Simple and Effective Approach for Consistent Training of Hierarchical Phrase-based Translation Models

In this paper, we present a simple ap-proach for consistent training of hierarchi-cal phrase-based translation models. In order to consistently train a translation model, we perform hierarchical phrase-based decoding on training data to find derivations between the source and tar-get sentences. This is done by syn-chronous parsing the given sentence pairs. After extracting k-best derivations, we reestimate the translation model proba-bilities based on collected rule counts. We show the effectiveness of our proce-dure on the IWSLT German→English and English→French translation tasks. Our results show improvements of up to 1.6 points BLEU.