30.01.2014 Views

Annual Report 2010 - Fachgruppe Informatik an der RWTH Aachen ...

Annual Report 2010 - Fachgruppe Informatik an der RWTH Aachen ...

Annual Report 2010 - Fachgruppe Informatik an der RWTH Aachen ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

• The phrase-based tr<strong>an</strong>slation system was improved with a focus on search org<strong>an</strong>ization,<br />

including new knowledge source <strong>an</strong>d better coupling with automatic speech recognition<br />

systems.<br />

• Additionally to the phrase-based tr<strong>an</strong>slation system, a hierarchical tr<strong>an</strong>slation system was<br />

implemented using the cube growing <strong>an</strong>d cube pruning algorithms in decoding. It<br />

performs similar to the phrase-based system <strong>an</strong>d thus has been extensively used in<br />

evaluations. Further extensions to this are being investigated.<br />

• Two extensions of st<strong>an</strong>dard word lexicons in machine tr<strong>an</strong>slation have been implemented:<br />

A discriminative word lexicon that uses sentence-level source information to predict the<br />

target words <strong>an</strong>d a trigger-based lexicon model that extends IBM model 1 with a second<br />

trigger, allowing for a more fine-grained lexical choice of target words.<br />

• A consistent phrase model training using a forced alignment procedure has been<br />

implemented. This novel method utilizes phrase-alignment data in or<strong>der</strong> to make training<br />

consistent with the tr<strong>an</strong>slation deco<strong>der</strong>.<br />

• Different possibilities for h<strong>an</strong>dling large l<strong>an</strong>guage models in the tr<strong>an</strong>slation process have<br />

been investigated. These approaches allow the usage of large l<strong>an</strong>guage models with a<br />

relatively small memory footprint <strong>an</strong>d have been successfully applied in the systems used<br />

in evaluations.<br />

• Our method for system combination for statistical machine tr<strong>an</strong>slation, inspired from<br />

methods in speech recognitions, was improved.<br />

• Research efforts were continued in the area of automatic tr<strong>an</strong>slation between Germ<strong>an</strong><br />

written text <strong>an</strong>d Germ<strong>an</strong> Sign L<strong>an</strong>guage. In April 2009 the SignSpeak project started.<br />

Speech Recognition<br />

Architecture of <strong>an</strong> automatic speech recognition system<br />

Today, state-of-the-art systems for automatic speech recognition are based on the statistical<br />

approach of Bayes decision rule. The implementation of Bayes decision rule for automatic<br />

speech recognition is based on two kinds of stochastic models: the acoustic model <strong>an</strong>d the<br />

l<strong>an</strong>guage model which together are the basis for the decision process itself, i.e. the search for<br />

240

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!