@PhdThesis{Fed2013b,
AUTHOR = {Federmann, Christian},
TITLE = {Hybrid machine translation using binary classification models trained on joint, binarised feature vectors},
YEAR = {2013},
SCHOOL = {Universität des Saarlandes},
ABSTRACT = {We describe the design and implementation of a system combination method for machine translation output. It is based on sentence selection using binary classification models estimated on joint, binarised feature vectors. By contrast to existing system combination methods which work by dividing candidate translations into n-grams, i.e., sequences of n words or tokens, our framework performs sentence selection which does not alter the selected, best translation. First, we investigate the potential performance gain attainable by optimal sentence selection. To do so, we conduct the largest meta-study on data released by the yearly Workshop on Statistical Machine Translation (WMT). Second, we introduce so-called joint, binarised feature vectors which explicitly model feature value comparison for two systems A, B. We compare different settings for training binary classifiers using single, joint, as well as joint, binarised feature vectors. After having shown the potential of both selection and binarisation as methodological paradigms, we combine these two into a combination framework which applies pairwise comparison of all candidate systems to determine the best translation for each individual sentence. Our system is able to outperform other state-of-the-art system combination approaches; this is confirmed by our experiments. We conclude by summarising the main findings and contributions of our thesis and by giving an outlook to future research directions.},
NOTE = {HU}
}
|
