|Adrià de Gispert
|Dr. Adrià de Gispert is a lecturer in Speech and Language Technologies in the Engineering Department of the University of Cambridge, U.K. He has been doing research in Statistical Machine Translation (SMT) since 2001. He obtained his PhD on integrating linguistic knowledge into SMT systems at the Universitat Politècnica de Catalunya (Barcelona) in 2006. In that time he participated in the development of the ngram-based approach to translation, and took active part in various research projects related to statistical translation of text and speech, such as FAME and TC-STAR (EU-funded), ALIADO (Spanish Government). In 2007 he took a research position at the University of Cambridge, where he has worked in the AGILE project of the DARPA-funded GALE program, developing various SMT systems (ngram-based, phrase-based, hierarchical phrase-based) and participating successfully in multiple international evaluation campaigns for several language pairs and conditions (IWSLT 2004-2006 tasks, ACL-WMT 2005-2010 shared tasks, NIST 2006-2009 evaluation), including the translation of automatic speech recognition outputs. His current research interests are: large-scale statistical translation systems with little pruning, grammar induction, morphology-aware models, distributed models, reordering models and MT system combination, among others.
|Addresses: www mail
- 14:10, Tuesday, July 20th, 2010
- Room FA1 - Pavilhão de Informática I (IST)
- Adrià de Gispert, Department of Engineering, University of Cambridge
In this talk I will describe HiFST, the Cambridge University
Engineering Department statistical machine translation system. I will
review hierarchical-phrase based translation, and explain why an
implementation based on Weighted Finite-State Transducers is
convenient to avoid search errors in decoding.
I will then give an overview of the current research lines of our SMT
group, focused on defining appropriate translation grammars,
exploiting the vast number of alternative hypotheses encoded in
translation lattices via restoring with large-scale language models,
and decoding under Minimum Bayes Risk for system combination.