so thank very much for the introduction

i'm going to talk about our recent work

at cambridge university

current dialogue models

are limited in the complexity of dialogue structure as they allow for example relations between

objects

i'm that'll work

we

raquel

we present work as well

so it all a

deliberately break those limitations

and

we're not all of the current to have complex relations between a

complex data structures but you don't x

and even more so we i name system to address

relations between objects and we allow

users to address relations between objects

and even though this model has been

developed for some dp based task oriented spoken dialogue systems

it should be applicable for to all kinds of data systems by you want to

have a symbolic representation of the dialog state

at first the creek we can on and if you based

spoken dialogue systems

so what is here is the architecture of a spoken dialogue system

we have the input processing

modules which yields a semantic representation of the

of the input of the user

it is then used in the belief tracking

module to update the

the belief state which is then used by the dialogue policy to use the next

system action

which is then

transfer back

by the language generation and synthesis module

two response or you response the user

and the task of each tracker is then to update the

the distribution of all possible dialog state based on the current user input the previous

system action and the previously stated

and

you using force them to find optimal policy which maps the current belief state index

system action using a reward signal from the user

and the optimal policy is then defined as

a policy which maximizes the expected future calculated reward

this architecture has been used for

so you go domain dialogues a lot

and also for multi domain dialogues

and

using the multi-domain dialogue model in psychology

here's an example dialog

about finding a little interested in cambridge

so

you saw what does one soldier in cambridge use us a few questions about that

price of instances are aligned and the system table two

give a response

k is right cambridge is expensive fortune the sensor

he's is quite a also on the restaurant in cambridge

in the centre

system continues asking questions to figure out what the actual

the rest of the lexicon meta-file the rest from the user wants and then it

is okay chromacity trying to system

in the sense

and

in the multi-domain dialogue model

the user input first n is the main tracker which you can start okay what's

the current domain of the of the dialog the user input

this case hotel restaurant

and then they are

independent

part of the dialog pipeline which handles than the request buddies

and this model has several as a dialog state

stuff that all

the communication between those two domains

or at least doesn't allow easy

way of doing this type of communication

so for example the user would say trade it also effective referent

in the same area which is a relation relating to information about hotel this was

would not be something this type of

a lot of attention

so instead of

of modeling domain

we are looking at entities that

so

the same

example dialogue

in the beginning

you the dialogue was about

a not sect of type hotel

and then later about object of interest ones

and

also and then entities in our dialogue topic entities

and

that are also contains an additional ldc the relation which is modeled as a separate

entity connecting

both

a text holes in the restaurant

and this is

the main idea of the new dialogue model

so a and all these all these energies countryside centre conversation world

which kind of wine sent again

so i will

go in more detail on but let me tell you more details about those the

dialogue model

and then either so you present your prototype implementation a showcase that we focus on

relation modeling

if you can so that

this has dialogue model

is actually very effective in handling relation

between opt

okay but first we talk about the conversation into the dialogue model in more detail

what are optics tex mex and its use of the real world

we have an object in cambodia the extra restaurant

and all the objects in heaven the matching

and you know database

and based on these entries

actual objects in our conversational world

created

and

of course to notice do that the type of that

also has to match

the database specification

and a relation then connects to objects

no example the restaurant again

and a different definition of this relation can also be derived from

the type definitions of those objects

so in this example

we just compared compatible attributes like area areas

present in both

it would list of the

object type

so this is

this can be something they can connect on for this is part of the relation

to finish

of course you can elaborate is if you

she one

and

each conversation entity which can be adopted or relation maintains a user goal belief state

and the context eight

so whether it's an object or a simple all have their own

own state definition

and the user goal belief represents information

shared by the user bits

as a set of uncertainty whether and

so it's not as prepared probability distribution over all possible user goal states

as stated she's

similar to what we already know

from completely based dialogue modeling

in general

and the connect state represents information which is offered by the system

so for example the system says

well

on the houses a chinese restaurant in the sensor

and this would be something which would be in the context and so the system

knows that it has access information does not uncertainty associated to that

and the clicks it is

very important because

braces of refer to the information we just been shown by the system instead of

and so the information

rather than a computer opticals been found yet

and it is it of conversational world and their entities

allows them to the belief tracking on two levels

on the word level there you can also have a behavioural

a system behavior for creating and entity disambiguation

and then on the energy level they actually

for about what specific to gonna can forward

protesters this property stuff this relation actually have

okay so

do not confuse you anymore i have as a small example but we can go

through step by step

the sex on the t c them before so shouldn't be too difficult the beginning

the user says okay and looking for hosting cambridge whether the system can infer okay

that must be an object

of type or tell

and we don't have any information about it yet

then the other continues there's more information share

and that should be the price range expensive so we cannot tell the user goal

belief state and the system share

the most video what the proposed

and so what cambridge's an expensive one

so this is when the context it is updated

then a user says greater also need cheap restaurant in the scenario

which allows the dialogue system to infer that there's another object of time restaurant

and that's a relation between a topic

so this is when this could come to live if you will

the data continues personal information acquired about the

but it's fun and two

system proposes a computer us plan to the user

which is when the context and the of the object and the relation is updated

so

if you to the properties of this of this

you dialogue model

and compared to the most of them are not only can see on several expected

it is more or flexible and the

so for example objects can be can be addressed by the use of the system

is something

also in common in a way but this new model also allows

to have multiple object of the same time so you could have put model dialogue

where you have

two hotels

all

three best friends or whatever you think of

and were to be able to maintain a dialog states

correctly and you could also have relations between those object even

it also allows the type hierarchy so

the kids okay the rest when the venue and these things and this might have

given you wanted to policy modeling for example

and

well then let the calm assessment of the dialogue model also can do is modeled

relation

it's can be addressed by using the system

poses al

state it the relation to have their own state

this allows the system to actually talk about those relations and address time in a

system response

and also the relations remain specified

even if the context of the object

change or if there's

no object context is if i don't all

so if i if you think of the dialog the user but at the beginning

i'm looking for a whole to and the rest and in the same area

there is no connected all the scenario wouldn't

couldn't be derived any and

in any other way then

like doing in the like having like this separate entities modeling the rest

so this is

well i'm going to tell you about the

how about how the model

where

how to continue the how to cook something to treat example of how we have

implemented this model

and

for experiments on

relation learning something that this is very powerful tool to actually

do this

so one idea of formal beyond us to do when we do implement a portable

to focus on relation labeling and to reuse

the bound on the conditions you already have for belief imposing modeling

we have from the domain based models

so we have a fixed position world

maybe africa will one object of each time

and there's only one conversational taking the focus of attention

which is which helps the system to figure out how to continue the dialogue

and is similar to the to the

the main tracking problem

the user goal belief is more of the marginal distribution for slot

and this is

but also has the effect that you can reuse

step something on the conditions for quality modeling for example but also

but the only possibly relation you can

actually model is the constellation this type of

from you state

another to open questions

but dialogue management

how can be incorporated the conversational objects and relations into the decision making process

and how can we derive an input stage

to the dialogue policies from the conversational entities

and

we implemented a few reinforcement learning approach for that

because this is very

useful as it

have hierarchical structure and it kind of

subdivide the probably the several steps

in the beginning in the based on the focus of attention an object is basically

a

selected whether it's also respond and then there's a master policy a basis master policy

a decision is made but also talk about the art techniques to talk about a

relation place of that object

only after that the system

and the second bullet is able

about ready to the object

or to the relation

so this means that each conversation and he has the all and policy

and is also allows to say a relation policies

because this policy is the same value

come from this

so i talking about a the whole circle the output from the start talking about

the rest

and information we use

two

to

for the policy as input to do this

is on the entity level

for the relation we just look at the state of the relation to but for

you all checked we have to find a way to combine those things

which is necessary to database access basically

and the last odyssey it also we need a combination of the objects in the

relation

as you want to do computer solution

because

the object you please state of the object

could state something which is

it can stick to what the relation says based on what object to have a

specified

and to figure out those conflicts and to allow the system to handle this situation

you need to combine them

we consummate experiments

of finding out and the rest and in cambridge

for which we use the pilot system

and the conversational world and in the six and experiments

consisted of two objects multiple do not have restaurant

and

user goals contain the rage

always

between the slot area pricerange so

every user goal had this relation but was not always use

and there was an additional parameter are which specify the property that the usual you

use a similar to actually use that relation

so the you using lattices and has may consist of a lot to talk about

the value of all that's also various sensor all the talk about the relation area

of course that you "'cause" every restaurant

and if are outside only use a simulated user would say area

every hold because every restaurant

and we'll of the two experiments

comparing the

conversational entity dialogue model with them multi-domain dialogue model

in the first we were interested of the effect of r

does this happen in fact we actually you need to model relation can you can

be able to cope with this problem different way

for this at a fixed all of conversational czech only the second one was learned

and then we also ones of you got to see what the actual effect a

of this type of model on the on the learning behaviour in general

so we sell

experiment it's in experiment one about the unknown to different semantic error level zero percent

fifteen percent

you see the success rate

and the other relation probability are

and

for c represent you could clearly see that with the high our

the composition is the other one of the form to go if you significantly higher

and then the multi-domain dialogue model

and this is also true for the fifteen percent

case the goddess much lower here because of how the

our remote operations

it is artificial because assimilated but based on that could call some of the

problems you would usually have

in this situation

note interesting here is not only that the that the concession digit our model

performs better

for high are also that follow lost it does not

so the model domain data model is able to cope with

with use a talk about relation to a certain extent

touched by basically repeating the question

so you looking for both what prices are looking for

yes n is also a classifier for

t

if it does restate a question

this is the model can learn all the call this

i'm going to a certain extent and on in terms of success rate or re

well which

goes the we want

but i would assume that if you look good

user to sex anything's it might be different

and

we also looked at how many of the system actions actually address

and

how many i'd how mean dialogue system the system actually interest

and the relation and then we the system utterances in twenty four point five dialogues

for twenty four point five percent of the dialog directly address a relation which i

think it is very interesting

and for experiment to a we have a similar situation

it's not the strong because they took turns and effects of the domain

if the rest know how to most part of the first

first object of the dialogue was about then it

there was a relation present

so

i think we can say that the new model breaks up the limitations of current

our models

and

i think the not next logical step towards more complex dialogs

a lot also more complex data structures like relations between objects

and that could shown approach have implementation which we use as normal voice models that

it outperforms the baseline

and the future work

we further need to increase

the complexity of dialogue structures

so what's more complex relation

and

what ended if it had to be ways and just to name only two

so you can you can find implementation on the website

and also we have a new meetings from high to a so if you just

that if you to assign a thank you very much

okay

now we have time for questions that one

so you mean if it if there is like

just multiple talk about multiple objects

and you say the for the third object or whatever this have been the same

area

and then difficulties with this rule a strong gonna

okay

and i mean this depends on how the ontology is modeled

i mean in this example dialogues we

have you have a simple way of modeling the area which is not so please

press centre

i think

for doing these things you would have to

have a different ontology is gifted different

of locating these things

and then you could if you model that i love this situation

also so you so you mean about the relation refers to the to what the

user request of the actual without loss

okay

i mean it's setup currently

and they said to try to go back to that slide

i said relations of references a context a feature space is what the system proposed

so whatever specified for by user

is not necessary take into account when you look at the relation

the relation looks at what the system

also there's an entity whatever the values

it's basically means if you haven't talk about the area at all

then you said i want something the same area then you would be able to

refer to that while the

system offered to the user as a result it is basically what you're saying

you want something which relates to what the system of a global the users admin

in general terms

and i think this is why it it's important to have this as two separate

the possible state

okay so we have a question i think you where the first

i mean this

so the idea of this one is to estimate the model information just circumvention values

of theirs and frequency about that

then a

you have to add things to the model a question is whether the model can

that is actually able to model these things and allows the system to address these

type can be greedy

and this was the goal of this i mean

it cannot solve everything because they would that just have some

some

the reasoning going on maybe or whatever out to be derived information

or you could add some something to the language

to the understanding component or whatever to do this type of thing

if you and but if the system would be unsure about what the correct way

of getting the situation would be you put

the model could the policy learnt how to handle that and i and aspect of

whatever

okay let's move the source it's you

yes

yes

well from my point of view rubs of nh

referred to dialogue intentions of dialogue act so you something you want to alter relations

and i think you could

have those of the object type definitions if you will to extend to have like

what specific actions you could actual dollars

and

since you're doing however a multimodal and belief tracking or also policy decision you could

if more general we have to this on the on the upper level and then

a more fine grained way of doing this on the lower level which would allow

you to handle these

these things

i would be my files the correctly with the status of errors

it's hard to say i mean i think they are computationally is used to that

it's i think that's just what to basically referring to

and i think you go you can also line i think it depends on the

smart

way of modeling the

you stated so the probability distribution i mean is very simple way of doing the

probability distribution and you want to have

more complex weights maybe you won't have hierarchical one and you have to figure out

how to do these five

it doesn't

but strict the general idea of modeling the dialogue and that way

so

would still allow you to define relations would still allow you to talk about those

relations and i still think is a good idea to have those relations because this

is

from my point of view the only way that the system is also able to

talk about those relations

but where how far this thing going it's hard to say because i don't know

how the new model to look like

no

i mean one thing i would i

the next that i would go to also be that have a probability of what

different types of relation

and also on the entity level which would no longer to have also things like

on one

this all that

i one

not this and how to do these things and i think this is what you

could do if you have like with the distribution operations sort of just the values

that last

but also what the chase them differently

without looking to lead to a given more for even good on something that

okay thank you