Conversational Understanding Model with MS Azure

Okay, so a little bit of introduction about myself

Yeah, Kuljod Desai. I am a first year undergrad at IIT Bombay at the Department of Environmental Science and Engineering

Okay, so what is the tutorial going to be about? The tutorial is going to be about creating a conversational language understanding model with Microsoft Azure platform

Okay, so a bit of introduction about what we are going to do in this tutorial

So to start with, a conversational language understanding model uses concepts such as

natural language processing and deep learning to make the bot behave like a human

So through this tutorial, what we are going to do is we are going to train the computer

or the model to behave just like a human and then when we give it a query it's

gonna recognize the intention behind it and then act accordingly for instance if

I give a command to print the time then it's gonna identify the intention behind

it and then act accordingly to print the current time okay so this slide is not

important but to keep the tutorial on a comical side or just to have a small

comical aspect of the tutorial I have kept this light just for the sake of

file so okay moving on to the next slide some important terminologies so before starting with

the model we are going we will need to learn some some term some important keywords and so the first

keyword is intent so intent represents a task that the user wants the computer to perform so

So intent can be as simple as telling the computer to print the time or, or, or as simple

as telling the computer to print what they, it is, for instance, what time is it

Tell me the time. What's the time now? So all these three examples are nothing but different, different utterances of the same

intent that is telling the computer to print the time. So the second word, the second important keyword is entities

Entities as the name suggests, refer to some object. For instance, I tell the computer to tell me what time is it in Pune

So the intent in this specific sentence would be telling the computer to print the time

That is the intent of my query would be that I want to know what time is it in Pune

and Pune the word that is highlighted or that is written in bold letters would be the entity here

and the type of this entity would be location similarly the second example is what day is it

in London so in this sentence the intent or the intention behind me inputting this query would be

that I need to know what day is it in London and the word London that is written in bold letters

would be the entity here and the type of this entity would be that of a location

okay so now moving on to different types of entities so the first entity the first type

entity is something known as list entities list entities have a very

limited usage and the model identifies these entities with only a specific set

of values so the crux of this whole tool Linus is that suppose I want the

computer to identify Monday as an entity of the type we did okay so if and only if in my query I

enter the word m o n or Monday as a whole then and only then when the computer identified that

word as an entity of the type and similarly for Tuesday we have used TUE or TUEs that means if in

my query I write the word TQSJ as a whole or use TUE or use TUEs then and only then will the computer

identify that word as an entity of the type we take okay so now the second type of entity is

learned entities now these entities are one of the most important types of entities and will be the

most commonly used entities so these entities work on a very simple principle the more you provide

the utterances of phrases for these entities the better the model gets trained to identify these

entities. That means to make your model more accurate in identifying these entities, you have

to feed the model with huge chunks of data. And the more you feed the model with data, the more

precise the model gets. Okay. So moving on to the last type of entity, these are something, these are

called as pre-built entities. So, as the name suggests, you can compare or relate this concept

with that of keywords in a language which are already defined in DSDK

For example, the number 6 will be identified as an integer or a numeric entity

and the word Microsoft will be identified as an organization entity. So these pre-built entities are already offered or are already offered by Microsoft as your platform. So you need not worry about these entities. You need not worry about, you know, already defining with defining these entities as learned entities and then going through the whole process of feeding data into the model and then training that model on that amount of

on that huge amount of data to identify these entities. I know this these terminologies are a bit difficult to grasp at once but believe me as we move forward with our tutorial as we move forward with our practical once you do the practical hands once you do the

hands-on lab you will get quite familiar or quite comfortable with these concepts so have patience

and please please do not ignore the lab or the practical that we will that they

will perform later in this video okay so that's just a slide to which

differentiates utterances based on their intent and entities so pause the video

for maybe five or ten minutes have a good look at it and dive deep into this slide have a good look

at it and I guess we'll be good to go for the hands-on tutorial okay so I guess the theory part

of this tutorial is completed now moving now we'll be moving on to the hands-on lab

so uh okay so the first step would be to open portal.azio.com

wait for some time yeah so the first step here would be to create a language resource

that we will use be using to create a model to create a conversation

understanding model yeah here you go click on create continue to create a resource so this is the subscription here which will

obviously differ on your counterparts resource group I have already created a

resource group or if you want to create a new one just click on the new one

create new resource loop and then move forward so I'll be choosing the

already created resource group that that that that RG region I guess each US

would be fine name I will keep it as new language resource pricing tier this

click on review plus create validation required okay something okay yeah click

on this checkbox review plus create So click on Kuljot, new language resource

Okay. So our language service has been created. So I guess 20 or 30% of the tutorial has been completed

So yeah, just make a note that go to resource management tab and under the resource management

it click on keys and endpoint and it's always a good habit to just copy the key

and the endpoint URL and save it in a notepad file or some text editor file

yeah so I am also going to save location for being the endpoint URL yeah

so now so now we are going to begin the major part of the tutorial so after this you have to go on

language.cognitive.azure.com which is the language studio or the place where we are

going to create, train and deploy our language, conversational language understanding model

Yeah. So here we are. Click on sign in. So now what we need to do is we need to connect our language to the language resource that

we have already created. So click on your Azure subscription resource type, sorry, language and the resource that

we just created here. the content now what we need to do is create a project click on create new and then

conversational language understanding our project name would be let's keep it dashboard looks like of the language studio so what we need to do next is

create some intents and entities and then move on to deploying a model

training a model and then deploying the model so to add an intent click on add

intent name would be get time click on it and we'll be adding some attrances here so forget time some three to four attrances

would be something like what time so what time is it

what the time they need the time and what is the time click on save changes

so let me have a look at my code what all intents do I have to create get day

get date okay click on create a new intent get date

what day is it what's the day then the page what is the day

click on save changes now the next intent would be get time also oh yeah

get date sorry click on add so I'm sick click on gate date select the date so

What's the date? What date is here? There will be the date

Click on Save Changes. So now we'll be creating some entities

So the first entity would be location type. Click on learned entity

Click on add entity. So now to train the model on these entities, follow these steps

Get date. Click on get date. and let's say what is the Paris and now click on Paris and if sorry click on

that is and the whole and define it as a location entity some another example

would be 10 May the day in London and another would be what's the day in Leo

save changes now we are gonna add another entity which we are gonna name as weekday

and this is gonna be a list type entity and new list Monday

on pre-built add entity and now click on add new pre-built click on select pre-built and from this

click on date time so to just have a look we have a choice entity date time entity email

general event general organization location ip address person name phone number and and stuff

like that so you have a lot of pre-built entities already defined in ms as your language studio

offering so you need not worry about uh you know defining first for example defining percentage

as a learned entity and then feeding data to train the model on identifying

percentage as an entity click on save so click on intent okay so big derived we have to define big day

what the day of 23 9 and identify this as a date identity okay I guess we are

done for that's pretty much it for training our data for training our model on this specific

dataset click on save changes click on training jobs now we are going to train our model so what

we have done till now is provided the data and now that now all that is left is training a model

Deploying a model and testing are deployed model. So to train a model click on training jobs

Start a training job into model name would be a new model

Standard yeah train And this might take some while so have patience

Okay, so our model has been trained now we have to deploy a model. So to deploy deploy a model click on deploying model and deployment

The name would be it's a new deploy men New model deploy

okay so hard deployment has been completed now let's test a more selected upon that new

deployment let's say what is the time he's run the test so so the intent then the model has

identified is get time with a confidence percentage of 99.99 and no entities

predicted as such and this is the graphical interface to view the JSON

interface click on JSON okay so this output is an adjacent format with

confidence scores of all categories so for example for the category get time

or for the intent get time the confidence score was 99.99 percent for the intent get date the

confidence score was 88.78 percent for the intent get day the confidence score was 87.56 percent

and for the intent none the confidence score was zero and the entities predicted here are none no

entity was predicted so let's run one more example

run the test okay so the intent that has been rated is get time with a

confidence score of 82.98% location is India with the confidence of 100% and

with all these examples I think our model is working just fine okay now so

coming to the final part of the tutorial in this part we'll be using the command

prompt to call our model and input the query and let's see and and and then see

what result we get so the first step would be to open the command prompt

Okay, so now in this deploying a model tab select your current deployment and then click

on get prediction URL then this is a post command a curl post comma copy this

then copy it in a notepad and now we I think we will have to make some changes in this yeah

so yeah there is the participant ID here this one takes your query here let's see up let's put it

what is the time in London that's or I think we should select a new day in India

window paste and then hit enter yeah so this is the final output that we have

got from the model so query was what is the time in India top intent is get time

confidence score of get time intent was 84.47 percent of get day was 75.3 percent of get date

was 64.96 percent and of none was zero so that means the intent detected is correct which was

which should have been get time and it is get time here with the most confidence score of 84.47

and the entities detected is India which is of the location type of which is of the

type location length is 5 confidence score is 1 so I guess that's pretty

much it about this tutorial and I hope you enjoyed the tutorial

yeah thank you