The present work reports system combination task for the Chinese-English statistical machine translation systems. We focus on the strategy to build the candidate systems to enhance the gain of BLEU score by introducing diversity at the early stage of the system combination. One of the most effective strategies is to carry out system combination of the various systems with different word alignment algorithm. Our approach differs from previous work in one important aspect that we report on the diversity of the alignment refinement heuristics of word alignment techniques that are complementary to each other for the system combination. This approach could harness several word alignment possibilities and proved to be beneficial in generating consensus translation where the acting backbone which determines the word order is permitted to switch after each word. We carried out experiments on candidate systems of phrasal and hierarchical paradigms and system combination of both the paradigms as well. To our surprise, the combo systems using the various word alignments with various symmetrization techniques of both the MT paradigms show gain of 0.8 to 2.07 absolute BLEU score against the best candidates of the respective test sets.
展开▼