[Univ of Cambridge]alt[Dept of Engineering]

MIL Speech Seminars 2004-2005

The MIL Speech Seminar series schedule for Lent Term 2005 was as follows:

February 22nd 2005 Bill Byrne (MIL) Current Research in Phrase-Based Statistical Machine Translation Current approaches to statistical machine translation are based on a source-channel modeling approach which will be familiar to speech recognition researchers. These statistical translations models are now accompanied by automatic performance metrics, large training collections, common evaluation sets, and competitive system evaluation, so that research in statistical MT has grown to resemble even more closely the current approaches to automatic speech recognition. This talk will review some recent trends in statistical translation in the context of developing a phrase-based statistical machine translation system. I will discuss performance metrics, statistical models of word and document alignment, and search strategies for translation. The presentation will be `speech recognition friendly', in that the approach and techniques used were borrowed or inspired by recent experience in building ASR systems.
March 8th 2005 Shankar Kumar (MIL) The Translation Template Modeling Framework for Statistical Machine Translation I will present the Translation Template Model (TTM) which is a phrase-based generative model of automatic translation. This model is formulated and implemented entirely using Weighted Finite State Transducers (WFSTs) and I will describe how this approach yields state-of-the-art translation while avoiding the need for specialized search procedures for generating lattices and N-best lists of translations and word alignments. I will also discuss some issues related to WFST modeling: how to describe local phrase movement with WFSTs; how Minimum Bayes-Risk Translation improves translation by minimizing empirical risk of errors under specific translation loss functions; and discriminative training can be used to refine TTM model parameters to improve translation. This talk serves as an introduction to the TTM Toolkit, which is available for use in the MIL.