um so today's session is is going to be about virtual agent using nlu how we can give uh additional advantage to our conversations by using an audio models and best practices yeah oh good now thanks awesome good so guys uh just quick reminder on the Safe Harbor notice um just in case if if I'm using any forward-looking statement that should not make your bike decision oh yeah so with this notice quickly move on to next slide uh just about me so I'm part of service now here uh teamfold product success Management Group uh I'm I'm a product success manager there based out of India Bangalore um just quick reminder in case if you want to join us in future sessions um then you can you can click the link which one of my colleague will will post here um so you can join us in future similar sessions uh very useful to learn more about servicenow product offerings a couple of reminders on the housekeeping so this session is going to record and fully available for um uh for the for uh you know for the feature uses uh in the community page uh after the session you will have a short survey please don't forget to fill the survey uh to it's a good way to share your feedback how we can improve further in future sessions uh quick poll so before going ahead I want to just quick check on the send demands um about uh you know um about the audience that we have about their familiarity in terms of energy so I'll just run one poll here and expecting your answer so let me launch a poll the poll one will have two questions the first question is more about your awareness on nlu how familiar you are the poll two is about if you're using conversations uh in VA then are you really using nlu now or or not at all using VA just wait for few seconds yeah that behind the pole seems like we have a majority audience somewhat familiar with nlu we're looking for tips and majority of them they are using uh Bas of now and with an Lu and the poll here see result good really good audience all right so with this let me move on to the next slide uh before we moving forward um I would like you to film there with couple of terms uh that we will use frequently in our session today and very useful to know about nlu terminology but what is nla in such uh natural language understanding um is is something uh something uh a great way to um to you know if you can use these analytic models with with the in conjunction of your conversations then it system will learn and respond to human expect intent so it's kind of a how human interact you can make your machine smart enough to interact in in somewhat similar way and if you use uh nlu with the combination of your VA conversations then nlu model can provide information to your VA conversations by determine determining what exactly user want to do what is the intent of the user you know uh behind the conversation uh that that we're doing so um how how nlu can do this by extracting couple of relevant information from the conversation that that VA is doing with the end user uh so that's about the nlu and how energy models will be helping the VA conversions to make them more smart couple of terms um that will use frequently in our session today um intent is something um that user want to do or user want to achieve right so for example if someone want to access of your Wi-Fi then there is intent to get access if someone want to report an issue that may be intent is to create an incident or create a ticket so so intent is something the meaning behind the conversation what is the reasoning behind it what exactly a user want to achieve utterances is um a different examples how a user can you know Express the intent so we might have a different ways to express express our intents so attendance this will basically it's kind of a way how ex how user is expressing the the need or intent entity is is something a very good concept uh where uh you know we can Define entities uh in the attenses uh just to give a context uh to those attences so that our model can learn what is uh what exactly the context of this conversation and also read the action behind it so that it can start serving better way to all those conversations we do have a different kind of entities uh let's go into that's not going to Deep dive but system entity is something predefined in the instance it comes as out of the box with the servicenow instance user defined entry something user can Define based on their conversations their Industries standard based on their user cases common entities or something again predefined something can be used across the organization for example quantity or in general currencies so those are pretty much common entities there is a concept called vocabulary which is again very useful sometimes when we want to Define um you know acronyms or synonyms then vocabulary is is a really good way to define all those synonyms so that system can understand what what is the common meaning behind all different words right so if you can see here example of Microsoft if you want to Define an acronym Ms then system can detect you know Ms means Microsoft data showing here so Walker load is another way to you know give a meaningful uh insights to your models and as I mentioned earlier energy model is nothing but collection of all those utterances entities uh is is something uh called remodel yeah so there is a story behind it one of our customers they they uh they were talking internally and they keep calling this term you know term call uh where is my zebra where is my zebra so in human in general conditions whenever we will hear this question probably what picture will come into our mind is is zebra like an animal right black and white color uh but you know what they were referring so they were referring actually uh zebra um scanner which is kind of a handheld device so they were looking for zebra scanner not not exactly the zebra so this is where uh you know um vocabulary or entities will come into picture uh and how we can translate this system so the system can also start differentiating um you know user want to talk about zebra scanner not about the zebra animal so this is what this is what nlu can do for our meaningful conversations to differentiate to you know give the learning to our model as a system to differentiate these two things uh at the same time so this is what the power of nlu with going forward if you're clear on these terminologies and we can move forward um so in in our uh you know traditional ways of uh this is how virtual agent work this is how our conversations work in service now if you are aware about the conversations then we have a different uh point to initiate the conversation then uh you know where the user can start the interaction with the virtual agent and the virtual agent is having a power to interact with the servicenow tables and all the different uh modules to you know get the information and pass on this information back to the end user to serve the better right so whether it's about the knowledge base or information about cases or outages a lot of internal Integrations with data tables so this is this is how the typically virtualization conversation Works to serving the user right if you go further down I would like to highlight the difference in the traditional ways of conversations versus the new new modern world in nlu so in traditional ways of virtual agent what we do um we simply have a kind of a flow in the designer we will Define all the steps uh you know how this conversation will initiate and how it will go from one point to another to the end right so uh in in a usual way uh VA without nlu uh it totally depends on topic Discovery based on the keyword search with based on the keyword mapping uh this is how we you know typically we thought about that VA conversation works so it's a step-by-step process in the meantime uh you know system will ask a couple of uh follow-up question to user hey uh do you mean laptop what do you mean by Hardware um do you want to express your uh your priority with this issue so all those conversations will go in between of this the moment we have analu comes to picture so it will give some mind to our models or machine to start thinking of a way beyond our traditional VA conversation works so here we have a context uh contest context of intent and entities so using nlu you know models they will start predicting the intent behind any conversation so if you take a simple example of update my email to John at sn.com so what is the intent granted if we really trained our models they can start predicting those those intents so intent here in this conversion is basically user want to update the profile right that is what the ultimate intent is and what is the item that user want to update basically the email and there is this new email ID this is what the value is so everything is available in single conversation and system will smart enough to extract all these information and then you know start providing the meaningful responses or actions further if you compare with the traditional models uh VA conversation so you know sometimes during the flow system will start asking where is your new email ID right so we can by using nlu we can bypass all those um you know sort of additional questions we can make conversations more smooth and short very quick relevant to the user so this is what this is what the power of nlu using with your VA conversation make your conversation smart with this uh I would like to have a quick check on the implementation cycle um so in case of nlu we usually follow this cycle we start with thinking about our model requirement what exactly we want to achieve by you know making any nlu model uh you have to keep in mind all the behavior conversation which you want to attach with this model and then you start working on the model building so it's nothing but creating your relevant intents and assigning the right utterances to them uh defining the entities and all and then you know test your model um see how it is performing is it like as per the expectation to new model if it is not performing as per expectation then finally publish it uh you know to your production or sub production stands for your end users once we have a model deployment done then it's good to ready to associate with your any VA conversations or AI search right in the meantime you keep eyes on your performance of your model if you see that performance is degrading or you see there are new ideas to improve the performance go and start tuning your model again right so this is the typically implementation Cycle Works before I going forward I will take a pause uh I do have a next short demo plan just to let you know how this uh you know analy workbench looks like uh is there any question it would be you can ask your question uh in the chat window or you can post your question in the Q a section will help you to okay good I'll take a pause before going demo in case if you have a question there there is one question with the nlu having access to the servicenow database will it be able to use AI to determine the intent and utterance um yes so there is a separate section um for AI uh I think servicenow is really aggressively working on the feature AI options to be available along with your conversations and how they can blend AI with nlu but yeah um it is part of the service yeah you know not not really sure I understand your question um so with I know you're having access to service now the other place will that not be able to use I I to determine them okay let's see so you're saying using AI to determine the internet and nutrients yeah okay there's another question how does energy fit in with the no assistant send it to Via controller capabilities um I think we do have um follow-up sessions basically on the Gen AI section I know it's one of the one of the you know interested interesting topic but as of now this session is highly focused on and edu capabilities that we already have available on the production instances and customer can start using is servicenow is aggressively working on the Gen part as well so in future we will have a focus session on gen AIS there you can learn how how we can really use the power of gen AI along with nlu and conversations yeah so let's keep keep the session limited to nlu and conversations uh yes Japanese we do have mandarin I need to check I think we do have some 17 languages I've seen supported awesome good okay so keep asking questions there's one more question in the Q a thank you any okay uh best practices yes we'll just be the best practices for sure in this session so let's hold on this question for now good in between those keep asking questions in the chat window uh few of my colleague they can answer you with chat itself I will move it forward and just show you a small conversation so if you go to your VA um just you know I want to just show you how this entire flow works ask a simple question what is Nadu yes I want to know about nlu so it is started giving me different information on the nlu by just simply asking a question of what is an Lu okay but the reason is because there is a model behind it which is really extinct in the intent so intent is nothing what I want to know about the nlu keyword if I want to know more yes [Music] so this is one of the simple conversation where I just use the the intent to know about nlu as such and see how VA is started responding with the relevant content so how this entire conversation works if you recall your VA experience then usually we go to uh to use the VA we usually go to the conversation interfaces where we have all the VA related uh options if I go to settings here uh you will see the tab called virtual agent so I'm not going to Deep dive in such a VA because this this session is all about using nlu with VA right so we'll skip couple of va topics we'll focus our conversations on the nlu so if you go to a virtual agent site you will see this section and this is how you can switch on your nlu on your instance usually it comes automatic switched on so you don't need to really click separately and this is the this is the one area where you can have all your settings uh in the under VA about nlu right so if I go to view settings quickly foreign so if you see by default we have a service provider service now nlu in case if you have more energy service providers like uh like Microsoft Luis or IBM Watson Google dialog flow so service now is capable enough to consume all those different service providers integration so if you have integration with other service providers you will see the option in this drop down list you can select the provider you want to use as of now I'm keeping one default service better as servicenow which comes inbuilt with the instance once we have this enable then probably another concept which I want to show you is what nlu workbench so how you can go to an early workbench by clicking here here right before I show you to the analy workbench I want to show you the conversation which we just completed right the nlu conversation so this is what the flow I was using very simple flow you know start the conversation information on the nlu um you know system will start predicting the intent so intent was to know about nlu mode and Supply the information uh and ask a question you know if it is really relevant you don't know more so this is pretty much very simple conversation how this is getting power from nlu if you see the properties behind this composition you will see this section called nlu and this is where the power comes see what I did here I just created a small model uh in the service now instance and I have attached this model here with this conversation right so this is what the model name is and this is what the intent that I'm uh that my user want to know and this is what the predictable intent behind the conversation so this is how you can integrate your models with three conversations if you integrate this way then you will have this additional tab here called NNU intent all right so here we have power this is this this is the addition section if you integrate your VA conversion with nlu then you will see all these possible addresses we will learn more about these utterances in in few seconds yeah so this is what the VA conversation looked like I was talking about the nlu workbench earlier right so uh this is the way so you can click here and go to nle workbench uh I will click have already open the workbench so this is how the workbench looked like uh One Stop Shop for all your nlu model activities workbench will give you option to create models to tune your models to check the performance your models to Define entities walkabilities or everything at one place right simply you have to navigate to workbench you can navigate to workplace by clicking on this setting or also you can search nlu here you will have models right so this is what the model looked like here you have option you can use our pre-built model which comes as out of the box part of your instance you can import your data from CSV and create your model from here otherwise you can create a model from blank data right so if I quickly show you I have to just give name of the model here say ing ticket all right I will adjust to the description what what exactly this model is going to do about I can select a language so earlier there was a question on the languages so I think we do have uh so many languages support it as part of nlu so you can so idea is you have to if you are using languages then you have to create a model in that language itself yeah so if you have a say three languages English Mandarin or Japanese you need to create your models in three different languages so here I'm going to take example of English model what is your purpose behind it so if I want to use this model with virtual region then I can select virtual agent or with AI search I can select the AI search as of now let's let me take a virtual agent option you can Define the business area if you want as of now I'm gonna select ID the reason is so that you know if you have said so many models then you can easily differentiate which model belongs to which business area click next so I think system is creating model for us so it will once it will create and it will notify us so this model is going to be a blank model because we have not created this model by copying some models or importing the data so it is going to be the the blank model yeah very cool oops okay so error maybe some test model to this and see next yeah this time I created this is what our blank model is asking me there is no content yeah I know because I have created black model so this is what the model is created and you will see the life cycle of this model right so this is what the first life cycle event you can manage your model or content here you can test your model you can publish your model later on right and this is where the the power of model comes you can Define your intents here what exactly you want this model to achieve what are the possible intents you can Define the entities you can define a walkability rights all those options if I want to add an intent so I want this model to predict uh you know open ID ticket so for example I want to predict if you use based on the user conversion conversation um whether user want to create a ticket not set the description so this is what the intent I have created now I have option to add multiple utterances right so attempts is you remember utterance is nothing but the various different ways user can show the intent right so one of the attends can be um a laptop issue right so you can add the currencies here um you remember so we created this model blank so I can I can add that and says uh manually uh you can also import the entrances or you can copy the addresses yeah so this is what the example of our Trends here and if we see this uh underline so what it is that system is detecting uh what do you mean by laptop so laptop may have different meanings right what what is your meaning for this particular intent so basically here you can Define the my meaning so basically laptop here I am referring uh Hardware computer type of hardware right so I can Define this so all these uh you know definitions will help model to predict the right uh intent similarly um there is a concept called entity as I as I remember earlier so entity is nothing but giving um you know context to conversation uh setting up the action so we can Define the entities here how we can Define we can create new entities these operation different type of entities that you can choose as of now let's say I want to create a new entity here say our chancy so I medium low um this medium and low so what basically I'm doing here it I want to Define an entity again this word issue and whenever this issue entity discovered in the conversation system is going to assign a priority for this entity so there is a action behind it yeah so I want anytime if user type a Word laptop issue or any issue it is going to assign a high priority for that entity right so we can skip a conversation to asking this question hey what is the urgency of this particular issue that you are talking about system will detect if there is a issue keyword mentioned it will assign a high priority for that one right or another example may be laptop or IR right if someone's laptop is on fire then you know it doesn't make sense to ask this question what is urgency of course it's it's a higher tendency so you can Define the entities you can give a meaningful uh conversation in meaningful inputs to your conversation yeah this is about entity and we learn how to create models uh there are different ways so this is our model creation you can test your model we'll we'll talk more about testing options um I don't have a slide over there so any questions till now creation of the model assigning virtual agent let me check yeah there's a couple of questions so uh one I just answered follow us do we recommend one model or multiple models with grouped intents so I did post the KB article there which you can refer to that basically we recommend you try not to have too many models because there is the the possibility of clashes between models okay so we do have a tool that you for you to help troubleshoot some of that if you do get clashes yes but uh we do recommend you you stick to you know fewer models for better awesome good keep asking questions with this I will move to my slides back there was also one other question in the Q a uh I think it was around licensing so you know what license you got for nlu so you basically need to need a pro license like idsm Pro or CSM Pro like that so if you have virtual agent not the virtual agent Knight then you should have uh llu questions good thanks all right good so I hope you can still see my desktop so since till now we covered how to create a model how to interact with your workbench how to you know assign your models to your virtual agent conversations I would like to talk about the various different ways how you can test your models once you created your model these are different options you can test of course you can you can follow your own method of testing but servicenow gives you a couple of inbuilt option that's something you can utilize and start testing the performance of your model right so if you can see my desktop the first option that we have here you can compare the draft model versus the last published version so if you have published your model already you have using you're using this energy model in your conversation and you recently to your model again to better perform then before publishing that new changes or new drafts you can test how it is performing as compared to the the the live model that we have or production model that we have right I'll tell you I'll show you how to do this uh the another option that we have is an Ado expert feedback loop so this is nothing but um it gives you option to extract the data from your virtual agent chat the production chats and it will start giving you option you know these are the potential chats that we have and you can you know one of your admin or someone can go through all those suggested uh chat chat logs and start using those chat options you know to assign those as utterances in your existing model so this is something about you know you can share your feedback as an admin user or on the real conversations happen on your production instance and see if those making sense to add your model that you're using right so to to improve your conversations in future the multi-model best testing give you option to test uh multiple models uh against a large set of data so if you recall the machine learning right so using machine Learning Works in such a way that uh 80 of your uh data where you want your machine models to learn and then you keep 20 percent of set of data for performance testing of that model right so same concept here you can keep your testing data separately you can upload this testing data on the instance and you can start comparing your model performance against those testing data points right so this concept we'll learn more about in a few seconds cross model conflict is is really a good option where you can see what are the utterances they are conflicting so for example if uh one of the one of the intent is create a ID ticket is having a tons of laptop issue you might have a same address into some other intent as well so for example uh you know speed slow um my laptop is slow it's kind of for intent so there are sometimes by mistake or while creating utenses you know people they keep same utensils into different different models or different intents then you know your models will start confusing what to break but which is the right intent to predict against those common utenses so this is where the way you can use and you can see all those conflict attences again you can see how it is basically how your model is performing against the actual VA conversations so um if you recall your VA whenever any user initiate the chat uh you can enable the option where user can give feedback whether virtual agent give us a relevant information or start asking the right question so if this is something you can you can get the data from your actual conversations and then compare your models again those things right intent Discovery is something in in a realistic world if you want to know what are the potential intents if I can include those intent into a model our users they will have a more benefit so system will discover intents for you based on your real conversations on your production virtual agent instance right so we'll we'll just quickly see these options before going to the next uh so how you can utilize these inbuilt options that we have so go to nlu and you will see these options right you will see the option to check the performance to see the intent Discovery uh multi-modal batch testing or even cross model conflicts if I take example of say performance so you can see how my model is performing um for the virtual agent right so if you can see here we have a virtual agent conversations where 55 percentage something user confirmed as virtual agent rightly predicted or physically you know that VA give us give give a right question or write content to the user and this is something based on the user confirmation how you can enable this option on the user confirmation go to your base settings so you can enable this option ask user if topic VA choose is correct so user will initiate the conversation then there will be an option to you know give us a feedback so user can give feedback to those conversations and this is where system will use to compare you know the the model performance with the user feedback so you can see here um eight percent something at the instances where user you know not shared this information or feedback 36 percent of our conversations where users they said uh you know VA predict a wrong uh intent so this is one way you can check the model performance how your model is performing uh against the actual feedback from the user right so you can use this uh this option within service now instance I will not go into deep dive but uh but yeah you can you can really see filter out your data based on uh further you can see what are the unsupported utenses you can see which are the utterances performed well and which are the utterances really where your VA not predicted the right intent right so this is one way the another way is okay uh intent Discovery multi-model best testing if I click this this is the option you can so you remember the machine learning concept keeping your 20 data for your testing or model so you can upload those 20 data set you can create the data sets here so I have created two data sets and just to check the performance of my models against these utterances right whether model is predicting a right intent against these data points so you can upload your data set you can uh you can run the analysis and if I show you the data here so you can see against this sample data 46 percent of my model it is predicted correct way um 33 is is something we are model confused uh multiple predictions happen uh zero percent missed and twenty percent uh time that you know my model predicted the incorrect intent so you can verify the data points and you can you can improve it further right so this is another 8-bit option if I go further down then we check the multi-model best testing now we can check the cross model conflict so you can see here I just run one analysis yesterday this is what my model is and you can see there are two critical conflict two moderate conflict so what this conflict basically are I have in my model I have two intents open ID ticket and hardware issues and what we saw here we have a common utterances so basically uh this is the way we can identify if we have a multiple common utterances you know the reason is if you have a common entrances then model will start confusing you know what to predict so you can utilize this option to check how your if you have any conflicting utenses in your models and you can improve them further right you can modify those utterances to make them more relevant about your intents so laptop is very generic it is predicting laptop because laptop defined as an entity under the hardware issues and since it is related to our laptop issue so it is indicating that user need to open IG ticket right so you can you can utilize this conflict review option so these are a couple of inbuilt options we do here intent Discovery as well where I tell you right so we can we can discover the intent system will suggest you that you know what are the best intents you can include in your models uh based on the real conversations uh on your production instance yeah so with this I'll just take a pause in case if you have questions yeah there was uh one question from John um so how does nlu assign an urgency to the utterance if you map to an entity so basically the you know the entity is just a variable um and then it you know it'll capture the value for you to use within your conversation so it's not it's not there's no logic to The Entity and then we just have another question that's come in can I ask how we collect this data and something where testing group will be users um so srita are you talking about um the utterances yes yeah so we have multiple ways uh either you can keep your uh test data um within you you can collect this test data from your own experiences otherwise as I mentioned right you will have a different option you can you can take a feedback from the users from your actual uh conversations you can correct this data and start using that so it is up to you if you have some data to test go ahead and see how your model is performing uh on on my own data right if it's creating right way or not otherwise you can trust on the service now instance service now insteads can collect data on your behalf from your real real virtual agent conversations on production yeah I hope this answer question okay so keep asking questions in Midway um so there was a question on the entity right so these are the entities defined so you can Define entities here um as I mentioned we have different type of entities we can see when it is urgency two uh there are system in building entities as well so it's kind of a reference so what system will do system will make a reference of those entities and whenever your VA conversation they will have a same references then nlu will you know give the will predict the right intent behind those uh variables so someone is type my laptop is on fire so fire is an entity already defined as a high urgency so you know your conversation can skip the question about urgency so to not ask a question hey what is the urgency of this issue of course system already know if laptop is on file it's on high urgency so it will directly start to create a ticket instead of asking a question from user so you can skip multiple uh you know questions uh that in our normal VA conversion we used to ask from user you can avoid all those unnecessary things and make your system smart enough to to you know make a decision okay good let me go back so we discuss about the various testing and options uh and of course you can publish your model uh by looking here if I go to my model so you can create a model you can add the assurances Define your entities here you can test your model performance if I just quickly show you right so this is the way to test and you can once we have satisfied with the performance you can publish your products so you can publish this model on on your instance yeah uh I want to cover a few additional things I want to cover the best practices that someone asked question about best practices so let me go back um so we we discuss about the life cycle of the implementation right this is about the life cycle of tuning your model uh you can plan the tuning work you can build your model test the model here and then if you see any performance degradation or if you see we have you know new entrances to increase the model prediction then you can go to the tuning cycle how you can theory model again you can use the different options that we already have for testing the performance of your model um and then then I'll tell you what are the best practices you can use to make tune your model and start getting the best responses right and once you tune your model deploy the model monitor the performance again and if it is performing well you can expand this model to you know your VA conversation or yeah I search and you can improve this conversation in future as well right so this is typically the lifecycle works I would like to touch base on couple of best practices um I'm not going to read everything but definitely we do have couple of out of the box models available which comes with your instance uh you know instead of creating your new models you can always try to use those models that we already have I'm not sure if you notice but if I go to home these are the out of the box models you will get along with your instance right about HR related itsm and you see these are the models in different languages so on a day one you don't need to create your fash model better to utilize those models of course you cannot change anything in this model because these are read only but what you can do you can duplicate these models and and you know make the relevant changes for yourself so you will increase to use the out of the box models always you know try to see the knowledge gaps in search content always try to Define intents and so that you know model can start predicting the right intent and with the help of utterances try to choose any other entities over separate intent you can use the vocabulary option that we have so not sure if we really cover the vocabulary but you can Define uh vocabulary for your own industry for your own company for on use cases um so if I go to this model you can see the vocabulary here so walkability is nothing but you know I'm giving uh your own industry meaning to a particular word for example I was talking about zebra right so zebra can be used in multiple sense so for example if any company they want to use zebra in terms of the printer or um uh you know those things then they can Define the various you know workable items Hey whenever we see Zebra printer then refer prefer the zebra that we are talking about zebra technology right so you can Define the workload you can Define your uh words to synonyms as well so people they call say a lot of different way yeah just uh just checking can everyone's still here looks like maybe maybe you might have lost audio is it just me I can certainly still hear you and other people still hear us well we have lost okay it might just be you Mimi can you can you hear me guys now yeah yeah I think we're still good keep going okay good awesome so yeah we're talking about the vocabulary so you can use the vocabulary to Define uh to give a different uh synonyms for to your words based on your industry based on your use cases so those are the couple of best practices uh quick on what the what are the different things you can do to perform your model in a better way uh you can avoid you know unknown words in your attenses so words like uh you know I am so I I am is not a correct word it should be I space am to avoid those unknown words uh you know acronyms like pin ASAP you can avoid these non-verse in your attenses if you want to use them then always you know use the entity and relevant vocabulary for synonyms yeah avoid using non-english word if you are creating an English model it will make your model confused and not perform well yeah avoid the model ambiguities so don't use the overlapping intents for example Outlook login or Outlook meetings it can make your model confused um don't use the low context utterances for example one word right passcode or single word reset it will not help your model to predict and the right intent don't combine the tenses so here you can see this example I need a new laptop and a printer so try to make two different addresses entrances instead of keeping a one one statement so keep it something like I need a new laptop I need a printer so these are recommended intent structure um you should have enough intents before publishing your model we recommend to keep at least five interns in your model and don't create too many intents right so it doesn't make sense to use so many intents though we are keeping an upper limit of 300 but yeah try to make it relevant intense you can also keep a eye on the number of attences we recommend to have at least you know 15 entrances in your intent make sure that your you have enough coverage when you are creating your utterances try to cover the attenses or different ways the user can ask same intent or so the intent our duplicates of course uh yeah so these are the couple of things that you keep in mind while while creating your uh models or two new models to perform at a best level with this I would like to go ahead a quick poll and of course keep asking question uh let me see if I have a question there are three open questions let's go should the format of address be in sentence question phrase format which format might have a better chance to capture the users and utterances yes sentences yes questions yes uh phrases keep those phrases short um you know it's not that three lines four lines of phrases so keep them short so that your model can learn and you know give us a best result um but yeah these are these are the good uh ways to Define your utterances and I mean I think it's it's in the it's in the it's in the title right natural language understanding so how are people asking for these how people ask that's how it should be that's how it should be phrased just make sure it's uh you know relatively short and uh I'm actually you don't have you know two phrases that are too similar yeah it should be natural it should be you should do if you maybe interview your your end users and ask them or look at the data and see how they're actually asking for things yeah okay um there's a question on the entities yes so entities are Global when you're defining your entities you can Define them as a global and then you can refer all these entities in in across all the models that you create the model specific but it is a model specifically you define it in the model and then they can be used within all that the the the intents within that models within a model yeah yeah so with the new model you create 10 entities uh then you know the then then you can use these entities across all the intents there's another question are there any capabilities of copy models from one language to another uh no no you have to correct you have you need to make a model for each language and system will not help you on the translation part of models but uh yeah I mean we do the another capability we didn't talk about today is dynamic translation so if you don't want to maintain multiple models one thing you could look at using is dynamic translation but for that you would need to connect to a translation service so with that you only need one model and a user enters whatever they enter that goes to translation service is interposite to the model language that you have uh or your system language and then it'll respond and then that gets translated back to the user's language so that's all transparent and you can use that if you don't want to maintain multiple language problems yeah so with this I just quickly launch a poll I would like to know um since you know a little bit more about the nlu capabilities how many of you are going to uh try it in your new business conversations just take seconds a couple of questions so we're going can an intent be used in more than one virtual Asian topic no no so that's where you would get a Clash um so you you wanna you want separate intents uh well you're a virtual agent topic should be dealing with one intent yeah that makes sense yeah so if you can see my desktop then uh only option I have to Define one intent I cannot Define I cannot assign multiple intents uh in my conversation cinemata I got the result uh I think yeah most of the folks they said you want to try the radio in their VA conversations okay uh let me go to questions more questions here if no more question then probably we can conclude this session here um in case if you want to join us in future sessions uh try to register yourself with the upcoming feature webinars and with this thank you so much thanks everyone for joining us today have a nice day thank you everyone