Building a reordering system using tree-to-string hierarchical model
This work addresses reordering challenges in machine translation for multiple language pairs, though it is incremental as it applies existing tools and methods to new data.
The paper tackles the problem of sentence reordering for statistical machine translation by building a system based on a tree-to-string hierarchical model, achieving significant improvements over baselines in BLEU, Kendall-Tau, and Hamming metrics for English-Farsi, English-Italian, and English-Urdu language pairs.
This paper describes our submission to the First Workshop on Reordering for Statistical Machine Translation. We have decided to build a reordering system based on tree-to-string model, using only publicly available tools to accomplish this task. With the provided training data we have built a translation model using Moses toolkit, and then we applied a chart decoder, implemented in Moses, to reorder the sentences. Even though our submission only covered English-Farsi language pair, we believe that the approach itself should work regardless of the choice of the languages, so we have also carried out the experiments for English-Italian and English-Urdu. For these language pairs we have noticed a significant improvement over the baseline in BLEU, Kendall-Tau and Hamming metrics. A detailed description is given, so that everyone can reproduce our results. Also, some possible directions for further improvements are discussed.