okay
so i'm not talking about understanding the use a user in social but conversations
and i really representing
the a team of students here so i wanna knowledge the students are really a
part of this
i also work we enjoy denotes the and that's faculty advisers
and that it's
it's been a lot of fun and the students i really
okay so it doesn't know about amazon like surprise and so i should point out
that name they are here a sounding board and then i don't think it is
again something
and is an important e
was a response to a caller competition
which was the l x imposed on a lot surprise and so the idea
back in twenty sixteen base elicited skull and they want university students
to build a social but i
and well as it is also but to converse quote unquote coherently and originally a
with people on a good topics inference and so it's very open domain
so my gratitude on with the team leader said you want to do this and
i think you're crazy but okay
and he got it together and they wrote a proposal and the
and the intended to select and then we can then the field of the system
and all that
at the end of that we had about ten million or more than ten million
conversations
with real users
and between that
and the fact that we're working with the new type of conversational i basically what
we where it is there is a lot of research problems in that is based
dialogue that it is i hadn't thought of before and so i'm the focus of
this type is gonna be understanding user
a particular including user modelling but i want to start out by saying this is
used once all these the overall big picture i'll give you a little of those
probably picture is just one small piece
so what it what i mean by social plots so and why do i mean
by think this is a new type of conversational ai
so a lot of work in conversational ai has two spaces and people often talk
about it as two different possible task
so there is the virtual system and what task oriented dialogue
and in that
type of dialogue system
you're executing can we have the answering questions it or something that is social back
and forth
on the opposite end of the spectrum is a chap which is oriented towards chitchat
kind of how are you know what you're doing today but it really limited content
to talk about
i like to think of these not is to different option
but as
a two different types of conversation you know broader space has at least two dimensions
probably more but there is the accomplished task dimension where the virtual assistants trying to
do something in the chat but is not and there is a social conversation dependent
where the jackpot is being social but doesn't have as much to talk about
so what we are trying to do us something that's in between
we do we're a little bit less social and a little bit less
a task oriented
then the other two
well i i'd argue that it is to some extent
a task goal oriented because you're providing information
so there's some so most social exchange and information so with that background
what i'm gonna talk about
is initially that then of the social but for our chi specifically and that is
that the conversational gateway
and all of us system overview i'm gonna kind is true that because this is
early days of working on social but that's and the architecture with you is not
gonna be the our architecture that anybody'll use a couple years from now but we
need to understand it to see how we're collecting the data and what we're doing
then they want to focus in on characteristics of real users and this is just
an analysis somewhat anecdotal but i think it's important to understand where we're going and
then i'll start in panel talk a little bit of our first steps in user
modelling and out in this was something queries
okay so this is also by as the conversational way
so what
we see
is that this social but when people come to talk at social but they are
not they don't have a specific task that they wanted you don't wanna work
a restaurant reservation for example that they do you come up with some sorta
ideas of what they might
and yes or conversely
and they were new information and their priorities are interested in a of all their
goals available
and so the social but is still indicating that a balding
so one of vocal set
the users are also in this case
coming to a little a little device
to talk to okay that our accessible dot so they know they're talking to about
we are not trying to pass a two pass into i
i would argue that users should know that they're talking to a box and so
making the lasso human like as to what the users may not be such a
thing to do
pretty much the systems
i know that you know in some
for some people channel after a little controversy all this is not a chat but
i think there really are applications for this
for example you could imagine in language learning having a conversational agent that can converse
was which is a good way to practise language tutoring systems a good way to
interact with learning about information at their own case with depending on your own interests
you know
are you we're using
a chapter information exploration interactive health information recommandations and just to give you a nice
you have how you can imagine that so when i come home i actually use
the my i'm not a power user but i actually use my
why alexi the and often times when i come home i want to listen to
the news well i'm at dinner well you can imagine if you could interact with
the you could tailor the news to the stuff that you're actually interested in
and that there is the notion of an exercise coach or your coach so we
end up teaching conversational a high course
screenrecorder a building on what we have learned the teams of students to read and
there was a great coaching a i system that
one of the student teams bill so a lot of actual applications like think this
technology can lead to and a lot of people are shown that interested in
okay so are you is that it's a conversational gateway timeline content so again when
you get when o
you might want to talk to a the in a rat we had your system
to learn about what's going on in the world
and in this particular case we're scraping client and the contact would be a new
source it is it could be video could be well actually no with all text
was so it's new sources the could be whether we're not using quarter of we
use a and b
we read from a red it's the discussion for that could be so all of
the stuff that's online you could interact with
so just to give you an example or even by
this is an actual dialog in all examples i'm gonna give you are actual examples
and exposing a lot of our system
so in the first case you have to start out with you says let's chat
that evokes a system because we're supposed to be anonymous in the competition everybody was
required just a this is the know what surprise social but and then added that
can just go on a chat and
you have to chat about topics you have just play games and chat about whether
so we are for or something
somebody will accept the
they will talk about that and try to leave the conversation for thirty eight somebody's
not saying too much
and so far in set with this case we're talking about movies and we might
talk about a director or there were we might go which you that sort of
like so that's how the dialogue going
in the beginning i'm showing here
a recognition error the person after house or a person says that alright reason we
can get that rat get that the answer responded that correctly is because actually we
have n best alternatives
and so we could do you get out in figure out based on probabilities and
based on the actual context responding to house k at the present actually said
okay
so i want to highlight why this and how this type of social but is
different from a virtual assistant that has much more research
well so that have i use that is a sort of conversational ai cyst
components and even if you're doing and to and
you're and i and would sort of rebuild we're often different stages that maybe training
and the and reduce different stages of the speech and language understanding
the dialogue management response generation but also every system is gonna have some sort of
backend application that you're interacting with
so and a virtual assistant
the speech and language understanding is constrained domain
can be and easier task you like task intends
oftentimes you're filling out forms a binding constraints to resolve with the person wants to
do on the social by the end
are more social
or information oriented i want information on this topic
so the entrance are a little bit your french
and in terms of understanding at the sentiment is gonna play a role
the dialogue management side on the virtual system you're trying to resolve ambiguities security and
options to figure out what's the best solution to this problem
and then executed task
and the roar would be timely completion of the task
a lot so i
you're actually trying to learn about the interests of the user
and the suggestions at least in our system but that's information oriented you one make
suggestions of things that might wanna hear about
and the reward is user satisfaction which is not so concrete
and that's very challenging
the backend per a virtual system with a and b is structured database are back
and is totally unstructured so we have data structure
and lastly because it's a constraint or maybe virtual assistant response generation is you are
then in our case
which is an open-domain because we could be presenting information on in there
okay so let me tell you a little bit about our system
and i'll give into a how we
the velocity cut it a little bit overview and then evaluate the system
so
again this is the new problem when we started we had no experience with the
lexus skills we didn't have our own dialogue system and
using their tools
well as it really a good solution because it will for designing for speech or
fine
and that's not what we were doing we're actually doing conversation
as opposed to you know the form filling task oriented things that people have designed
apps
so
that was a little hard
and find that
there was no data no people often chair challenge is of that no amazon had
data they just they should have given that you know there was no data amazon
did not have data they had interaction straight and transactional interaction like such a kitchen
timer
you know plane using
they did not have conversations
this was one of the reasons i'm sure this part of the competition
and i
after
the performance or so getting the data from other teams in the recognizer recognition error
rate went down a according to them in a paper three percent
so i really didn't have the data
so it's a new it was unusual you're a new probable and what that means
was there's no existing degraded entering so we started out thinking that's what we would
do when we started out with do we present a sequence modeling with whiskers it
doesn't work
because it's all data
so we have read a yes in terms of starting from scratch
i think
because we're starting from scratch our system was you see that in
so that data that we collected in the beginning
you know was good retrain your recognizer
what was not so good morning how to improve our system
so this is all the say we're at the beginning the system wasn't so good
it of all it had to well okay so that setting the state probably to
the system design
alright so we when we first started building a system we first started getting data
we realised have that it was we side effect okay what we wanna think about
in terms of designing this just
so
i think that people what makes someone a good conversationalist
so you know to a perceptron and you looking for people to talk to you
generally want to talk to somebody has something interesting to say
okay
and how we also want to talk to somebody listening to you and
joint we are interested what
you have set
okay
the principle seem reasonable to apply to a social but and in fact i think
they really work for us your some examples
so
we saw that users would react positively children something you will tell you later how
we have got that information so for example around christmas time
a people what like to talk about christmas and we in calling our content have
undefined
this little tidbit space accent beer ingredients to the international space and station just in
time for christmas and a lot of people that was kind of interest and they
like that piece of information data and also like sort of
cool size of our a lot of the users are turkeys and so they like
the fact that babies as you are ten months get is that how much someone
values a particular goal
by observing how hard they are willing to work to achieve that
i interesting people that was interesting and like that
they do not like all news
so that we had a fixed that problem really early on we tell me something
that's two years old that gave us better use
the also didn't like unpleasantly then it you know it turns out there's a lot
of bad news in terms of current events i mean that if you're scraping you
will get plane accidents where people die and things like that
so we started hearing or and you are visiting us that reactions
but filtering is really hard problem
so we can filter for people dying but we are a piece of news that
people really didn't like was something about cutting the dog's head off so that's really
unpleasant we wanna with that
so another thing that we want to try to do show interest in what the
user says of course they're gonna lose interest if you're not
if you get too much stuff they
that you don't want to talk about they wanna get acknowledgement
something that's really working in these conversations they need to get encouragement to express their
opinions does not be used to this
so we ask questions like have seen superman
it's layer
which part did you like best
so that's important part of the dialogue
and fortunately to ask questions you need a little bit of knowledge of the work
so you can ask seven standard questions about movies but once the domain gets brighter
we might ask questions like this article mentioned google have you heard of
yes
i generated this happened to us in the demo
unit we we're doing this averaged ml so in this case you know everybody last
but sometimes you know what are the actual uses a gets annoying
alright so this leads to our design philosophy of just summarise briefly
we're content driven and user central
so we had to do daily and i need to keep are
are information price
so we had a large and dynamic content collect collection and represent with the knowledge
graph
and dialogue manager that promotes popular content and diverse sources
or the user centred side we had language understanding that incorporates
sentiment analysis
we try to learn a user personality in the world around topic changes and tracking
j engagement and on the language durations so i
we tried to use prosody appropriate grounding
so
this is the system and i'm not gonna tell you everything i'm just giving you
to the lecture but you can see is a language understanding component dialogue management component
language generation there's this back and where we're doing content management
we're using and
and question answering system that
in this are provided
we're using not expert we're using eight of us for
some text analysis
so that's a big picture there's lots of modules because we're at the beginning stages
were constantly swapping in and changing things
and enhancing things so it is a modular architecture to be able to about the
rapid development
so very quickly aren't each of the different components
natural language understanding is multidimensional
we're trying to capture different things some responses can be long and in capture both
questions and commands
we have to cut taxes topics that people are trying to talk about and the
user reactions
the dialogue manager is hierarchical l
so we have a master and minions and the master is trying to control the
overall conversation negotiate and right topics to talk about
thinking about coherence of topics
engagement of the user and of course it's important to the since work on trent
content driven
two are considered content availability you don't want to suggest talking about something that you
don't have anything to say about it
the minutes it'll are focused things
for related to social aspects of the conversation and different types of news sources "'cause"
different types of news sources
or information sources
come with different types of
metadata an extra information so with movies we have relations between you know actors and
movies well for a general news source we just have the news and the metadata
about the top
this is
back to the example it you before
and in this example there's stages of negotiation and that would be handled by the
master and
different types of information sources that were jumping around the n
that are handled by the different
go so are different many skills so the movie is one skill
we great from a celebrated that skulls channel is
and so that the last hole
those often are willie
and then we also sh great from another source it's giving us a and that's
we're that job you're between skills
and the language understanding so i
basically we get dialogue acts
for
the dialogue manager and we get information that's to be presented from the dialogue manager
and the response generation is gonna take those internet into the actual texture got it
you're gonna say that includes a brace generation but also prosody adjustment
the tricky thing for the so for the things use a lot you just the
prosody in the speech synthesis
so we have no control over audio but we do you have control
i'm using s m l
so you can
make your
i'm like enthusiastic
of which you have to do with the prosody instead of having the above three
d
intonation
by for the is that we present in
news we actually just read as it is we rebuilt or it to get things
that are covered more conversational
but we're
but that's text
pretty domain and that's really hard to control prosody for
actually we also do some filtering in the response generation which will see later
content management has this end we crawl online content
we have to filter inappropriate and depressing content
then we index to index to using some language some parsing and entity detection
we use metadata that we get from the source
for topic information but also use popularity metadata
and then we
good at all into a big knowledge graph our knowledge graph and eighty thousand entries
and three thousand topics so in and you can have multiple topics
so here's a idea
so we would take for example over in
e upper left inside
are a bunch of news article or
bits of content that mention ut austin over here it is a bunch of things
that men mentioned google
et cetera
okay so the system is evaluated
by what amazon decided and basically that was really one to five user ratings that
was the most important thing and then in terms of the final that there is
that i it it's duration the ultimate goal
if we had made it to twenty minutes
with all the judges then we the team would've gotten a million dollars
so we actually did really well
i didn't expect us to get your five minutes
so ten minutes was pretty good
it's a hard it's really hard problem
but the interesting thing is the other judges that we're not so all of the
development was the amazon users
but they are three people for interactive interactors and three people for judges
where for the finals and they were people who were motivated to improve the system
people who were like news reporters you're conversational it is
and so the motivated conversationalist
actually last a lot longer than the average amazon user however there are more critical
so the average amazon user divas higher score so that's basically how it works
so what we
i actually pretty balanced and
is the average the amazon users
but the rating is at the end of the conversation
you have a huge amount of variance
and some of them
declines rate is actually more than half of them inclined to rate the system
so the ratings are expensive noisy and sparse
and i haven't that
you can have you know we're not occur between the states we get word sense
this then you in a weird sense ambiguities can lead you to do something that's
off topic
and so you can have guide conversations you can get is i can get that
depressing news you can have sections of the conversation that are working well
and sections that don't work so well
so you're or a score
is not a equally representing all parts of the conversation
and so in order to actually use that overall score
to meaningfully do design
we have taken a and then to the fact that users give us more information
they actually accept or reject topics that we propose
they proposed topics
and the reaction to the content is important
so what we actually do
it's we take the conversation level recognition and we projected back to dialogue segment we
can segment just because we know the topics from the system's perspective
and we project that using the information of user engagement
so you could be projected that non-uniformly
and once we have those segment level estimated ratings
then we can aggregate across conversations for example we can aggregate across topic we can
add aggregate cross specific content
or we can apply across eventually accurately aggregated cost use it right
so this is how we could figure out a this is the content a lot
of people like
this is a constant a lot of people don't work so that's basically it
so what i'm a bunch of the user's task just some kind regarding constraints
we could not you
and i think we have a audio side
so speech recognition
all we got is text we get an audio for privacy reasons
asr is imperfect
we don't get any audio so we don't get
pauses we don't have sentence segmentation that's been changed in the version but we didn't
have that
we don't have intonation so there's a lot of things that we can is
detect
and this is it we can do u s and all but that's all we
can do
so there are some constraints so that just to say
a lot of the errors are false alarm errors are all gonna show you have
any examples you can appreciate
okay so i'm just several conversations
so what i wanna say here
is used some observations and then talk about personal implications
and then all the three of these i can talk about the user modeling
so
there are
for dinner points and wanna make for users have different interests
they may have opinions on a different opinion on the thing is a
and use were example in the us
news about from
little is the whole or opposite reactions from users
they have different senses of humour
some people like our jobs and some people don't
there is they have different interaction styles different well and they're different ages isn't of
family so just a you example how this impacts the system
one of the things that we found
was people like to talk about vampires for some reason
so this was the piece of information that a presented a lot to people and
that
basically says did you know that relation vampires are tiny monsters that perot into people's
heads
and for some the talk about that
now we don't control the prosody on this because this is general content so it's
basically read prosody
and so when people are listening to this
if there actually listening they are often amused as a kind of an so but
sometimes
they think it's
a bad
okay so they're not of
or
they what they had
because this is didn't make sense to them
i times you can tell they're not really listen
so far well
citrus
but last three
there are a user community is a little more complicated
there are also the callipers
and so this would and
resulting in topic changes for those people like that
they are different interaction styles so this is one user
talking about vampires all kind of this was i useful user i'll come back to
this for other examples
and then we know that she user which is actually more frequent category
where a lot of the answers when one word so
this is important to appreciate that it affects
language unit
so
the a type of user actually is a lot harder for language understanding
because
there are there is more recognition errors
we're not
you know it's harder to get intent
this type of user actually is also hard for language understanding because
we don't have prosody
so what it's saying no in a way that
so it
if i ask a question
do you want to hear more about this and the person says no
that means they do not want to hear more about this if you a request
if you say something and there are a lot and pairs as know that
if you wanna hear more about this
and so it's important because we don't have prosody
that we use state dependent dialogue and language understanding but even that doesn't always got
it
so this is my argument for
right industry we give us project
okay so they have different calls to the information seeking goal
the information some people just generally want to know more others ask specific questions others
is really hard questions like why
i there is
a like maybe empire percent
well i'll laugh are on and start asking a relevant question to the topic of
vampires
but not
and the user to call or the that are that we were talking about is
it really true that are like tv vampires and then there is a speech recognition
here
opinions sharing
some people would like to spark a lot like to also share their opinions that
actually not so hard to deal with because you can you that is you might
in a party and not in huh
and then there's other people who want to get to know each other they want
to find out
why a lexus favourite x is tell us about their favourite axes and so those
are different levels you have to accommodate
a we also have an adversarial user is we share suppose three family friendly
if we do things that are not in we currently we got taken offline as
this is really in the field
for us
did not use everything
so we did not wanna get taken offline
so we work really hard and we did many times we worked really hard though
to build content filters in the come up with strategies to handle adversarial users
so in this particular case we're not supposed to talk about anything
related to pornography or sacks or anything like that
so you just but a lot of users so you just have to have a
strategy for dealing with that so
in this case
we just tell people are much as well
when they have a sense of language one time we got taken offline because if
you didn't understand what they said sometimes a good strategies to repeat what they set
and that
and so what we were doing this we were filtering all the concept we were
presenting but we forgot to filter what the people's
so our solution there was to take the babble heard and replace it with random
funny words is one of my students came up with this that i would never
i thought it was a really stupid idea but it actually people laugh so it
people really liked
so we say things like unicorn i imagine you record or it's actually more funny
if it's in the middle of a conversation and its you know butterfly open your
whatever it is
and change then we change the subject and then there's a lot of people who
manage and control it just have a strategy
which
i don't understand or whatever okay
the last problem is working with children and you have a lot of children and
problem one in working with children it is a that speech recognition just doesn't work
as well for young children everybody knows that
it companies
have included some stuff to get it h o ring age found in him to
lower things but really young children it doesn't work as well
i'm quite sure i'm looking at the and bass
this is a kid talking about the pet hamster but other than that it's really
hard to figure out what they were talking about in this case asking them to
repeat
is not gonna solve the problem it's better to just change the topic
they think it's content filtering
so when you're talking to a kid at christmas time
a lot of times in the us a lot of people want to talk about
class
fortunately a lot of the contents
that we were scraping from was or at all
and
we take it also i because we set sail a class was a lot i
another concept sequences that
i we were not only people this i two is that would so ever saw
that what points that start talking about the other class
so results actually
okay so we have a user personality not well
it's based on the fly factor model that's based on the we ask questions based
on this two questions but the real world readable more conversational
weaker ones and things questions that we don't actually used to make it
more engaging the people but we can ask human because this is the and the
interaction where we're supposed talk about topics of people to want to just
you know do you with all sorry
so the data we have is very noisy and impoverished we're not asking that many
questions
buy tickets is it doesn't give us some information so what we can see
is that personality for the things that we explored
thus correlate certain types of personality correlates with higher user ratings
so people who are extroverted
agreeable
or in haven't you give us high ratings okay sort of make sense
i think that's interesting is there is a statistically significant correlation
we owe personality traits and some of the topics that they like
you know not for the topics a lot of people use
not everything but there is system
this correlation
for certain types like kindergarten actually hurts
there was the that data seem to be pretty good some extra perks like recent
fashion introvert like a i-th routing task
if you are and imaginative you like
and you like things like a i've time travel anyway and
low conscientious now as was explained as you know you don't like to in your
home or a and those work with those people like pokemon be in one craft
so that data actually sort of it sounds
okay so just summary here
the implications are that
age and dialect
that the implications are the user characteristics okay every single component of the system
that age trial are dialect verbosity a pack language understanding your interests the fact that
dialogue management and the types of if you're you talk a lot more errors that
affects the dialogue management strategy
you're interested that content management
you're h does because of how you filtering is
as we begin to user modeling we wanna multidimensional content
in that so we can get ratings the different user trials
and lastly the phrasing that we use the generation
if we have more information about the user
should be adjusted based on
so a user modeling
this is really early work
so this is a preliminary so nothing public but i thought it would be under
talk about in this audience
so i'm gonna talk a little bit about why we care for content nanking and
the user but in future embedding models
so
while we wanted to the task that were interested in and is given a particular
the contents
a project whether the user is going to engage positively or negatively
or slowly with that content
and so the time span is gonna be characterized in terms of the information source
topic entities
at some point later sentence and valence but we haven't done that yet
the user engagement is characterized in terms of what topics as the user suggest
what topics
what does the user accept or reject
positive or negative sentiment in reaction to the content but also a positive or negative
sentiment in reaction to the ball
because that reflects an overall being unhappy with content but maybe not a specific font
probably generally
so the types of features were using
include both some user independent stuff that's like the bias term
so relatedness the current topic and general popularity in dialogues
but then the user specific features for mapping these different types of measures of engagement
into a few additional features
and then the work trying to use the light cues
the user to capture things like age personality
not the issue here is
we have very little data so we don't know
we have to treat each conversation independently conversations we know that no the conversation came
from the same device
but these devices are used by families and oftentimes use more than one person so
you cannot assume that
the person is the same problem
conversation to conversation
for specific device
in the future you can still have that information but this is we have to
use only a conversation
so that it is very sparse
so you have to learn from other users
so
just this is just a motivational slide
this is just say that the user is really important so when we're predicting the
final rating of the conversation if we consider topic factors
i didn't factors and user factor so topic factors are what the topics or the
topic coherent stuff like that
who was it's just by the agent that there is there are things that the
agent's is
how they say that and then the user factors are user engagement and
the robot's them and things like that
user factors are alone
you better performance than everything together
in predicting the final conversation level so the user is really of work
okay so
i do not mention neural networks except to say that we didn't you and training
so i'm gonna now mentioned in that it doesn't mean that in fact are used
because everything has to be passed et cetera but we are using them in terms
of finding user embeddings
so the first thing we did was actually not be used a neural network
well as latent dirichlet allocation
a which is a standard way to do topic modeling that works modeling for any
task
so what we're thinking about what we think about this is each user is a
bag of words
and that would be a document like a documents
and we're gonna come up where represent lda the clusters the different what about topics
of lda would be user type so unsupervised learning user types
so we just had to just do let's just use hand what topics or clusters
because we don't think there's that many different user types and this would be undercut
somewhat interpretable
and that if you look at the most frequent words
you the following phenomena
people who like interact with certain types of things the people like to know what's
one particular cluster people talk about music was another particular cluster
and the personality quiz
like this
shows that another cluster
interesting and
a lot interest in the let's the
shows that
and another cluster
interesting you know be oriented so with the legs what your name what's your favourite
but analysis self oriented person i think i am
there's people who are generally positive
a whole one interesting
and there's people who are interested immediately
so that i l
it's so first of all you play traffic the lda in order to get the
interesting interpretable cluster you have clusters you have to do some frames
you have dropped frequent words it turns out i really that we needed to keep
yes and no in there is a positive people and negative people
but because you get yes no a questions are just so i'm gonna get those
in there
that you have to for them out
so uniqueness to make it work and there is you know there's this class and
i have fundamentally in a perplexity of that's what we're doing
without is that the right objective take your right users
well trained on another a problem we played around the different objective
to learn user embeddings and this was user we identification this is also unsupervised
and then it is
you're gonna take a bunch of sentences from user and bunch of other senses orchards
from the same user
and try to learn embeddings that make those things from the same user closer together
and things
to a user
farther apart
okay so we have
distance to sell
we want to minimize
and distance to others we're gonna maximizes it's a minus sign
so when somebody's talking about tasks and they keep talking about task
we want those to be close
and when they talk about something totally different that's gonna be five away
that is another way of dealing with drawing up things
so
if we this work was actually done related
and we have this problem where we're gonna let cid each and what are you
serious and i say finally somebody else like that
you from their tweets
so using this unsupervised learning which we call reality it turns out and you're picking
in from forty one person in forty three thousand random people we evaluated with mean
reciprocal rank
so basically the mean rain
with our best
just which was initialized with worked about
and then use the identification is twelve that well at a forty three thousand is
pretty good
lda is a five hundred
so this type of user adding i think is very promising
very for dealing with learning about user types
okay so how do we evaluate them that's with this task of embedding channel project
engagement
and a conversation level ratings x
okay so in summary
the a unit summarize the sounding board stuff and then the user stuff so basically
the social by
as a conversational gateway
involves not
accomplishing tasks
i in hearing about helping the user of all the goals and collaborating to learn
interest
and the user what the user is doing is learning new fast
exploring information ensuring opinions
so that the end of your conversational a system
the radical system components are basically related to the user into the common to tracking
the user intents
and engagement
but may also managing and evolving collection of contents
with you can think about a social chat knowledge
and as i said in the beginning
million conversations with real users and this new form of conversational it i
least menu problems of this is just the tip of the ester
okay so the sociological asr that user group that information exploration
re a user variation so
you know i'm sure that either conversationally i get a lot of user variation but
it
but with a lot
understanding the user involves no
not just what they said you can send that but also they are and lastly
that use amount has implications for all components of the dialogue system
and for evaluation
so lots of open issues this is the typical shape of the iceberg
a user and reward functions dialogue policy learning
user response generation and so we have a context where language modeling that use the
user model as an input
and rate for user simulators of those times the things you could do we haven't
started out and you have this
but well as dependent the word function we have anyway it so it's a at
this platform for language processing research and that i will still
so that is stuff that i know best about and they're definitely where other people
you
participate who are interested in user modelling
so wouldn't be so the system we feel that had no user modeling this is
the coast
value that this is close to competition
using our data
okay
so we had no user modelling and they're
we didn't have a the detection of engagement and the personality stuff
and we did that actually started with you we use personality to predict topics
so we had a little bit
but not the not about it
so there were other people interested in user modelling i don't know specifically what it
what they did
the presentations that were so i know more about the three finalists because of their
presentations
i
don't think there was
much using modelling and in that
so
so i would say i don't know as much
more of that was so we did less the
trying to use reinforcement learning and that sort of stuff because
we just that we don't have the day
so the people to more of
that approach so i think there is a difference
in terms of the silence of the approaches
i you know when and the thing is
everything is important
so you know button most important
you know that i think the user modeling definitely the user
centric stuff so that the thing is in terms of being user centred we will
change topics quickly
and to if things were going style
so i think that helped us i think the prosody sensitive generation process
but i think most importantly having lots of topic
contents
interesting content
helpless
but you know the other stuff that people did it probably would have helped us
if we had incorporated it just that was not always some time
so it's hard to compare what was more important
across teams
exactly and that was indeed the strategy
i
i agree and so we don't do very often
so what we did is we had
a series of strategies
for when we didn't understand what the person said
that was one of them
we also have the strategy of asking
for repetition
we also have the strategy of saying we don't understand
so there was
i think there is at least five different strategies
we would cycle between with some randomness a but also some use a the sentiments
of that percent to figure out
the detected sentiment to figure out
which to prioritise
tobias
between the different strategies so our way of dealing with it is to sample between
different strategies
there were actually at least one t maybe more than one team that actually used
a lighter
and incorporated adam's in the same way isn't it wasn't like our are many skills
with a little bit like harmony still so they allow to take a shall use
the lights the slot eliza into the conversation
we did not do that
we just had that as one
particular strategy our own implementation of it
very few
but people do assets order to take
and
the ask questions that are a little bit more difficult so that's the that's like
the why question
there you people do that that's really hard you don't have a we don't have
a solution for that right now
more or and
they'll ask them or slightly more specific question
and we can come up with not a great response the least
better than i don't know
the thing i don't know when you say what did you find interesting is a
it can valid but not great response
wonderful question that you are asking that question they are not and because we don't
have the prosody we can't tell
and so a unit at different version of this talk i those examples and it's
very frustrating
you will know you would have a i mean prosody analysis is not perfect right
but you would have a much better ideas so you could it would be easier
to get sarcasm
no request
so are now natural language generation is not at all sophisticated
that's an area where i would definitely want to improve it's just
in my own mind it's not the highest priority so when we were generating the
content
about you know the news of the information or whatever it was
basically you take what we got from read it and we that with minimal transformations
so there there's transformations to make it
shorter
there's transformations to there are some simple things
to make it a little bit more suited to a conversation
all but mostly things that are really not suited to conversation we just for well
so that strictly just
the wrappers around the
are generated but that's fairly straightforward
so this is an area
that
we could do a whole lot better
so the knowledge crowd
basically provides links
we we've it's
man it's on you want the details the actual technical details
they use dynamo db on the amazon cloud stuff and i can point you to
my grad student how we do that it it's really important because we have to
handle lots of conversations when we're alive we have to handle conversations all over the
country
so everything had to be super efficient
within a conversation you have to respond quickly so everything has to be super efficient
so the what the knowledge graph allows you do years
say from this point
if i want to stay on topic
or keep with related topics
this is
the region of the set of things that i could go to and that we
have a content ranking