Learning to Find Out What the User Wants: LSTM Dialog State Tracker

Speaker:
Lukáš Žilka
Abstract:
The core component of virtually any dialogue system is a dialog state tracker. Its purpose is to monitor dialog progress and provide a compact representation of its history in the form of a dialog state. Promising advances in the field of artificial neural networks suggest that using recurrent neural networks can substantially improve the capabilities of the state of the art dialog state tracking component in terms of training cost, accuracy and variety of the trackable information. I will introduce my dialog state tracker based on Long-Short Term Memory neural network, show how it is trained, and present some preliminary results. Also I will mention the main ideas of some of the techniques used in modern neural networks that I use.
Length:
00:45:10
Date:
23/02/2015
views:

Images:
Preview of IMG_0103.JPG
Image IMG_0103.JPG
Preview of IMG_0104.JPG
Image IMG_0104.JPG
Preview of IMG_0105.JPG
Image IMG_0105.JPG
Attachments: (video, slides, etc.)
63 MB
146 downloads
800 MB
146 downloads
296 MB
164 downloads
157 MB
151 downloads
115 MB
145 downloads