BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Sentence-level Topic Models - Kris Cao (University of Cambridge)
DTSTART:20170602T110000Z
DTEND:20170602T120000Z
UID:TALK72613@talks.cam.ac.uk
CONTACT:Anita Verő
DESCRIPTION:We present two generative models of documents which generate w
 hole sentences from underlying topics. This relaxes the word exchangeabili
 ty assumption of traditional generative models of documents to sentence ex
 changeability\, and can hence capture inter-word dependencies that LDA mis
 ses. Despite the additional model complexity\, model training and inferenc
 e is still feasible using state-of-the-art approximate inference technique
 s. We show that both our proposed models achieve lower perplexities than a
  standard LDA topic model and a strong LSTM language model on held-out doc
 uments. We also manually inspect samples from the topics learnt\, and show
  that the topics both models learn are coherent. Finally\, we show that on
  a shallow document classification task\, LDA outperforms our models\, and
  analyse the reasons behind this.
LOCATION:FW26\, Computer Laboratory
END:VEVENT
END:VCALENDAR