Dependency-enhanced reordering model for Chinese-English SMT

This paper proposes a dependency-enhanced pre-reordering method for Chinese-English statistical machine translation (SMT). Firstly, two kinds of dependency structure-based rules are extracted based on the source-side dependency tree and corresponding word alignments between the source-side and the target-side sentences. Then a maximum entropy classifier is used to calculate the orientation probability in terms of swap or monotone between two rules. As a result, a reordering rule set is obtained. Two different ways are proposed to filter out the rule set. Afterwards, the dependency parsing trees of the training data, development set and the test set are traversed, and if the syntactic sub-tree structure matches the rules in the rule set, the word orders will be adjusted. Thus, a reordered source-side sentence is generated and then fed into an SMT system for translation. Experiments conducted on NIST Chinese-English MT data sets show that the proposed method significantly improves translation performance by 0.46 BLEU compared to the baseline system.