ISI’s Participation in the Romanian-English Alignment Task

摘要

We discuss results on the shared task ofRomanian-English word alignment. Thebaseline technique is that of symmetrizingtwo word alignments automatically generatedusing IBM Model 4. A simple vocabularyreduction technique results in animprovement in performance. We alsoreport on a new alignment model and anew training algorithm based on alternatingmaximization of likelihood with minimizationof error rate.