Training Nondeficient Variants of IBM-3 and IBM-4 for Word Alignment
Thomas Schoenemann · Meeting of the Association for Computational Linguistics · 2013
We derive variants of the fertility based models IBM-3 and IBM-4 that, while maintaining their zero and first order parameters, are nondeficient. Subsequently, we proceed to derive a method to compute a likely alignment and its neighbors as well as give a solution of EM training. The arising M-step energies are non-trivial and handled via projected gradient ascent. Our evaluation on gold alignments shows substantial improvements (in weighted Fmeasure) for the IBM-3. For the IBM4 there are no consistent improvements. Training the nondeficient IBM-5 in the regular way gives surprisingly good results. Using the resulting alignments for phrasebased translation systems offers no clear insights w.r.t. BLEU scores.