Abstract
We study the challenges raised by Arabic verb and subject detection and reordering in Statistical Machine Translation (SMT). We show that post-verbal subject (VS) constructions are hard to translate because they have highly ambiguous reordering patterns when translated to English. In addition, implementing reordering is difficult because the boundaries of VS constructions are hard to detect accurately, even with a state-of-the-art Arabic dependency parser. We therefore propose to reorder VS constructions into SV order for SMT word alignment only. This strategy significantly improves BLEU and TER scores, even on a strong large-scale baseline and despite noisy parses.
Introduction
In Statistical Machine Translation (SMT), structural divergence between languages remains a significant hurdle. Modern Standard Arabic frequently utilizes a Verb-Subject-Object (VSO) order, whereas English almost exclusively follows a Subject-Verb-Object (SVO) structure. This divergence leads to two primary issues: data sparsity, where the system fails to generalize the relationship between verbs and subjects across different orders, and alignment errors, where the word aligner fails to link distant subjects and verbs correctly during training.
Methodology: Reordering for Alignment
The authors propose a "low-risk" reordering strategy that specifically targets the word alignment stage of training. Unlike previous methods that permanently reorder the source text—which can be catastrophic if the linguistic parser makes an error—this method only uses reordered text to generate better word alignments.
- Detection: A dependency parser identifies VS constructions and their boundaries.
- Reordering: A temporary version of the training data is created where VS sequences are flipped to SV order to match English syntax.
- Alignment: Standard word alignment (GIZA++) is performed on this reordered text, resulting in more accurate and denser links.
- Mapping Back: These high-quality alignments are projected back onto the original Arabic word order for the final decoding phase.
This approach ensures the system benefits from superior word-level correspondence without being forced to rely on potentially erroneous reordered input during live translation.
Key Insights: The VS Construction Challenge
Analysis of hand-aligned parallel treebanks reveals that post-verbal subjects are the norm in Arabic, appearing 70% of the time. However, their translation into English is not a simple inversion.
As shown in the data, while 98.2% of Arabic SV sentences remain monotone in English, VS constructions are more complex: 64.7% invert to SV, but 27.3% remain monotone (often due to passive voice translations). This ambiguity, combined with the difficulty of detecting subject boundaries (achieving only a 74% F-score), makes VS reordering a high-stakes task that is best handled at the alignment stage rather than the decoding stage.
Results
The proposed method yielded consistent improvements across multiple benchmarks. By improving the quality of the word alignments, the system produced significantly more links—up to 330,000 additional links in the large-scale system—which in turn created more comprehensive phrase tables.
On the large-scale baseline, the technique achieved an average improvement of +0.25 BLEU points and a reduction of -0.24 TER. These gains demonstrate that the method is robust to parser noise; even if a parse is imperfect, the resulting alignment is generally better than one produced from divergent word orders.
Conclusion
By focusing reordering efforts on the alignment phase, researchers can mitigate the risks associated with automated parsing errors while still capturing the structural benefits of syntactic reordering. Future work may involve using multiple subject boundary hypotheses or extending this logic to other syntactic divergences like adjective placement.
Summary updated