TU Berlin

Quality and Usability Lab2018_09_03_Griess

Inhalt des Dokuments

zur Navigation

Training and evaluating a policy of a statistical speech dialogue manager

LOCATION:  TEL, Auditorium 3 (20th floor), Ernst-Reuter-Platz 7, 10587 Berlin

Date/Time: 03.09.2018, 14:15-14:45   

SPEAKER: Katharina Griess (TUB)


The multi-domain Statistical Spoken Dialogue System toolkit PyDial provides statistical approaches implemented for all dialogue system modules and is used for scientific purposes. With the toolkit, policies can be trained to find an optimal action which the system executes. This thesis gives an overview of Spoken Dialogue Systems and Statistical Spoken Dialogue Systems. The main focus lies on training policies with different settings and with the inbuilt user simulator to test the performance of the toolkit. Furthermore, two of the trained policies are chosen for a user experiment to perform an additional performance test and see how users perceive the system. The two policies provided good results with the user simulator, even though compared with the user experiment the good performance could not be reproduced. The hypothesis that the policy with a higher success rate would be rated higher in the user experiment, could not be proven.



Schnellnavigation zur Seite über Nummerneingabe