Thesis (Ph.D.) - Cairo University - Faculty of Computers and Information - Department of Computer Science

A statistical Machine Translation, the state-of-the-art of MT approach nowadays, can learn from a huge amount of data, but originally designed as a batch model. Retraining SMT existing models, using human edits to MT output, are dominating the research field. Traditionally, user{u2019}s feedback is linked to commercial applications, when a review is written or a product is rated similarly, translator{u2019}s feedback is used to improve SMT; therefore, the user could receive better translation learnt from his feedback. This dissertation proposes an online incremental method for statistical machine translation system, in a scenario utilizing experts edit and correction for the SMT output. By updating the model by new translation rules, learning new vocabulary or adapting the MT system to a human translator style. This dissertation presents a new method to improve SMT using post-edits. The proposed method compares the post-edit sentences with the hypotheses translation output in order to automatically detect where the decoder made a mistake and learn from it. Once the errors have been detected, new word alignments are computed between input and post-edit sentences, proposing a set of similarity features, to extract translation units that are then merged online into the system to fix those errors for future translations

Issued also as CD