Chinese Syntactic Reordering for Statistical Machine Translation

Chao Wang, Michael Collins, Philipp KoehnEMNLP 2007 · OpenReview

Chao Wang, Michael Collins, Philipp Koehn Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)

Abstract

Statistical Machine Translation (SMT) systems often struggle with significant word order differences between source and target languages. This paper presents a preprocessing method for Chinese-to-English translation that uses linguistically-informed rules to reorder Chinese parse trees into an English-like sequence. By addressing long-distance structural divergences—such as the placement of relative clauses and prepositional phrases—before the translation phase, the authors achieve a significant improvement in translation quality, with a +2.34 BLEU point gain on the NIST 2006 evaluation set.

The Challenge of Word Order Divergence

A primary hurdle in SMT is "word order divergence." Chinese and English exhibit systematic structural differences; for instance, Chinese relative clauses and prepositional phrases typically precede the heads they modify, whereas English usually places them after. While phrase-based SMT systems utilize distortion models to manage reordering during decoding, these models are often computationally expensive and struggle with the long-distance movements required for Chinese-English pairs. This research explores a deterministic preprocessing step to align the source sentence structure with the target language's syntax before it reaches the translation engine.

Methodology

The proposed approach involves a four-stage pipeline:

  1. Parsing: Source Chinese sentences are processed into constituent trees using a Treebank-trained parser.
  2. Transformation: Hand-crafted rules are applied to these trees to relocate specific syntactic constituents.
  3. Flattening: The modified trees are converted back into linear word sequences.
  4. Translation: The reordered "English-like" Chinese string is translated using a standard phrase-based SMT system (MOSES).

The reordering rules target three main linguistic categories:

  • Verb Phrases (VP): Modifiers like prepositional phrases (PP) and temporal nouns are moved from pre-verbal to post-verbal positions.
  • Noun Phrases (NP): Relative clauses (CP) and "DEG" phrases (DNP) are moved to follow the nouns they modify.
  • Localizers (LCP): Post-positions (e.g., "room inside") are flipped to function as prepositions (e.g., "inside room").
Example of syntactic reordering transforming a Chinese parse tree into an English-like word order.

The syntactic labels used in these transformations follow standard Treebank conventions, which define the various phrase types and parts of speech involved in the reordering process.

Common syntactic labels used in the reordering rules.

Results and Discussion

The system was trained on approximately 637,000 sentence pairs and evaluated on official NIST MT datasets. The reordering approach significantly outperformed the baseline MOSES system, raising the NIST 2006 BLEU score from 28.52 to 30.86 (a gain of +2.34 points).

Key insights from the analysis include:

  • VP Reordering is Most Impactful: Among all rules, moving Verb Phrase modifiers had the highest individual impact on translation quality (+1.14 BLEU).
  • Robustness to Noise: Despite the reordering being "correct" only 62% of the time (often due to initial parser inaccuracies), the SMT engine still benefited significantly. This suggests that even imperfect syntactic reordering provides a better starting point for phrase-based translation than the original source order.
  • Alignment Synergy: By making the source and target sentences look more alike, the system creates better word-to-word alignments during training, leading to a cleaner and more accurate phrase table.

Conclusion and Future Work

Linguistically-informed syntactic reordering is a powerful tool for improving Chinese-to-English SMT. By simplifying the task for the translation engine's distortion model and improving word alignment during training, this preprocessing step yields substantial quality gains.

While the current system makes "hard" reordering decisions, the authors suggest that future work could involve reordered lattices. This would allow the translation engine to consider multiple possible reorderings simultaneously, effectively accounting for parsing uncertainty and potentially further increasing translation accuracy.

Summary updated

BibTeX
@inproceedings{wang2007chinese,
  author = {Chao Wang and Michael Collins and Philipp Koehn},
  title = {{Chinese Syntactic Reordering for Statistical Machine Translation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2007},
  url = {https://openreview.net/forum?id=SyZYyMfOWS},
}
Figure