📄 Clause Restructuring for Statistical Machine Translation

Summary

Phrase-based statistical MT handles local reordering reasonably well, but it struggles when source and target languages differ systematically in clause-level word order. German and English are a hard case because of verb-second matrix clauses, verb-final subordinate clauses and clause-final infinitives. The authors propose a preprocessing approach. They parse each German sentence and apply a small set of hand-written, linguistically motivated transformations to the tree, producing a German word sequence that follows English order more closely. This reordering is applied to both the training data and the test input. The modified text is then fed to an unchanged phrase-based system (Koehn et al., 2003). On German-to-English Europarl translation, Bleu rose from 25.2% to 26.8%, a statistically significant gain. Human judgements also favoured the reordered output. The work shows that explicit syntax can help a strong phrase-based system without redesigning the decoder or the model.

Key points

• The task is German→English translation on the Europarl corpus. The underlying system is the phrase-based model of Koehn et al. (2003), and the German parser was provided by Amit Dubey.
• Bleu improved from 25.2% (baseline) to 26.8% (with clause restructuring), and the difference is statistically significant.
• The rules target several German phenomena:
  • finite verbs in second position in main clauses and in final position in subordinate clauses;
  • infinitival verbs at the end of their VP;
  • subjects appearing after the finite verb;
  • the placement of negation (nicht vs. not).
• Human evaluation:
  • One annotator judged the reordered system better on 40 of 100 sampled sentences.
  • In a larger per-sentence comparison over 1785 test sentences, about 52.85% improved, 36.4% got worse and 10.75% stayed the same.
• Xia and McCord (2004) learned over 56,000 reordering rules automatically. This paper uses only a handful of manual transformations and reports comparable gains, though on a different language pair.

Interesting ideas

• I1. Reorder the source, keep the translator unchanged. Syntax enters only as a preprocessing step applied to both training and test data, so any existing system can benefit without modifying its internals. This could carry over to other pipelines where reworking the core model is expensive.
• I2. Few general rules instead of many local ones. The parser produces flat German VPs and clauses, which yield a huge number of distinct context-free productions. A single rule such as "move the infinitival verb to the front of the VP" covers all of them, whereas production-level rules would need a separate transformation for each. This matters for any language with free word order and flat treebank structures, which is relevant to Ukrainian.
• I3. Linguistic analysis as a design source. The transformations follow directly from a description of how German clause structure differs from English. A short contrastive grammar sketch can be turned into effective, interpretable engineering rules before trying data-driven alternatives.
• I4. Pairwise human judgements with sign-test-style statistics. Besides Bleu, the authors count per-sentence better/worse/equal judgements and compute confidence intervals. They also question whether bootstrap tests are valid for Bleu. This is a useful complement whenever automatic metric gains are small.