have host-related that kind of thing, 1:17 whether it's the Treehouse Blog or elsewhere. 1:20 Does that make sense? 1:23 >> Absolutely, what a great example that was, yes. 1:24 >> Okay, so we're back at our page and we can test this out. 1:28 Let's try asking our chatbot to find me an apartment. 1:31 Okay, so notice that we see the response about our chatbot's capabilities. 1:37 Let's take a look at the intent back here in our Conversation tab. 1:42 So if you remember, these capabilities are what we imported from the sample. 1:45 And actually, if you take a look at what we imported, 1:50 you can see there's a bunch of extra statements in here that 1:52 don't really have anything to do with what our chatbot's capabilities are, so 1:56 this is a little bit messy, because it's a test, a sample instance. 1:59 So, at this point let's just delete this. 2:04 We don't really need this anyhow, although you could pretty easily recreate this with 2:06 a more simplified set of intents here. 2:09 So I'm just going to click on the trash can over here to delete the intent, 2:13 yes, delete it. 2:17 And we'd like to do the same thing here for the entities, 2:18 let's delete anything related to capabilities. 2:21 And finally, in our dialog node, we did create an extra node for capabilities. 2:26 I'm going to delete this as well. 2:31 All right, let's try this again, find me an apartment. 2:36 Okay, so this is interesting. 2:42 I actually didn't expect this to match and 2:43 flip over to the Discovery Service just yet. 2:46 So let's take a look at the output from Conversation Service to see what happened. 2:48 Okay, so here in the terminal console, 2:53 this is where we are logging output from the Conversation Service. 2:56 And here we can see that it matched on the turn on intent with a confidence of, 3:00 yeah, 0.29, it was a really low confidence. 3:06 So we said find me an apartment and our chatbot understood what we were saying, 3:08 and it was like maybe they're trying to turn something on. 3:13 I guess it's very low confidence, it's a very low score. 3:17 But that is below the threshold, so 3:20 sure enough it went over to the Discovery Service. 3:21 In the conversation flow it ended up in the anything else node, 3:25 which it would have said, I didn't understand, you can try rephrasing. 3:28 But because we passed it over to Discovery it did run the query 3:31 that said find me an apartment. 3:35 So we want to train our Conversation Service at least a little bit 3:37 on how to handle this new kind of request. 3:41 So back here let's add a new intent, and 3:44 what we're going to do is create a new one to find things for us. 3:48 So let's call it a find intent, and lets add a few examples. 3:53 We'll say find me something, look for something, I need, maybe. 3:58 Locate, and how about find a. 4:07 Just a couple different ways. 4:10 And looks like I have a typo so I'm going to come back here. 4:11 Gonna delete find me and retype it as find me. 4:15 Yeah, okay, so this can be a little bit confusing. 4:20 In fact, I had some trouble trying to understand exactly when to use intents 4:23 like this and when and how to flip over to the Discovery Service. 4:27 And this is one of the things that there's no, or 4:33 at least there's not always a clear right answer for, in developing a chatbot. 4:36 When we know exactly what the user is looking for, 4:41 we do wanna be programmatic about it, like this, on the Conversation side of things. 4:44 When we are less certain in how we want it to respond, that's when we need to flip it 4:48 over to Discovery and somehow tie into a broader data set. 4:53 So what we're doing here is creating this new intent, and 4:58 we're going to do an example to find a couple of things, 5:02 very basic finding, but then when we ask for lodging or an apartment, 5:05 the confidence will fall and it will flip over to Discovery. 5:10 Alright so to continue this example, let's add some entities here, 5:14 we'll create a new one. 5:17 And we'll call these locations, and lets add a couple things in here first. 5:19 Lets do a gas station. 5:24 So for the synonyms lets say gas, fuels, 5:26 fueling station, and then lets add another value. 5:30 Lets do a hotel. 5:35 And for synonyms let's say motel, resort, and 5:38 how about place to stay, and bed and breakfast. 5:43 So, let's complete this with a dialogue and we'll make this real simple. 5:51 I'm gonna add a new node below turn on. 5:54 This one, let's name this based on the intent, name this find, 5:58 and the bot recognizes our find intent and 6:03 it matches on one of our location entities, at locations, any of those. 6:09 Then we will respond with just something generic. 6:16 Let me see what I can find. 6:19 Okay, so if we open this, we can see that Watson is training on this new data. 6:25 And now it's done, and we can try this out. 6:31 Say find me a gas station. 6:33 And it matches on the find intent, knows the location was gas station, 6:37 and it says let me see what I can find okay. 6:42 All right so I'm gonna refresh to clear the page, and 6:45 let's try that down here the same thing, find me a gas station. 6:48 And we get the same response, let me see what I can find. 6:53 All right how about locate a hotel. 6:55 Okay, cool, so again, this is exactly what we just programmed. 6:59 But now, let's try it with something a little more vague. 7:03 Let's say, I'd like a place to stay near Central Park. 7:06 Well, interesting, it says it doesn't understand. 7:15 So, what exactly is happening? 7:17 If we scroll down, It didn't match on any intent at all. 7:18 So because I said I'd like, it doesn't know that we're trying to find something, 7:23 and it certainly doesn't match the turn on intent, 7:28 so it's kind of just lost without any kind of intent. 7:30 Interestingly, it did match our locations and hotel, 7:33 and that's because place to stay is a synonym for hotel. 7:36 So we could go and add this to our intent, or we could maybe try to control for 7:41 this a little bit differently. 7:46 But for now, I just want to try asking this in a different way. 7:47 Let's say, I'd like an AirBnb near Central Park. 7:51 Okay cool success. 7:59 All right check it out, we have in the responses, we have the title 8:01 of an AirBnb listing, and we have a review, this is the actual review text. 8:05 This is what we fed to our Discovery Service, and 8:09 there should be three of these. 8:12 So here's the second one, best deal, and if we scroll down there 8:13 should be one more, and yeah notice that these are all by Central Park. 8:16 In fact because these are individual reviews, we have two different reviews for 8:21 the same one. 8:25 So first, we need to look at the console. 8:26 So here we have the response from the Conversation Service and 8:29 it matched on the find intent, but as we wanted it to with a very low confidence. 8:32 So again, because we didn't have the language to match I'd like, and 8:37 because we didn't say hotel, or place to stay, or gas station. 8:41 It's trying to figure out what we want, but 8:45 it's just not very confident that it knows. 8:46 And in that case, it's like, alright, 8:48 well I'm gonna go over to the Discovery Service. 8:49 So then, over here, on the right side, 8:52 we can see the response from the Discovery Service. 8:54 And it has our input text, and it has a bit of information about the service and 8:56 the collection, and here is the output. 9:00 And we asked for three results which we got. 9:03 We have a listing name, we have some details about the listing, and 9:05 if we scroll through there are different things that are interesting to look about 9:09 for how to Discovery returns data. 9:13 There are different entities, you can see there's a person entity, 9:16 there are location entities, information about New York City. 9:19 Again, this was all pulled from the documents, 9:23 from the review documents that we fed to our collection. 9:26 And a lot of this extra information is because the Discovery Service 9:29 is hooked into other information sources. 9:33 So you can see there are even links to different pieces from the reviews that we 9:35 can tie into if we wanted to. 9:38 Let's try another example over here, let's say this time, 9:41 show me some good places to stay near Times Square. 9:44 Okay, and again, this was vague enough to flip over to Discovery, so 9:51 here we have a little bit of a different result. 9:55 So it tried to match on the turn on intent, because we said show me, it really 9:58 wasn't exactly sure what we were after and it was really low confidence in that. 10:02 So again, it flipped over to Discovery and, once again, 10:06 if we take a look on the right if we scroll down a bit, 10:11 there are different parts here that show the different entities. 10:13 And if you look, you can see things like it'll mention Time Square like this. 10:17 It matched it from the review. 10:22 So you should try this out with some different landmarks in New York City, 10:24 maybe some different ways of asking the question. 10:28 In fact, figuring out how to ask the question and how not to ask the question 10:31 was a little bit confusing for me and still trips me up from time to time. 10:35 And that's one of the things that has been hard for me to wrap my head around. 10:38 Because programming for this type of stuff is a little bit different than 10:42 more typical programming that I'm used to at least where we are giving 10:46 very specific questions and answers, very specific control flow. 10:50 So this kind of chatbot development requires a lot of testing and tweaking and 10:55 modifying, so if you think of different ways to say something 10:59 then you may want to program that into the Conversation Service. 11:03 It's also worth talking about how there isn't just one way to use 11:07 the Discovery Service. 11:10 You should follow your intuition and make it work for your chatbot or for your uses, 11:11 however you need to. 11:16 So in this case we are looking at the intent confidence, 11:18 and we would want to build up different intents and add more user examples. 11:21 But for now, we can see how we can get to a low confidence in a matched intent, 11:27 and we can query Discovery for help. 11:32 Now another way you might tackle this problem, 11:34 just if you have data that you're usign from Discovery. 11:35 You may want to set up an intent specifically 11:38 to relay the request to Discovery. 11:41 So you could imagine expanding our scenario, we could take this, find intent. 11:43 We could look at the location that was matched, and if it was a hotel or 11:47 a place to stay, if we made the synonyms match on that, we could look at that, 11:51 analyze it, and then we could In that case, pass over the attributes and 11:56 go into the Discovery Service. 12:00 This is actually how the food coach example is setup, 12:01 the food coach that we saw earlier. 12:04 It uses a dialog node to branch conversation into a Discovery Service. 12:06