Building a reordering system using tree-to-string hierarchical model

TL;DR

This paper presents a tree-to-string hierarchical reordering system for statistical machine translation, demonstrating significant improvements across multiple language pairs using publicly available tools and a detailed, reproducible methodology.

Contribution

The paper introduces a reordering approach based on a tree-to-string model using the Moses toolkit, applicable to various language pairs, with comprehensive experimental results and discussion.

Findings

Significant BLEU, Kendall-Tau, and Hamming improvements over baseline.

Effective reordering for English-Farsi, English-Italian, and English-Urdu.

Reproducible methodology using publicly available tools.

Abstract

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.