Building Compassionate Conversational Systems via User Modeling

[Music] uh Welcome to The Accelerated Discovery Forum distinguished speaker series I'm today I'm happy to introduce

raraju messed upu it right yeah that's good she's an IBM distinguished engineer and master inventor uh Roa is currently leading multiple projects in IBM's Watson division uh on conversational

systems that can provide natural personalized and even compassionate uh interactions with users um this involves inferring uh users emotions tone personality and much more uh through

text analytics and machine learning um I've had the pleasure of working with Raa and her team on some of these projects a bit uh and so I'm sure uh you'll you will um enjoyed the talk today uh the title is compass building

compassionate and personalized conversational systems on thank you Bob I'm eager to find how you to find passion I'll find out sure yeah yeah that's what we here today

for um well thank you once again Bob for the introduction Michelle and moo for organizing this and um I have to say you all work in a great lab and it's

difficult to find this lab that's what most visitors say right when they start the talk that's supposed to be a joke by the way for who didn't know I work in this building um so it's good to be among

friendly audience uh all right so we'll get started I'll tell you about the work that we are doing in Watson about um building compassionate natural and more

personalized conversations uh and also the a point of view that uh we currently hold a big shout out to all our team members several of uh uh whom are here

in the room and uh some of them maybe on the lab so thank you everybody for all the collaborators and we also have collaborators from hia research who we work closely with On Emotion modeling

and uh there also have been over the years many people who worked on several of these projects who are either no longer with IBM or have moved on to other project so I want to also take a

moment to acknowledge them and uh thank them for their contributions so just so we level set what do we mean by conversational systems by conversational systems here we are talking about you know the things

that you encounter with when you when you're doing chat with chat Bots or intelligent personal assistants like Siri um Cortana or um you know Amazon's

uh Alexa or even Google's Google Assistant now which was formerly Google Now um any kind of digital virtual agents or even robots and this could be assistants in car this could be you know

health and wellness heartline um systems and I'll talk a little bit more about those or could even be exercise trackers and uh motivators so these are all examples of conversational systems or

they're also alternatively called as um dialogue systems but I'll tell you that this talk is not about conversational systems per se or the architecture of those

conversational systems you know that would be a very interesting talk it would be about you know who has control in the conversation how do you detect intent do you use single initiatives or

mixed initiatives is it a slot framework kind of architecture all of those would be great things to talk about but today I'm not talking about any of those what we are talking about today is about user

modeling in conversation so it's about really how do we incorporate the perspective of the user with whom the conversation system is having a conversation with and the the context

around the user who the person actually is and the the personality the context the the intent and uh the emotions of the person and such how do we bring all

of those aspects of users into the conversation to make an ongoing convers ation into a more empathetic compassionate you know natural and more

personalized so that's what we'll be focusing on and we'll also assume that there is a base conversational platform that we can work with and uh so we'll assume that we'll be building on top of

that such a conversational platform to incorporate these user models so consider this conversation so you're ask Jan is a person who is uh let's say a hotel guest

and she's asking what's a good place to eat in San Jose California and Watson is the technology or the CH agent let's say conversational agent it says what kind of Cuisine do you prefer Jo either

Italian or Mexican or Indian or Chinese any of those would be fine Watson says are you a vegetarian Joan says yes let's say Watson says do you like fine dining

experience or fast food let's say Jan says Fine Dining Watson says are you driving or taking taxi or public transportation Joan says I'm driving Watson says do you prefer a restaurant

that has a parking space of its own or are you okay to find street parking on your own all right I'm exaggerating here but uh uh Jan says yeah parking space would be good and then Watson comes with

recommendations let's say it says whatever there is this three restaurants one is an Indian one is a Mexican one is a Chinese you might consider these all of these have ample parking okay a typical conversation agent that we build

for making restaurant recommendations might look like this today remember that parking is a Criterion for for restaurant oh in big cities it it could very well be right in

New York or in other places yeah but as I said I exaggerated a little bit because I wanted to kind of show how it asks different questions before it an agent can give a

response it works but it takes many turns to arrive at the response right and the solution is not personalized to Joan now consider this revised conversation so Joan again starts what's a good place to eat what kind of Cuisine

do you prefer today that's what Watson asked and Joan says well any of these four is fine and that's it let's say this conversational agent now comes back and says Johan because you're a vegetarian you're open

to experiences and you prefer fine dining and you are likely to drive given that you are currently renting a car I recommend this restaurant the ganas which is a fictionary restaurant it

doesn't exist slow cooking Indian restaurant on such and such uh Avenue in San Jose and it has ample parking so what exactly Happening Here is the

conversational agent actually understands Joan's preferences and who Joan uh is as a person in terms of her being open to experiences and so she may try out this slow slow cooking

restaurant um this time and uh and also knows from the context that she's currently renting a car and she's on a business trip and she has uh and she's so she's likely to drive there so taking

all of this information into account um Watson is making a recommendation so you know cognitive assistant is demonstrating all these preferences and understanding of user and user context it takes far fewer terms in the conversation to get to you know the

recommendation and it's also a lot more personalized so that's where we want to get to with user modeling by incorporating user modeling into conversational systems it's like a conversation with your friend yeah that

that understands you who you are and has a context and has your history of your preferences and such right so let's consider the second

conversation um this is a health um hline scenario Kathy is a a teenager who is in um maybe you know middle school or you know early High School uh and she is

talking to this hotline that's been set up by her school and um uh so she starts out the conversation it's an agent that it's not a human it's an agent that she's talking to I'm getting bullied in

school I'm so distressed Watson says Kathy there are proven effective ways to deal with bullying and Bullies would you like me to help you and she says yes please she the system asks because it's been

configured for bullying says are you facing verbal physical or social bullying right Kathy says okay I guess she doesn't know what those things are but I guess it's social and verbal bullying I'm unable to focus on my

studies my situation is so helpless Watson says where has this been happening you know C says in school bus in recess after school you know everywhere why do you think they're bullying you I think it's because I'm

overweight they call me all these names it's so demeaning have you tried avoiding those who are bullying you I can't they're my teammates because I have and I have to do a project with them now whatson says well all right

here is a short video that helps you uh with different techniques to deal with bullying right this works this is typically how health heartlines you know

a regular one you could imagine would be built but it lacks compassion this person is really distressed and she's saying I'm distressed my situation is hopeless it's so demeaning and uh you

know if you're talking to a person who the nurse counselor the nurse counselor would be a lot more empathetic or at least compassionate toward understanding the problem and and um she may come up

with same kind of responses but um a human would behave lot more differently in terms of the responses um that he or she would make to Kathy's situation so let's consider this revised scenario I'm

getting bullied in school I'm so distressed Kathy starts out Watson says Kathy I'm so sorry to hear that you're facing bullying in school don't feel distressed you're not alone in this fight there are proven and effective

ways to deal with bullies and bullying I'm here to listen and help you in any way I can would you like me to help you she says yes please so goes on same thing questions about verbal physical

bullying and she says my situation is so hopeless so here says oh that is indeed unfortunate don't lose hope Kathy you can manage the situation and I'm here to help you and goes on let me first ask

you where is this bully happening and she answer it um so it's a compassionate companion and you know the system is saying I see I understand I see like a human would say um it's so demeaning when she said that she said I'm sorry to

hear that Kathy this is that is indeed very insensitive of them to say that right have you tried avoiding and um and what's and again says I see I understand the situation and and finally it says after you had a

chance to you know watch this video and the resources that I'm going to give you you know I I I can play this um role play with you that way you can be more prepared to deal with the situation next

time so how can this role play be enabled in in this conversational situation this role play can be enabled when the the chatbot or the agent or the hotline agent now can take a a Persona

of its own and behave like a bully right as and and in the meantime maybe pause and give guidance to her and saying when the bully says this you say that and you know those kind of things so in order to

do that even there the chatbot needs to have some kind of a Persona of its own right at this point in time in order to do this role play um I need to take on the role of a bully so these are the

kinds of you know personification of um these chat agents and making the conversations a lot more compassionate and more natural and also personalizing them for people is what we mean when we

say you know user modeling in conversations so now let's yeah let's talk about um and I already stated that this is this is the goal you know this is about inferring

users personal personality their context their emotions their sentiments and incorporating all of those into conversations so let's um see how can we

measure let's say we built such a system how do we measure the performance of such a system what metrics do use right so it has to act the the agent this when

I say agent here I mean usually the the automated agent digital virtual agent it has to be able to come back with more compassionate responses in a fewer turns

you know right get to the point quickly because you have all the context and you understand the person and uh it of course the time taken not only the number of turns the time taken to get back should also be shorter and the

measure of usage of the suggested recommendations right does the user eventually end up liking the recommendations being made uh would they give thumbs up or would that convert into their purchases if it is a

recommendation or actions being taken as in the person actually goes to the restaurant recommended by the system right that that would be a good way to measure that the recommendations that we're making are actually useful and that's a standard measurement for any

personal any personalized or recommended systems and that would apply just the same here right and um and we have to be able to increase the usage of the conversation systems because they're so useful and they're actually like talking

to you know a helpful person who understands their context people come back to that system a lot more right so those that would be another criteria as well and eventually you can also do a survey and find out that um whe whether

customers who interacted with that agent liked it or not so these are some of the ways in which we want to be able to measure the system that we eventually built and I'll talk about some of the things that we have done uh and of

course we need to do a lot more um so now let's talk about where when a when a uh conversation system is built typical conversation systems have these kinds of

components right a user is interacting with some kind of a input module of the conversation system the input module is really doing the speech recognition if it is speech based or text recognition

if it is text based and such uh and doing gesture recognition and all that and then translating the the said or spoken or written content into the intent of the user right what is the

user actually saying so that's the natural language understanding part once you understand the intent and what the user is saying then there is the the dialogue management typically that happens behind the scenes which is all

about um given this intent and this particular kind of question what might be the right answer to get back with and managing that dialogue process and coming back again with a response that

includes framing the natural language um response and that's could involve natural language generation or maybe just simply pulling the response from a set of documents or pre-and responses that you have and coming back

with it and eventually output if it is a speech based you want to convert the text information that you collected back into the speech or whatever the modality of the user that you so these are typical

steps in any conversation system so when we talk about incorporating personalization compassion and all these user models into conversation systems we look at all of these aspects in the

pipeline because the the bottom portion of the the the T the the rectangles here are showing the components of a user model that play throughout this

conversation process so when it's when the speech or text or gesture is being made those signals need to be detected in a conversation because they have they have useful information about whether the user is distressed whether the

useful is user is excited about something or is angry or what is the tone and intonation and those sort of things and in understanding the intent and emotions we need to be able to really you know go beyond the said and

spoken things to understand what is being spoken in the content and therefore understand emotions and such expens press and eventually when you get back it has to match the tone you know

for example if an agent if a customer came in very angry about a product uh capability and said you know I'm I'm really upset with this particular feature you guys your service is

horrible and all of that the agent agent's most natural response to that has to be more professional courteous and apologetic right because they are having a problem so knowing what is the

right kind of emotional response to get back with you in response to the the emotions that are being expressed by the user is also important and it's and that is part of really framing the response

and getting back to the user so we argue that it it you know user models play an important role in all of the the aspects of the the conversational

systems so now we talked about uh some examples we got an idea of what it means to have user Models Incorporated into conversation systems and we talked about where all it applies now what are the

core models that we need to build in order to enable such such systems right so let's take a look at that so when um when we talk about cognitive

systems we usually talk about this cognitive systems have the ability to interact naturally with people right we say that as a very um important aspect of conversational systems now let's

break that down into these different parts and actually examine what it means interact naturally with people right so it has three different parts there is people aspect there is interaction aspect there is the natural aspect so we

argue that there are these three core models that need to be built so the first model is about interacting with people so that is who are the people have models for understanding people at a deeper level then the second aspect is

about understanding the interaction patterns between humans to humans and humans to computers and really and bring those that knowledge of that interaction into into these models and finally the

natural interaction part involves you not only you know one kind of mod ity like text it could involve speech um you know text audio video all of these so we need to incorporate all of these into

into the system so now the rest of the talk is organized into these three different chapters the people the interactions and um the natural mediums and how we can bring all these things

together into user models so I'll talk about the models that we are building for people and I'll show live demos of it and I'll talk about some of the the techniques we've used and the accuracy and the measurements and such in this

section of the talk and the the remaining ones are work in progress so there would be more sort of what we intend to do kind of a thing so people you know people are complex beings we know that right they have they exhibit

different emotions they have you know different skills they there's different context in which they operate they have different preferences they OTE differently given that people are such complex beings how

do we have a model of people right how can we build a model of course you know there has been a lot of work in this area so the question is we looked at

what are the different uh um models that were built by others in the past to understand people there are starting with the field of psychology there are many Marvels in the field of psychology to understand people from the point of

view of their personality their emotions their uh needs values and all of that and we we use several of them and there is this um Effective computing field which focuses more on the the notion of

um the feelings and emotions of uh of people there is decision theory that has looked at how people decide and what are the different criteria the rationality versus irrational decisions and all of that lot of work in that area and and

marketing folks look at consumer Behavior right there is uh lot of work in understanding consumers from the point of view of what they buy how they um gravitate towards certain things and

not so there are many models that we can rely on to to build um a model of people but the question is which one is the right one for us or is there a um a

holistic one instead of just looking from the point of view of how the user buys or how the user decides or how the user behaves can we have a more holistic view of the user is what we we asked ourselves but then why do you need a

holistic model is that really important is another question right maybe you just need a model that works for the current situation that you're in um why why to boil the ocean well we argue that uh for

us specifically in Watson when since we are building building blocks and common platforms that are being you know used by developers and everybody out there for various purposes and various use

cases having a we we cannot predict which use case a particular client will use it for you know for example when we built personality inside technology which I'll tell you more about we um

knew that it would be applicable to to match people with people but little did we know that there would be a a dating company that would be using our service to to match people up so there

are many situations that we would encounter um and for which different kinds of user models would be required so we we thought well we need a at least for us to think about we need a holistic

user model and um because we are releasing this general purpose platform on on for developers so we created one by assimilating all of these so we use this as a guide guideline and obviously

this is going to be a journey and we we're not complete in this journey so users so we created uh model thinking about what is the inner state that

user um has and how does the user interact with the environment and what is the outer state that eventually these actions manifesting so the interstate part is about um how do users who are

the users what is their existential being part of it how do they feel how do they think and the environment aspect of it is that what is the context around them how do they interact and what are the the different options and products

and features and things that they're exposed to and uh the outer aspect is taking all of this information into account how do users explore and decide right and based on

all of that they they're taking actions those actions could be purchases anything so let's open it up a little bit so what does being mean what does feeling mean let's take a look a little

bit so being we look at being as not only the the demographics personality aspects of individuals it's about the personality their needs intrinsic needs their intrinsic values their um

Ambitions goals motivations all of those kind of things are part of who an individual is and the feeling aspect is not only about sentiments emotions but also about you know the feelings the attitudes they have about for different

things and and such thinking is about how do they think what is their cognitive style um what Knowledge and Skills do they have what do they bring to the table and uh if you look at context it's about what is their lifestyle what are

the life events that are happening that influence their decisions and there could be several social political technological economic things that may be influencing the context that U that the user is in and the standard

marketing Mak type of options right what is the what are the sensitive to in terms of the products crisis that they are being exposed to Promotions and such and uh how do the taking all this how do they search what preferences do they

form how do they communicate um their intent and what choices do they eventually consider and how do they decide and those actions manifest in purchases commitments um decisions and

you know all kinds of things which will have consequences and in user models we need to have this feedback loop that when we observe that those actions we should be able to come back and refine

our models of who they are how they feel how they think and such so this is a model that we use um and I haven't specified the transformation function at this point on how do you take these

emotions personality needs values opinions knowledge skills into really predicting those those actions right that's where a lot of interesting work happens right I mean having a model to

be able to infer all these attributes is is great and that's what we will uh the work that we are doing but in the context of the actual applications where these models are being deployed and used

a transformation function needs to be specified so that you take all this information that you're able to infer about the users so that you can predict the actions and such so we'll talk about some of the things that we have done but

obviously those who are using these in in their systems need to Applications need to be able to BU build those transformation functions so here are some models in literature for

understanding people uh we'll go through them very quickly to get to what we have done and some of the the results so several of you may have seen this

Iceberg model uh um which talks about how people's emotions thinking values and beliefs really Drive what manifests externally in terms of their behaviors and actions right and there are several

variations of it so basically what this is saying is that it's yes it's important to observe like how we do right now in all of the the recommended systems of what people purchase and what

they are doing but what marketers are really after is not only an understanding of what they bought so you could recommend the next set of things that people bought but really about why

is it that they buy what they buy if you understand that drivers those drivers you can actually construct your marketing messages a lot more effectively and also probably consider those types of

product features that would appeal to a particular uh user and and talk about those in depth right and there are several of these other models Big Five

um in marketing um which which is about the openness how consci conscientious how extroverted how um agreeable and how neurotic and individually is and then the models

about human values needs and and emotions yeah so I've seen the big five stuff for quite a while since the project started what is there

independent work that actually says understand let's say that you can actually model emotional state these there separate psychological work that

if you understand the modeling and you affect the advertising message such and such a way you actually increase the result you're trying to get yeah I will show some of those results today is that

something that for a long time I never nobody could could show right we we we we have some of those um um studies and I'll talk about those and one is an in

Market experiment that we did with um U Bloomingdales and Ruchi is here she probably you can comment on that as well so just kind of quickly going over some

of these um um there are models um that talk about personality types Meer Briggs models several of you may have taken some of these in your um in your U jobs

uh or in student life uh that tell you about what kind of a person and what your persona might be and there are models like um uh uh these dis which which classify people into foure

categories based on whether they're dominant or influential uh steady or cautious so there are many models out there and marketing T psychology folks typically look at um surveys as as the

mechanism for for getting to all of this information um to understand people but we know that surveys don't scale so we use uh machine learning type of techniques because people are leing digital Trails all over the place these

days right on social media and everywhere else so it's almost like instead of you sitting and watching on Spanish Steps who people are and observing them and and who they are um you can now read their social media

content and get to know them better and understand who they are and and uh um you know what their personality and what they likes and dislikes are so I tell you sometimes I find that a little creepy you know Ron we know that you

like sports and I think God read my email or something something of Interest anyway I do find that a little bit creepy sometimes that if they know a little bit too much about me they don't care for them to know right so I might

as well come to that point right away yeah privacy is a big issue in this area of user modeling in general um and the way to get around it is by using optin kind of uh approaches and also

anonymizing it so let me give you examples in an opt-in approach if I have on my mobile device um this travel assistant or a digital agent that will help me in my situations let's say with

travel reservations and all that and I'm I'm actually more likely to give it my my access to my social media data to use that to personal give more more personalized travel reservations and

travel uh plans for me and I know that it's used only by that agent and it's not passing that information to others right so in that case I'm more likely to share the information because it's helping me it's adding value for me and

I know my data is secure right so in those kind of situations privacy hopefully should not be such a concern um and in other cases where you can actually anonymize uh understand the parents in in uh large amounts of groups

as to how people behave and why they behave that way and um uh and when a particular individual comes again based on optin you can personalize the recommendations based on that you confident they're not showing other people because they say they're not

going to do it or why no these are applications well um applications that uh you can control and you can EXP explicitly give access or deny access to right saying you know no I don't want you to use any of my I give them access

but how do I know they won't share somebody else yeah I mean they legal laws and you know those things can can cover that right I mean that's a standard question with anything I mean if you have a contract with a company

saying that you know you would uh use it only for this purpose you you're bound by that contract and you know loss could could come after you to if you violate

that right so all right so what do we have so let's get into some of the the uh the actual cognitive building blocks we have built so far um personality

insights of course tone analysis and uh and emotion analysis which is part of alchemy uh language API for those of you who are not familiar with Alchemy language API it's it's one of the

umbrella building blocks of cognitive services that we have on Watson developer cloud and this image is taken from wats and developer Cloud again for those of you are interns who are not familiar with um if you go to IBM

developer watch and developer Cloud you'll see this services and you can try them out okay um so uh I'll show some quick

live demos and then I'll immediately come to um Brent's question on how do we actually make these things actionable do

these things actually make so this is the personality insights again if so I'll go through this quickly you can all try them out yourselves um this is a service that's out there you

can uh So currently Opera is selected and her Twitter account is being analyzed and here you are um with um her personality uh based on her tweets um

these are the the ones that you see the personality needs and values these different dimensions that you see are the the the traits that we have inferred or analyzed and from her Twitter data

and are able to infer and um her scores are compared um to people who are tweeting other people who are tweeting on Twitter so she's 19

the 92 percentile 91 percentile for extrav verion and so on that's the way to interpret these scores and same thing for needs and values and up here here is

where we are we are um slowly getting into with our work um as to what do people do with what you know it's good to know that this is their personality but that by itself is not actionable right how do you um yeah there was a

question so uh these bins for Consumer needs uh you came up with them based on a preexisting model or you derive them from yeah these are from the models that

I had shown earlier psychology models and um yeah we we use that and but we're applying machine learning techniques to learn these

traits so what what do we do when you know that somebody is conscientious or open or uh you know is um uh somebody who prefers tradition how does that trans at into their consumer uh

behaviors and buying parents is what uh we are beginning to address with various studies we are doing and in some of the recent studies we did we were able to um correlate these traits with these

actionable behaviors for example person who is high and agreeableness high extra verion is likely to buy co-friendly products likely to adapt to situations spend on health and fitness and he's

unlikely to put her health at risk and unlikely to change careers so in Oprah's case uh that seems to be true she has been in that entertainment business for quite some time she hasn't changed her

careers so and then there is a summary as well so um what we are in fact this is going to be uh something that we are releasing as part of the personality insights API so right now you only get

these these traits and their scores uh but very soon um you are going to be getting a list of those attributes as well as to what you're likely to do based on what your what the scores are

high for which traits you you will know what they are likely to do in in certain of the consumption preference Dimensions obviously you know we cannot do it for everything there are so many of

those uh so that's one that's about personality insights now I'll I'll bring all these things into the user models and what we are trying to do in just a bit so then we asked ourselves can we infer you know in order for personality

insights to work we need to have about uh uh 4,000 or so words um text uh that's written by individual and often we find out that it's kind of difficult we don't often have that much so we

asked ourselves can we infer one's personality from the text that people from sorry from the images that people are sharing and um so this is a project uh that is a currently ongoing one um

the results uh um although there is a demo we are still trying to refine the the technology Tom Zimmerman leads this work um so here it is um uh so if you

specify let's say Bill Gates so it's now sending the the images um retrieved from Bill Gates's uh um uh Twitter profile and uh is passing

through the vision apis and uh get a list of those uh um interests and what the images are about type of uh tags and

using those tags to run this um machine learning algorithm to correlate those tags with um with the personality traits so when this is done uh we'll get back

the personality of Bill Gates as infered from images so those were the top 10 images that uh top uh nine that are being shown

retrieved from his Twitter account and uh um here we are showing on these different dimensions where does b gets fall so some of the questions that we

are trying to address now is uh do the personalities of individuals collaborate when you infer them from text and images uh when you get it from Twitter versus

Facebook versus blogs versus emails is they are they the same uh if there are differences what are the differences and why so these are some of the things that we have only preliminary results we

don't have good results um yet to talk about um but these are a couple of projects and now I'll show tone analyzer and then we'll get into where we are with some of these Technologies so tone

analyzer yeah you comment that um some people try to cultivate an image with take Donal Trump but we infer things from that very cultivation that they're trying to do things they they would not

be happy about are concluding well yeah so people can choose the Persona that they want to project on different

platforms um and um of course that is always the case and and we'll be able to only in for the the the Persona that

they chose to protect in that case but if really kind of over a period of time if we have enough text it shows the field of psycho

Linguistics shows that you are likely to converge the results are likely to converge to your true personality the true personality being inferred

from um the surveys now one can of course also fake the surveys by answering them incorrectly or not be genuine uh if that is the case

then yeah we have a problem but uh I think that other than asking people some non not so direct questions and inferring their personality we don't have other ways of really understanding

one's personality right so one of the things that we assume is that there there are some basic axioms for our work and one of the axioms is that um that the surveys are the ground truth that we

rely on yes the surveys could be totally um you know misrepresented by the users in which case we would have a problem but to the extent that the surveys can be relied upon which the psychologists

over the years have done and said that you know these are the the best measurements that we have of infering one's personality I'm just saying something a little bit different Trump might say over and over I am

important and he thinks we're going to infer from that I'm important instead we infer from that he's a guy who tries is very insecure is trying to convince us he's important well just because somebody uses the word important doesn't

mean that that person is important in the personality traits you know all these words are used and uh their relationship to the personality traits is is done through studying multiple people and their personalities right so

it's not necessarily the case that important word occurrence will result in person's personality being a particular type that's associated with

importance so we move from personality to understanding one's tone and uh one's emotions um so here is the service um for uh those of you who haven't seen it again for those who have seen it I'll go

through quickly um there's a customer conversation going on customer is um agent and a customer are conversing this is a chat uh session

um so we analyze the the tone in it and um we get back what is the what are the emotions being reflected at the document level and what is what is and what is the language Style and what are the

social tendencies in the in the document that are being expressed so um in this particular one anger seems to be the dominant emotion and um there is some

amount evidence of analytical and some tentative tones and the the entire conversation is expressing um quite a bit of emotional range and agreeableness and we can break this down into what did

the agent say what did the customer say um and uh get those different kinds of um individual um individualized um tones

and we can also uh get the information at um at um sentence level so if you want to know why what is anger what is reflecting

um anger in this sentence which sentences so where the customer is saying you can't spell do you understand the problem I'm having or is this a lost cause um this is ridiculous so the

customer is really being very angry in this uh conversation and it's picking up those sentences and classifying them as angry sentences so this is the tone

analysis so now let's come talk about um all right so what how can we really say that understanding these personalities and understanding these tones and Communications and emotions actually um

is making any difference so here is one um demo that we did with uh Richi who is over there um it was uh done with a

large retailer in San Francisco and um so knowing the demographics here is important because the kind of products that you recommend depends on that so

you select the female let's say and the age group and um if you give the personal if you give the the Twitter idea of the person to whom you

are say offering a gift um it infers the personality of that person and knowing that the person is of this type of traits What kinds of products can we recommend and um here is what where you

know from the catalog of that particular retailer a particular set of products are being recommended because the personality of the person has been matched with the the brand personality

of the the product here and um and we we have bu built in some um matching rules that says personality of this type is likely to buy products with that have

this kind of brand and image and uh so these are the things that are being recommended for Adele in this particular case so now for this particular uh thing we um

actually Richi and team the marketing team ran uh an in Market experiment where they said if we made a recommendation um using the the personality traits of the individual did

these users actually uh choosing that product as for purchase so on the kiosks that were put up in front of the stores

um about 11% of the the people actually ended up going through the recommendations all the way and said I will buy this product and when they clicked on that buy it showed them where

the product was in the stores and that 11% compares to typically online about 5% conversion rate from browsing

searching to actually getting to to uh um considering buy so this is one such an in Market experiment we did that shows that you know you can get two fact

factor of two kind of uh Improvement in conversion rate when you actually personalize products so let's um get back and say what is the technology behind some of

these um so I won't spend too much time on personality Insight several of you have heard about it before um but we'll just kind of show this um latest things so we started out with

using linguistic models um there was a dictionary called linguistic inquiry word count that's very broadly used in uh many of um uh these these Works where psychology and linguistic areas come

together so we started out with using that dictionary and um but more recently we wanted to move to more machine learning based approach and not necessarily use any particular dictionaries to guide the system and and

um here are some of the latest results on the mean absolute error as compared to the psychometric surveys that we got on people so um you know when compared to the the linguistic uh the model that

uses those linguistic the dictionary versus the new model uh we are pretty comparable and in most cases actually do slightly better and uh that shows the

correlations for these five Big Five Dimensions so moving on um this is about how many words do we need to INF personality uh and we did experiments on that and uh we also show

that if we use the the more machine learned model um you can actually converge even with about 30 tweets to 95% of the accuracy that is

possible go ahead Anis kind of tax analytics first conver

each into tax and then tax anal Anis currently the one that's released so the question was uh do we still do speech to text translation for inferring

emotions um the the one on um uh that we have released yes it still does machine translate you have to do the speech to text translation before you use the um the the emotion the text based emotion

model we have but uh uh folks in research are working on uh uh models from speech directly and uh gestures and other things which will be coming up

later on what is the the sort of sucess rate me accuracy for emotions let me switch to that oh uh okay let me switch to that

and I'll come back to addressing one more question that um so these are the the the results on emotions we are right now at 66% uh you know 67% accuracy uh

for emotion uh for the five different emotions that we are um labeling there's two flavors of emotions there's immediate emotions and there's

long-term emotions you longterm you're a happy person short term you're an angry person right so this is um the dynamic the what's happening based on each specific statement sentence you are

making so this is short short term Yeah the more fleeting one the long-term emotional disposition actually translates into one of the personality tricks which is the emotional range what

is the typical emotional range you tend to have as a person in general um yeah this one is more sort of the immediate one okay just one last question so I

find this started soci and I'm to know if there's any place

especially uh with the teenage and the young kids right to have this implemented as a do

you know if there are anys place for demo or um so I know internally the one of the cognitive build projects that won the

cognitive build competition has been looking at this kind of Healthline um and discrimination type of hotline scenario um so that's one internal

project I know which is looking to use some of this uh other than the immediate set of mini demos that we are building some of our ecosystem Partners like ID

avatars um have been using uh considering this technology emotions and tones and personality to um give a Persona to those avatars which are being

deployed in health hline kinds of scenarios so there are some client startups which are exploring these Technologies there was a question there

so uh you said you used uh what was what was the oh that was uh the stand Stanford U

glove uh pre-trained pre pre-trained word to yeah okay so one quick again one quick thing to answer Brent's question um

there were many studies that we did over the years since the original personality insights to correlate people's personality to actionable behaviors and here are some of the studies and uh

several more are getting released as part of the API um so you we know um that uh people with those specific kind of Trends are more likely to respond to tweets which means they're likely to

collect information who are who are more likely to spread information who's likely to respond to unsolicited advertising types of things what kind of people redeem coupons more often um uh

more likely to redeem coupons and who are the kind of people that show brand Affinity versus who don't care about Brands um who are the people who are likely to have a sense of Satisfaction by being part of a community and um you

know who are the people who express interest from uh gaining Knowledge from content uh and what would their personality traits be and these are um some of the studies that we did and many

in the literature that others have done um but we are only releasing the ones that we have done as part of the API uh which hopefully we think will make the system lot more actionable now the the

traits in building these recommended systems and conversational agents um so yeah I mean I'll skip some of these um uh yeah so if you have for

example you know typically you know there is Lisa a person who shops at Target Amazon Whole Food she's been buying a lot of maternity clothes recently because she's pregnant typically you know you would get disposable diapers kind of

advertisements but if you knew that she was high in altruistic and is high in Morality and conscientious person and high in moderation you may actually offer her coupons for cloth diapers

instead of that so these are some of the the examples where knowing the personality actually makes a difference in making recommendations so we talked about the images let's talk about emotions as a as a as a technology so

given a piece of short text about three to four sentences the goal is to detect emotions um you know these uh five different emotions and um we are using an ensemble

model which has uh different kinds of uh um algorithms in it and uh the best of the best are getting chosen inside The Ensemble here and um you you know here

are the results as um uh and we we collected about 100,000 um annotated text um uh sentences that have been annotated on these different

emotions and they include semal ier types of data sets and also Twitter hand label data sets um some of our members who lead that project pram is here for

example and J as well um and uh here are the example here are the the the accuracy I mean we we have ways to go but this is where you know most of the

emotion systems are at uh with our Ensemble models with by mixing up different data sets that we have we at about

66% accuracy in predicting these these um five different emotion labels and we're working on more um positive labels

and additional ones as well so why a lot of people ask me this question on why emotions how do they differ from sentiments and how can you actually act on these emotions so here

are some of the things that we are looking at you know there is some you know this is really pulled out from uh the uh product reviews um online

somebody said you know why would locating a store with mileage feature be taken away from this app at least sorting you know whatever was put back on these are just stupid decisions on

your part you guys even try out the app before you put it on the market so sentiment analysis shows that's a negative emotional analysis say that's angry um and you can infer that the

person is unhappy with the product but here is one that's actually interesting so why would locating a store with mileage be taken away from the app at least sarting was put back on the fact that these things are being taken off and put back makes me nervous you know

what are these guys doing do they know what kind of product they building so of course the sentiment is still negative but the emotion is fear right I mean they the person is nervous about using your products so the kind of feedback

that you would get is that it's not only that you're unhappy with the product right but actually you have a little more fine granular feedback on the product now that there is some kind of a trust issue with the company that the

customer is expect expressing right and that's that's important information to know and um and where else could it be used I mean in this customer loyalty

Spectrum people often want to know where exactly is my customer what can I do with this customer right is this customer being optimistic thankful satisfied or is the person really you know in that zone of frustrated and it's

not just easy important it's not just enough to say that you know negative positive you want to know these pine granular sets of emotions states that users are

expressing because actually now marketers are starting to really take advantage of this fine granular information and some of these companies like Kya for example to you know label

you know this person is a loyalist this person is an enthusiastic this person is a refiner and that person is actually a Defector and is likely to go so it's important to know where where a person

is in the Loyalty Continuum so you can engage with that person better the kind of message the kind of marketing that you want to have is different based on that so overall user modeling with user

modeling people have a lot of privacy concerns you know Ron brought it up I commented on it that you can either do it with optin or by being anonymized in a conversation setting though it's

usually kind of a one-on-one thing so hopefully when the user optins the opts in their um context and user information can be used better in in um serving them

better so now let's talk about you know that's the work that we have done some of the work that we are doing I'll briefly give an overview of it um how do we do understand human interactions better so when people are interacting

with one another we know right when that person is angry or that person is happy that person is really being um you know just indifferent to you um How would how can we make that an a conversational

agent really understand when the person is conversing uh what kind of particular situation the user is in right how do we um you know what are the patterns of these human conversations that we can

learn from so that when a an a person is speaking we understand that pattern of communication and have the the agent the chat system or the automated agent

really understand the context and the pattern and and act based on that so there are again several theories uh of human interaction models there is the human interaction Styles which is a

Myers B person model we already covered that and this is about understanding the personality but also what kind of style or be what kind of actions what kind of

um response types work better for people of these personality types right that's what we are after now in the interaction style knowing the person is the first part that's good but we want to go beyond that and say what kind of things

work for this person of this personality type what style of interaction so for example for somebody who is uh if we go back to this one who is high who is dominant and is visionary and you know

who uh you know just wants to get the high level picture and move on for that kind of person you don't want to give too many details right for somebody who is you know very much into detail oriented you want to provide as many details about the product and the

situation as possible in order to make a condensing um argument so really kind of knowing those patterns of human interaction and incorporating them into conversational systems also makes the

systems a lot more um natural and personalized so there is this other one by you know Linda burn she talks about uh another model of her own um about uh

you know how um you know whether this person is the kind who charts the course or what's behind the scenes or is in charge of the situations or get things going and in literature you know there is techniques that people talk about as

to how exactly you would interact with such people and the speech act Theory um some of you may be familiar talks about um really understanding humans spe humans

from the speech perspective right there is um three different main kinds called locutionary locutionary per locutionary locutionary talks about you know literally what is being said you know if

you say uh you know you know don't go into water that means yeah don't go into water you know that's it's exactly as it said and that's what the person meant in the elocutionary one one when somebody

says there is a snake under you it actually means be careful right so there is a statement being made but there is an indirect there is a social function asso with and in the perlocutionary kind

of speech Act it's about the effect of what is being said so if you say you know and these are examples from um literature and Wikipedia you know if you say I have a CD you know would you like to borrow it you you may actually mean

it but you may also be showing off a little bit right that I have this CD and I listen to this type of music so there are some unstated types of actions in in when people speak and when I'm talking

to somebody that may be an an internal agenda that I have and that person has an agenda and these these the Dynamics of the conversation really plays out based on how people are interacting based on their context and what kind of

what they are bringing into the conversation so what we are doing currently is really looking at these um types of human interaction patterns and trying to model these and trying to

understand if we can derive these from um the conversation systems and patterns so there are these you know for example um you know agree when the person said something is this a statement that's about agreeing or disagreeing is this a

person that's apologizing statement is this a person that's accepting or is this a statement that's uh of the type thaning is this about advising asking so if we knew exactly what kind of what

category we can put this um statement that a person made um at an abstract level although this is an actual conversation is domain specific but at a at an abstract level we are

understanding the kind of conversation that's going on this person said something this person answered it this person said Thank you in return or that person came out angry this person said apologetic statement that person gave

something else and answer ended up in so problem being resolved and eventually the customer satisfaction uh led to customer satisfaction so we want to understand these parents in speech that lead to whatever the outcome of the

conversation is whether the use of being satisfied or dissatisfied use of being um you know um liking the recommendation made or whatnot so these are some of the

the types of um you know speech Acts or dialogue acts is the follow on that comes from it um that we are now learning from uh from conversations so this could be you know typically the

patterns in dialogue act that they look at are is this a statement is this a acknowledgement opinion or is it an appreciation it's no question nonverbal question yes answers um and a and a slew

of things that were found to be more relevant to detect in conversations that would have an impact in understanding what's going on and how the Dynamics

played out so we're mining um several conversations in um uh customer uh agent um uh call logs and uh

discussion forums and uh and such to understand these types of parents and hopefully we'll be able to share more results so this is uh comes from the work that our um collaborators in Hyer

team are doing and they came up with these interesting dialogue strategies um you know based on so if you knew that somebody is you know thanking acknowledging you know all of those so

what do you do right it's good to know first what uh to do what is being being said the next thing is for the conversational agent must know what needs to be done afterwards right in

order to really respond more appropriately to that particular pattern of conversation that was detected so for example you know if you um sense

selflessness you may offer choices or you may uh Empower them you may you know provide expert recommendation if there is anger maybe the right thing to do is to allow them to V for some time right

strategy after that you may want to acknowledge you know without encouraging them to go further or distract them if it is a kind of a scenario where distraction is the best medicine for anger there or you know take a time up

right there could be different strategies uh and ways of responding to these emotions and and U the conversation patterns that to detect in conversations and respond accordingly uh

and so these are some of the things that we are now exploring by mining these call logs and other things to incorporate um different dialogue strategies is there a reflective

component because you may have misdiagnosed a state of confusion or embarrassment follow along of half and

realize that is you're misdiagnosis and you're making it worse or you're causing additional problems is there a reflective watching the conversation to

see if it's playing out the way probabilties that's a very good question so how do we make sure that what we are saying that the our initial um understanding of the way the

conversation is going is right and the strategy that we are taking is working or not this happens with I mean this there's a whole underlying issue of semantics with machine learning and I use semantics in the programming

language semantics right correctness in programming is either it is or it isn't correct right in this world it's all probabilities right and right women if

we are Hing around 66 70% accuracy we could have mislabeled a particular emotion and we could have been taking the user down the path of uh you know some kind of a strategy that is actually

inappropriate um one way to do this is uh you know actually get a confirmation from the user uh but then you start having those much longer conversations like your very slide where it starts

feeling really stilted I don't know one of the things that we are trying to do that's why is actually mine a large number of conversations in

customer uh care scenarios that have happened to see deriv from the data what kind of strategies to did those agents who with whom the calls ended up more

satisfactorily for the customers resulted in what were the parents can we detect but you still sound like a computer scientist that's trying to build the correct system yeah I know I

think you know we have Fally never going to be a perfectly correct system so you have to adapt to the built-in errors right right yeah mean there I think

checkpoints in between to validate from the user that this is indeed the case or uh uh say you know exactly like it may add to few extra turns but uh you know

that could be a simple initial way in which we could incorporate it saying is this helping you or do you I I sense that you are angry with such and such uh would you like me to help you so if the

user says yes then move on with the conversation can tell a conversation was wrong there just always constant like does he understand what I'm really

saying is he back on P versus NP yeah sure yeah I know I that Bob you a conversational uh agent too if you have any comments on that I know that's an important um problem that uh from a

computer science perspective it's always be there will never be perfect right yeah I mean there's the issue of are you correct or not and there's the issue of

did you uh did you understand the the uh the user or the the other speaker and I mean you you know you can give you can give the the user an answer that is

objectively correct but if they don't understand it then you haven't yet fully you know accomplished your goal of getting this particular user who has different state of knowledge than another another user to understand the

answer to the question so yeah it's it's and you you start off with one answer and see what happens right and you see if it's you know because you don't know what they know and so you need to have you know sometimes they'll get it like that other times you may have to do some

repair so that this person can take some extra turns to repair that this person understands what that answer means even if it you know even if it's perfectly correct at the

start people exactly that's what I want to say it happens even with people so it's I I make assumptions about what you know and then I try something right and see what happens and sometimes I get it right first time and other time I have

to adapt in ways that I could never predict until we get into the N so there may be an analogy in teaching scenarios because you're doing the same thing with students I mean you're you're projecting

what you think is the most likely way that they will understand a subject some of them will get it right away others won't and if you go down the path if you

think they know and you you on that path you're probably making it harder to step back I don't know maybe there's another area yeah that's a good point yeah there

there might be strategies to learn from from other fields yeah like education and teaching you had a question and I'll come back to you

so that not that intend I had a question the conversation there's always a possibility but again you have to kind

of balance that with how um you would annoy ending up annoying the user you know like that uh Microsoft clip example that everybody sites right you could you could really like it or if it's showing

different faces this and that you could get really turned off by it so it's a balance we have to strike I mean I guess it's part of it is in experimentation and part of it is the domain and the kind of uh application if it's a professional application you probably

don't want to do those and if it is uh you know more sort of with the children maybe you teaching Aid maybe you want to incorporate some of those and such yeah

there was a question oh so I had an idea what do you think of um using the same thing to build a better gaming

experience so for a person who's playing games you know to predict what they'll do maybe create a better environment for them I mean has anyone thought about

that yeah that's a good good point so I mean we we we are building the core technology uh we don't I IBM as a company you know we won't be able to build applications in all domains so

when we build out the platform and release it anybody who can make use of this for their domain can use it hopefully gaming um there is already quite a bit of interest in general in in

Watson and cognitive Technologies for gaming um hopefully they they will be able to make use of this and build their own applications with it you know I IBM itself is building industry applications

in only select set of domains where we know we know the domain well well enough and we have the route to Market and such um gaming is probably one thing that we would work with our business partners to bring those applications to Market but

yeah that's could be a very useful one and especially um in multiplayer games you know there are different kinds of Dynamics going on and

different things going on um yeah how to adapt the system to the next level or downgrade it uh depending on the level that the user is at and all of those things could be

important yeah so uh in the first few slides you talked about how uh helping with compassion would be helpful uh in

the conversation so uh is it one Zer or do you twe how much compassion you can add because like you said uh if like the 50 example it may be annoying uh and not

in all cases you want to show compassion yeah absolutely so this is one of the ways in which we are you know working with our conversation platform to integrate some of this is by making it configurable so you say I want to use

this or not or you say you know whether this uh particular capability of user modeling personalized recommendation or emotion based um compassionate response or empathetic responses would you like

to have it in your application or not so depending on the domain it it may you may be something you want it may be something that you want to use or not um there may be some applications where it would be very helpful in some

applications where you are primarily giving factoid type of responses back it may not be as important to have user model behind mind because no matter who the user is you have the same stock set

of answers but there are large number of them and we're doing you know um conversation based information retrieval and so yeah so it use it when it applies

is what so is it one Zer or you can bre how much yeah so the the gradation of uh good question I don't know we haven't thought about it but I think that could

be something that uh yeah when we detect it uh depending on users own uh you know intens level for you know emotion reception and and such um you know in

the emotion model itself there is this whole notion of you know detecting the intensity level with um Valance dominance and um you know those kind of models but uh

um we yeah do we expose it I think that's a good point you're bringing up I think something to consider after the initial sets of are gone out and we do some we start getting some feedback so let me kind of wrap up I

won't say much about mediums other than than to say that uh you know there is of course you know gestures there is speech based audio video there is a lot of more work that needs to be done we worked on

text based we collaborate with our colleagues in research who are working on other things that we'll bring um so how do we bring it all together um so

these again going back to the question that um uh Brent asked uh we did some in Market studies to see whether actually do these things actually make sense right so I talked about this coupon

redemption do how do we know that these personality traits make a difference people who are of these personality traits who are highly uh orderly self-disciplined cautious and moderate are more likely to redeem coupons this

is something that we have done um as a postmart uh study with a large retailer so we actually uh showed them that if they had sent their coupons to these set

of people 47% more people would have redeem coupons as opposed to sending them to random set of people so this was one of the studies do they want people

to redeem redeem coupons sorry do they really want people to redeem coupons well so this is the kind of scenario where you want people to come to the store to buy more things right you want

you know yeah there are some kind of coupons that like surrogate for more people com to the store and not actually yeah this is exactly to bring more people to the store

yeah and we also did some studies to see if you actually detect tones and if you actually understand uh the tones that people are conveying in a conversation

um um does does the tones that people are conveying in a conversation have any impact on customer satisfaction or any correlation to customer satisfaction right so that's one uh thing that we needed to know and based on that can we

predict customer satisfaction as in How likely this call is going to end uh based on the tones that were reflected in the conversation so those were the two things we looked at and um uh yeah

so you know Q&A Forum we looked at and we we noted that when uh confident responses were given um customers are more likely to give kudos to those

responses uh so it's the same response the content is the same but the tone is more more is delivered in a more confident tone um again tentative responses are less likely to receive

Kudos that's a correlation we've detected um and uh you know this more very obvious ones but at least we have validation of these angry customer more angry customers are less likely to be

satisfied after the conversation so this is you know based on the tone how the conversation might end customers who showed disgust uh we noted that those conversations are likely to end less likely to end in satisfaction

interestingly if the agent shows High emotional range the call uh likely to end in unsatisfactory manner as well so what this means is that um agents are expected to be professional always and

courteous no matter what the a customer is saying they cannot lose their calm we know this it's pretty uh intuitive there is no groundbreaking result as such but what's important is that the correlations are important that yes I

mean cont you think content is the most important thing in a conversation yes that's the primary one but after content the way in which you deliver that content to a person is also just as

important um so we all have been in these situations where you know we got a um response uh to some call that we made to a customer service agent but when we put the phone down we have a you know

bad test in our mouth right could be because the agent was rude could be because the tone was condescending could be because uh was very T and not polite enough whatever the case may be um the

answer is uh the problem is resolved but yet you are not So Satisfied you don't want to give a you know excellent five on a survey of 1 to five you probably would give a three and a half or a four

so in those situations it's this unspoken things the conversation style and such that make a difference so these are some of the studies that we have done and we also did the

prediction uh did talk about the prediction yeah so we try to predict so we don't know anything about what's being discussed all we said is let's take the tone and can we predict whether this call is going to result in a

customer satisfaction or not and with uh about 66% accuracy we can predict and I think that's uh pretty good outcome given that we it's nothing to do with content it's just based on the tone uh

and we are able to predict whether or not this call is likely to end you satisfactory ter or not all right so bringing it all together what does it mean how should a conversational system incorporate us a

models in it U again we assume that the conversational system is ongoing and we say this is how we are thinking about it then this is how we are going to be working on to build it um get the user

Global context and user context first and once you the dialogue service and intent are detected you want to see is this about personalized recommendation if so know based on the user model

personality and all the information that we have about the user model provide the personalized recomendation it could even be that the the answer itself has to be if it's not a Fati type of question the answer itself may have to be different

for this user based on who they are right for example the the restaurant case right I mean there is no one good restaurant you can recommend there are many and U you use that user context to

make the recommendation and also um we incorporating tone tone uh into conversations so detect the tone and also if the tone needs to be modified

like in the example um where where you say um I I didn't show one of the demos but we had one where user says I lost my luggage I'm so mad and the systems comes

back and says oh I'm sorry that you're so frustrated and um you know do you need clothing store to replace the clothes that uh you need to go to your business meeting next right so

acknowledging that you're frustrated and is something important and uh and that's a it's not part of the answer or the recommendation it's a response that the conversation agent needs to come back with

and that's what we call as tone modification so it's not just detecting the tone it's modifying the tone um and the the response either by prepending a pending or by by changing the whole

response itself to come back to the user with um a more personalized and compassionate respon and eventually when all this is done you know you w't be able to um of course summarize and

capture all this information for next times when the user starts interacting with the system so let me pause there and um and say come back

to yeah I mean there are interesting questions that uh you know we can all um debate about but these are pretty standard to any kind of a um

personalized system you know how tightly and you were alluding to some of those about you know how the gradation and you know how tightly should they be integrated should the users give consent uh should the users be aware that there

is a user model behind the scenes in making this uh conversation more personalized um should uh they be given an option and it all these are all application Level questions and

depending on what domain what application You're Building you can expose them and let the user know or not and from our perspective we want to kind of build a core technology and make it

possible for application developers to build build these systems more effectively so let me pause there so the conclude by saying you know goal is to you know enable these more natural

personalized and compassionate conversations and in in doing so you know we're focusing on the user model aspect of it and so today I talked about some of the user models that we've built

the personality emotions and tone associated with um uh with uh people in which in how they convey things and also the future work that we are doing which is about really understanding the user

interaction parents about how to model how what are the different abstract level conversational patterns going on and um uh be able to adapt the conversation based on that using diog

strategies and um some of these cognitive services that I have shown today are um available on IBM developer Cloud so feel free to go try them out additional ones hopefully

we'll be integrated into the IBM's conversational API which is also something that's uh um an experimental version of it is available right now to play with which mostly does intent

detection at the moment with with the disguise of a dialogue specification behind the scenes but it will be hopefully getting more richer in functionality in the

future any further questions from anybody I can take them now okay all right I think we're done then