i speaker is
second
so
who will talk about interaction into that i still not only learning in i
hello everyone
and we only one
hello everyone
so mining station with an hour and i'm going to present the work like noninteractive
learning of actual nondigit model
and is a joint work lead to be you should leave one and the ensemble
so a list also human conversation is always knowledge-driven and human off and like most
the time ground facts to generate meaningful and intelligent responses
so do batteries on actually in recent years many words once a mini course has
inspired to walk on the particular area of knowledge from the decomposition modeling
so here to prominent works in this topic so one the first work is actually
upper extends the sickness to seek this model where the a response generation is conditioned
on the
pointing squandered twenty seven past selected from a knowledge base apart from the context to
make their lander from the conversation history
and the sake in the one is about the response a ranking model for right
of its dialogue systems
where they actually used a fact encoder and
combine it with their shared utterance and whistle response in order to generate some prediction
scores
so that usually these existing works is like the work so it fix knowledge bases
the knowledge bases so as we don't know like they are highly incomplete
and the knowledge is just one group over time during the conversation
so how an agent can learn new knowledge in the conversation processes with this two
major an approach that we can follow one is the passive learning approach that is
the learning by extracting information from we corpus or like past conversation history
and interactive learning approaches learning to interacting multiple dialogues for that is our focus in
this paper
so there's have an example here like the human law incentive original it in a
lifelong manner so for example user too large that spock home is that the capital
of sweden from user one and then leverages the knowledge in that now another station
to recall made user t v is the stockholm
so
this kind of knowledge learning happens it want to know what's where and with you
know multiuser environment setup
so based on this like idea actually we propose a new but i'd be more
lifelong interactive learning and inference
so the motivation is actually knowledge is not enough for the it in has to
learn the semantics continuously otherwise
in that it does not learn how to realise that knowledge in downstream applications
so the better team actually consist of three states and that actually the dependent on
each other so the first is formulate an executed very specific interests in to infer
a strategy that image processing and interactive actions
and then
execute after executing state easy they'd alarms the interaction but as that is gain some
knowledgeable
deciding what to ask and when to us
and then
back to our knowledge is actually a whereas in the current and future interest and
inference process
so in this work our focus is to develop an system that and lower new
knowledge effectively
from user
when the system is unable to answer the user's wh portion for example if the
user task like to annoy
which country boston is look at that
so than set may not be directly present in the knowledge base but the system
can ask some
after some supporting facts from the user to be tied and so
so it but with the facts that and we expect to both like hit relation
and they're like boston is the hidden to do you listen is look at england
three and with is that in entity
and here we focus on developing that y in g in this study
a building a full fledged dialogue system that also lines
tuning on portions and conversation like semantic question other stuff is like which task
and i
we leave it in the future work
so the challenges in developing this kind of this database you need
not knowledge acquisition and learning model that should
be a train that should be learned simultaneously and set it may not be present
in the kb but can be inferred so the
it's in has a lot of somebody's any probabilities
and
by aligning could easily over existing techniques also allows the semantics of the knowledge that
and replace models
conversational applications
so at it usually is like a group in one line intelligence ripples query relation
and in to be maybe on one because considering the open and in is all
the conversation new concepts and relations that you know that i
and then set or may not be present they can be so the it's and
has to learn how to reject
that is
see i don't know more and more confident enough
so the formally the problem can be formulated like this would be even enables the
query relation so we build that structured query
so the goal is two fold against fading the user query or in this thing
that weighted we on it can answer
you this direct answer is believed not exist in the kb
and landing or a putting some knowledge in the process that and we live it
is in the future like interaction and are also inference data
and we probably thing we should pay for creating a local queries basically where all
tickle one is at the destination unknown to the query
and we'll been
one queries is either of them can be on
so far that we propose an inching for one clean one for an interactive an
additional that in short silk
so the main idea of this system is tree where we use a to convert
the open what we're using the proposed one creating by equating some supporting facts
the user and so that it actually with during the conversation process is called supporting
and then inferring that variance or by realizing the were supporting facts an existing fess
in the kb
so the inference process is basically like this so
actually use this each in td that is present in the kb to form we
can do to triple
and then a it discourse each triple and chooses the indeed be with maximum score
as an so the creating like the confidence a more logical reasoning process
so it's can see stop like three conformance by knowledge base that stores the knowledge
that need to get started it want to see what time
and then we have interaction model that executives the skip one that is decides
when to ask and what to ask basically
unlike inference model that performs the inference
that or it can basically not the semantics
the expert knowledge
so less an example here sort on the left hand side we show the interfacing
all the silk system with the variational components
and on the right-hand side we show some things that example interaction all where a
system lance some something fast
so for example if user asking what country boston is located then submits a i
don't know why look at it can't remains can provide mean
an example so it is asking for a clue for that relation to that of
the semantics of the relation
and also you the entreaties unknown suppose it does not know
what is faster
it can ask for some additional fast
from the user
a two norm or able to the indian the semantics before it goes
for the inference
so as you can reduce it is not a teacher student learning setup because
the
like user does user may not know the actual and several so he can provide
the exact evaluation but again it was found related facts from weeks
the system has to learn
and be like that so
so the inference model is
like constructed based on the knowledge base completion problem
so basic idea is to infer the new x one existing fast in the kb
so the k v c actually explore as well as an option
so that is like so that so it is not directly applicable to the contour
knowledge learning in conversation
because it can only handle queries with no one relation and in today's and also
does not do any rejection
so where we actually remove social as well as i'm probably and follows the action
in the inference
so we use then you'll only just completion models with signaling so you need models
can be used
to solve this problem was specially adopt
the model
the idea of the model is like this so we actually convert each in detail
in section and
the like introduced in the relation in a candidate realistically into a like one hot
encoding vectors and then we'll on the remaining okay "'cause" for
indeed en relation and then rescore the triple based on some by linear scoring function
as we can should see the we going
and so does the overall model and we basically training with minimizing the maximizing the
ranking loss
so it actually maximize the positive triples and minimize the
a score of production rules
this because some option where here we use
so how to do rejection in the kb inference so
the same propose that a short above are basically that's forced
i in p d and relations basically prediction threshold and is continuously updated over time
so is computed like this so the facial for an entity or relation is basically
that original means course of triple involving positive integers and mean score or triple since
looking at it is
that belongs to a particular like hal initially very tuples testing probably
that part indeed you're relation so this basically validation body doubles it is basically a
like you just any minus top also q is that we at trying to be
a hit or maybe
and fifty plus is that says set for that gradient minuses the negative instances
for the practically
so
deciding the interaction
step is to be sufficiently so
as you all know like silk aspect we're supporting fast
tool and the embedding or entity and relation but use that can only rule high
very few sub for supporting perspire session which may not be subfunction line would be
meetings
and also a reading too many supporting taxes no into the user and is also
and i guess at even the model has already large
to do resenting only a given entity and relation
apart from that we also need submachine legal validation dataset to lower the thresholds
so
the min strategies ask for supporting facts for
no one integers and relation that is not for who is the system is not
confident enough because done on one for which you have to act with some first
one on them
so for which we propose the performance bar for those for the power from in
statistics of the
inference module basically over time
so where no
but and b are basically d north average a moderate so that is the mean
reciprocal rank achieved way they are the interaction inference model
and that is evaluated on the validation credited to sit at a sample from a
knowledge base
as at each dialogue station like this
still basically detects the second order cone like percent of the query and it is
a relation based on the immoral score so basically these basically shows you but if
you didn't in tdoa relation set and relations that for the nist dialogue session
so that means the one what set of entities and relation the model is not
are following very well
and then ask user for supporting fast eve the query and at all relations will
also die the diffusion state or is on the second so that is the basics
that is
so putting it all together so we have three states so the knowledge acquisition step
where
the system interacts with the user quicker supporting facts
and then knowledge of it sounded fess where we store
the supporting facing the knowledge base and also modeled aim in terms of training and
validation triple
and then inference model and one for a bit fess where we in like sam
illicit of training data and in addition to make a data model and also a
bit apart from as and like special of course
and this actually happens bar in one session and the proposals on a lifelong when
in multiple in with interaction with multiple users
so the overall evaluation setup is like this so as to outsource this evaluation as
a long is very difficult to conduct
i
also very time consuming for this is a specific set of because it needs point
in was introduction to the user
and also is not necessary given we are like
deal with the structured query
which is basically an as i mentioned but so we don't have a proper dataset
collected
for
like real interactions and not only that like static that does it will not work
because it has to be evaluated in a steaming weight in want investment
so we created a simulated program basically that we
call similar to do that actually to effect with this and it has to
components the knowledge base
so for task setting portions that is asked way so that is to provide the
supporting fast
and the query dataset from utt issues query to the system
and we have to convert it to well known knowledge base dataset like wordnet and
now
and we created a large table store out of that and then split it into
that users knowledge base and the base kb also and equating dataset to for evaluation
so the be a scary can be regarded as the initial maybe when the system
is deployed and begins
the knowledge learning when interacting with the user
so the overall evaluation setup is
based on to face is the initial training phase where we
actually in and initiate a inference model
only on like norm query lutions an interface that is present in the base kb
that is the initial kb that the system has been deployed we
and then online training and evaluation face ready to interest with the simulated user and
actually a supporting facts in the process and access the query
and the did the
d q actually
consist of both on
queries involving with more than a known relations and entities
so these shows the overall statistics of the data after the conversants the details of
the details of the
creation process in this in the paper
so we have almost one heard relations that we mean and basically by randomly deleting
the triples in the original graph
and also this the statistics of the unknown in tedious and relations over the case
where this
it's clings would be
and is there is no existing works that apart from the same fast so we
basically did some evolution study here
so we propose two kinds of variance of the model one is that spatial variance
and the agenda-based on the
that doesn't sampling strategy
so the facial variances of four is introduce a change its articles in a only
need to take spatial
to learn the
as a prediction threshold
and then relation special variance and the mean fiddler those two and the max threshold
and in dataset sampling strategy we have
like the partition of where we actually the inference one and only we triples involving
gradient it is and of the setup we train we really nation triples
here is the
training with a relation and entities
so this overall far from a statistics all if it's of audience
that is mean thirty long so humorous question set that it to one and he
said that case
so he said that it one means
the how often the contact
the answer a like one p denotes a discarding fifty million features and install then
how often the true answer exist in top ten pretty
so as you can see the max
do you gbd a single e x is the overall performance
the car noises little bit lower basically how well
because it's very hard problem and you need some
chance far off knowledge or early in the processes that will work on currently
and specially for a
near basically it is but little or than word and because the label nears
cost is k often stop lot of sound field relational organisation means the relations for
reason the number of pupils as relays so it is not
is very hard to learn proper semantics for them
so if we consider the detection performance so we actually evaluated based on the probability
of predicting an alpha given difference so existing the kb
and probability of rejecting there is a given then so does not exist in q
v and we want
should be high basically
so as you can see the max
threshold btr
partition basically
five at all based on what was it is the indy d facial and relations
are scored one
i'm showing some order to study z prediction and rejection behaviour
to study the effect on use any direction so we actually did experiments a high
paying the number of clues and three d face activated by the system
per station over time and so if i want their one point one please so
if we used a number obviously fashionable forcing be this in general especially for word
right
that shows like usually fess are much more important than lose
clues that speakers each
because the relation maybe
be are in future also quite frequently but
the comparatively introduce
will be with the lowest frequency the gradient is if you has been austin past
and we also compare the horse and very desirable the for problems but a four
and the performance drops in that case that shows like
the landing
the apart from the statistics over time leading the knowledge any glazing in interaction
but
indeed close
so this shows the overall performance improvement over time given the model has made some
prediction
and as we can see
like in the streaming very to the set like from fifty percent hundred percent these
results of safety
the performance of rollie was spatially for open one where it is because in open
one
it always
a gathers o supporting facts
where is in overall performance also improves efficiently considering he's that the one
because it also lots of facts
for a known
in this annotation
for was it is not from
well
so to someone like three proposed a condom lost knowledge i need in for dialogue
systems
and evaluated on the simulated in
user setting basically
and also some promising results
and this kind of system can be useful for call information-seeking on position or a
common a system that
deals with real world fax occipital so it's getting any and i will always that
can be input in terms of triples and required is any basically
can be
maybe they can use this kind of system apart from that all is around the
conversation morning is another potential use case
and in apparently and also in future working on larry
the mean college london component of the system to believe that will start problem because
when this relation has made it is no more of triples
so it is often hard to learn the semantics properly so to some extent like
if we can get a summit knowledge
from similar relations that we have already learned in the basket again
most of the performance
and also jointly learning
of all the different components with the core lining engine
so for each other
finally collect and some datasets
to train
so time you like if you have any
we have tens or
one of two questions
i one from rows a fine for the talk
and i'm just wondering where you give an example are just write a phrase by
harvard university
you're likely to get from you like this
are there
and i just wondered given it's actually
when you do this learning using no ellipses indicate prove parser
as opposed to just partly comes
is definitely
the semantic parsing and knowledge in any also quite interrelated because one can like influenza
other in terms of alignment
so a semantic parsing is hard actually because as we can is likely say like
be easily us what differences solution also
like in the process and also considering the conversation history and i need a copy
that the use of using their dean about
so currently are working that process but i can be used a segment here
how to like is also an issue and how clean to get to
like both the models to wire missing
but we are so there's a challenging problem
what are working on
i think interesting talk so i'm really i think one of things are averaged thing
about this
work interaction module i know was very simple but design but maybe you can talk
about well your design decisions in what would you want as like a baseline for
that sort of interaction
okay so a fortnight interests no one will design like
a so we actually also be some preliminary study and was in archival bizarrely original
paper where we used a rule based like strategy to learn the
interaction behaviour so it is very specific to the application individually applying the system
so that so you proposed original like
model and then
some example scenario where we then design some simple interest and strategy
so but
it's is based on like specific application you need to think about
a currently we so apparently we actually be with something a finite state automata or
like some if else kind of dialogue flow kind of systems to design the intuition
strategy
and then that is parameterized you can think and we'll aren't that bad i mean
the to like adapted it collection
well hard i
from the knowledge we gain in the interaction process
so that is a kind of still far away to looking at that but obviously
the problem is much more complex and lot of things can be done in future
so like the end of you know that
but the for ten seconds
thank you let's speaker