Bootstrapping Method for Chunk Alignment in Phrase Based SMT

The processing of parallel corpus plays very crucial role for improving the overall performance in Phrase Based Statistical Machine Translation systems (PBSMT). In this paper the automatic alignments of different kind of chunks have been studied that boosts up the word alignment as well as the machine translation quality. Single-tokenization of Noun-noun MWEs, phrasal preposition (source side only) and reduplicated phrases (target side only) and the alignment of named entities and complex predicates provide the best SMT model for bootstrapping. Automatic bootstrapping on the alignment of various chunks makes significant gains over the previous best English-Bengali PB-SMT system. The source chunks are translated into the target language using the PB-SMT system and the translated chunks are compared with the original target chunk. The aligned chunks increase the size of the parallel corpus. The processes are run in a bootstrapping manner until all the source chunks have been aligned with the target chunks or no new chunk alignment is identified by the bootstrapping process. The proposed system achieves significant improvements (2.25 BLEU over the best System and 8.63 BLEU points absolute over the baseline system, 98.74% relative improvement over the baseline system) on an EnglishBengali translation task.