Phrase Alignments in Parallel Corpus Using Bootstrapping Approach

Word choice and word order problems are considered as fundamental barriers in statistical machine translation (SMT). These barriers are more pronounced in deficiencies of training corpus. Phrase-Based SMT has advantages in word choice and local word ordering process; so phrase alignment is effective in improving translation quality. In this paper, an approach for automatic alignment is proposed in which boosts up the machine translation quality. Since, alignment problem is more problematic with lack of training data, so we make corpus of phrase alignment with high precision. For this purpose, a novel phrase alignment approach in a bootstrapping manner is proposed. By bootstrapping on alignment model via using a number of features, the accuracy of the phrase table is improved iteratively. These features are based on the phrase table extracted from Moses, IBM Model ۳ alignment probabilities, Google translator and fertility of candidates. Our experiments on English-Persian translation show an improvement about ۴.۱۷ BLEU points over the PB-SMT as baseline system.