Learning Translations from Comparable Corpora

This thesis examines the possibility of using comparable corpora to augment statistical models of translation. Treating comparable corpora as marginal samples from an aligned bilingual joint distribution, the estimation of translation models from a combination of bilingual parallel and comparable corpora is seen as a variation of the labelled-unlabelled problem [Seeger, 2000b]. Results on synthetic data confirm that successful re-estimation within the EM framework [Dempster et al., 1977] is highly-dependent on the balance between complete and incomplete data [Nigam, 2001]. Here we show that the utility of re-estimation with additional incomplete data is highly-dependent on the accuracy of initial parameters estimated from the complete data alone. We propose a method for constraining the re-estimation procedure in relation to the degree of comparability between marginal samples. This is seen to result in better conditional models when the assumption of comparability is valid. Finally, we consider how more complex marginal models could be used to further constrain the re-estimation of the conditional.