You are here

Scalable Deep Language Models by Prof. Xunying Liu

Nov 18, 2016, Prof. Xunying Liu is invited to come and give lectures about “Scalable Deep Language Models”.

Xunying Liu received his PhD degree in speech recognition and MPhil degree in computer speech and language processing both from University of Cambridge, after his undergraduate study at Shanghai Jiao Tong University. He was a Senior Research Associate at the Machine Intelligence Laboratory of the Cambridge University Engineering Department, prior to joining the Department of Systems Engineering and Engineering Management, Chinese University of Hong Kong, as an Associate Professor in 2016. He was the recipient of best paper award at ISCA Interspeech2010 for his paper titled "Language Model Cross Adaptation For LVCSR System Combination". He is a co-author of the widely used HTK toolkit and has continued to contribute to its current development in deep neural network based acoustic and language modeling. His research outputs led to several large scale speech recognition systems that were top ranked in a series of international research evaluations. These include the Cambridge Mandarin Chinese broadcast and conversational telephone speech recognition systems developed for DARPA sponsored GALE and BOLT speech translation evaluations from 2006 to 2014, and the Cambridge 2015 multi-genre broadcast speech transcription system. His current research interests include large vocabulary continuous speech recognition,  machine learning, statistical language modelling, noise robust speech recognition, speech synthesis, speech and language processing. He is a regular reviewer for journals including IEEE/ACM Transactions on Audio, Speech and Language Processing, Computer Speech and Language, Speech Communication, the Journal of the Acoustical Society of America Express Letters, Language Resources and Evaluation, and Natural Language Engineering. He has served as a member of the scientific committee and session chair for conferences including IEEE ICASSP and ISCA Interspeech. Dr. Xunying Liu is a member of IEEE and ISCA.

The abstract of the lecture is as below:

Statistical language models (LMs) form key parts of many human language technology applications including speech recognition, machine translation, natural language processing and handwriting recognition. Key research problems are modeling long range context dependencies and handling data sparsity. Deep language modeling approaches represented by recurrent neural network (RNNs) are becoming increasingly popular for current speech and language technology applications due to their inherently strong sequence modeling ability and generalization performance. This talk presents a series of recent research efforts aiming to improve the scalability and performance of RNN language models (RNNLMs) on large data sets. A noise contrastive estimation (NCE) based RNNLM training criterion combined with an efficient GPU based bunch mode training algorithm obtained over 50 times training and evaluation time speed up over the publicly available RNNLM toolkit. Two history clustering schemes based efficient RNNLM lattice rescoring approaches produced over 70% more compact decoding network size than tree structured 10k-best lists with comparable performance. Novel approaches

modeling multiple paraphrase alternatives and topic variation increased the total RNNLM improvements over baseline n-gram LMs by a factor of 2.5. Experimental results are presented for multiple state-of-the-art large vocabulary speech recognition tasks.