hi everybody and then attention from cmu this the joint work done with my colleagues
and one vice the mixing and so they might talk is going to talk about
how to making a charity slot filling systems using the encoder decoder models
so the slot filling system is the focus of this talk and it has been
a device critically in recent years because we have better models more data
but the is the one limitation of a just a two-pass approach is most of
the proposed approach is sometimes of dialogue system is the with domain specific and it
is restricted to one type of dialogue act states and there is not very domain-dependent
so because of that we have difficulty of extending existing operating system
when you don't lengths on your skills
and a second difficulties sometime in a way deploy a system in real life this
always out-of-domain a request and it's gonna be very limited to handling those out-of-domain request
so our goal is we want to develop a domain general system that is not
restricted to a the define predefine acts of state even this task
and the encoder-decoder model things a promising a choice because it's a very powerful and
its prosody model and it has been successfully applied to translation and the open domain
training
so the idea is quite simple we have a encoder that encodes the dialogue history
and of generating a dialog state representation and we have a decoder that you called
the system responds spoken by token
i