BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Paraphrastic Language Models / Structured SVMs for ASR - Andrew Li
 u\, Austin Zhang
DTSTART:20130201T130000Z
DTEND:20130201T140000Z
UID:TALK43182@talks.cam.ac.uk
CONTACT:Catherine Breslin
DESCRIPTION:PARAPHRASTIC LANGUAGE MODELS\, Andrew Liu\n\nIn natural langua
 ges multiple word sequences can represent the same underlying meaning. Onl
 y modelling the observed surface word sequence can result in poor context 
 coverage\, for example\, when using N-gram language models (LM). To handle
  this issue\, this paper presents a novel form of language model\, the par
 aphrastic LM. A phrase level paraphrase model that is statistically learne
 d from standard text data is used to generate paraphrase variants. LM prob
 abilities are then estimated by maximizing their marginal probability. Sig
 nificant error rate reductions of 0.5%-0.6% absolute were obtained over th
 e baseline N-gram LMs on two state-of-the-art recognition tasks for Englis
 h conversational telephone speech and Mandarin Chinese broadcast speech us
 ing a paraphrastic multi-level LM modelling both word and phrase sequences
 . When it is further combined with word and phrase level neural network LM
 s\, significant error rate reduction of 0.9% absolute (9% relative) and 0.
 5% absolute (5% relative) were obtained over the baseline N-gram and neura
 l network LMs respectively.\n\nSTRUCTURED SVMs FOR ASR\, Austin Zhang\n\nC
 ombining generative and discriminative models offers a flexible sequence c
 lassification framework.  This talk describes a structured support vector 
 machines (S-SVM) approach in this framework suitable for medium to large v
 ocabulary speech recognition.  An important aspect of S-SVMs is the form o
 f the joint feature spaces. Here\, generative models\, hidden Markov model
 s\, are used to obtain the features. To apply this form of combined genera
 tive and discriminative model to speech recognition tasks\, a number of is
 sues need to be addressed. First\, the features extracted are a function o
 f the segmentation of the utterance. A Viterbi-like scheme for obtaining t
 he "optimal" segmentation is described. Second\, we will show that S-SVMs 
 can be viewed as large margin trained log linear models.  Finally to speed
  up the training process\, a 1-slack algorithm\, caching competing hypothe
 ses and parallelization strategies will also be presented.
LOCATION:Department of Engineering - LR12
END:VEVENT
END:VCALENDAR