Factored Language Models

"Factored Language Models"Katrin Kirchhoff
Department of Electrical Engineering
University of Washington, SeattleAbstractLanguage modeling for large-vocabulary automatic speech recognition is often faced with the problem of sparse training data. This is a particularly severe problem for languages with rich morphology, which results in a high vocabulary growth rate and a large number of unseen word contexts. One way of addressing data sparsity is to adopt a feature-based word representation, where words are not considered atomic units but feature vectors consisting of more fine-grained modeling units. This representation can be used to develop more robust language models (so-called factored language models) that are able to share training data across words. I will describe how to develop factored word representations and the resulting probability estimation methods and present results on language modeling for both speech recognition and machine translations.BiographyKatrin Kirchhoff studied Linguistics and Computer Science at the Universities of Bielefeld (Germany) and Edinburgh (UK) and obtained her PhD in Computer Science from the University of Bielefeld in 1999. Since then she has been a Research Assistant Professor in the Electrical Engineering Department at the University of Washington. Her main research interests are speech recognition, language modeling, multilingual speech processing, and machine learning.Date: April 29, 2005 (Friday)
Time: 10:00 am
Host: Shri Narayanan, EE-Systems (SIPI)
Contact: shri@sipi.usc.edu (x 06432)

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248

Audiences: Everyone Is Invited

Contact: Regina Morton