BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Semantically Conditioned LSTM-based Natural Language Generation fo
 r Spoken Dialogue Systems  - Shawn T-H. Wen\, University of Cambridge
DTSTART:20150911T113000Z
DTEND:20150911T120000Z
UID:TALK60617@talks.cam.ac.uk
CONTACT:Tamara Polajnar
DESCRIPTION:Natural language generation (NLG) is a critical component of s
 poken dialogue and it has a significant impact both on usability and perce
 ived quality. Most NLG systems in common use employ rules and heuristics a
 nd tend to generate rigid and stylised responses without the natural varia
 tion of human language. They are also not easily scaled to systems coverin
 g multiple domains and languages.\nThis paper presents a statistical langu
 age generator based on a semantically controlled Long Short-term Memory (L
 STM) structure. The LSTM generator can learn from unaligned data by jointl
 y optimising sentence planning and surface realisation using a simple cros
 s entropy training criterion\, and language variation can be easily achiev
 ed by sampling from output candidates. With fewer heuristics\, an objectiv
 e evaluation in two differing test domains showed the proposed method impr
 oved performance compared to previous methods. Human judges scored the LST
 M system higher on informativeness and naturalness and overall preferred i
 t to the other systems. 
LOCATION:FW26\, Computer Laboratory
END:VEVENT
END:VCALENDAR