Introduction
In Statistical Machine Translation (SMT), "reordering"—the process of determining the correct word order in the target language—remains a primary challenge. Current approaches typically face a trade-off between efficiency and accuracy. Contiguous models are computationally efficient but struggle with complex structural shifts. Conversely, non-contiguous or syntax-based models, such as the Hierarchical Phrase-based Translation Model (HPTM), offer high accuracy for complex reordering but suffer from significant computational overhead, particularly as sentence length increases.
This research introduces the PRML (Phrase Reordering based on Multi-Layer phrase) model. PRML addresses the "one-size-fits-all" inefficiency of standard models by intelligently applying different levels of reordering complexity to different segments of a sentence, achieving high-quality translation with a fraction of the decoding time.
Methodology: The Multi-Layer Approach
The core innovation of PRML is its hierarchical segmentation strategy. Rather than processing a whole sentence through a single complex model, the system decomposes the translation task into three distinct layers:
- Segmentation Module: Utilizing punctuation and syntactic "dependency integrity" (derived from an MST-based lexical dependency parser), the system splits the source sentence into sub-sentences and smaller chunks. A segment is considered to have dependency integrity if its internal words relate more strongly to each other than to external words.
- Layer 1: Chunks (Micro Level): The most computationally intensive model, HPTM, is applied only to these small chunks. This captures intricate local reorderings where they are most needed.
- Layer 2: Sub-sentences (Meso Level): The translated chunks are then processed by a Maximum Entropy-based Bracket Transduction Grammar (MEBTG) model. This layer determines whether adjacent blocks should be swapped or remain in order.
- Layer 3: Whole Sentence (Macro Level): Finally, a simple Monotone reordering model integrates the sub-sentence translations to produce the final output.
By treating the output of one layer as the input for the next, the PRML model focuses its computational power on small, complex segments while using simpler logic for the overall sentence structure.
Experimental Results
The model was evaluated on Chinese-to-English translation tasks using the NIST MT 2005, 2006, and 2008 benchmarks. The training data consisted of approximately 297,069 sentence pairs from the LDC corpus.
Translation Quality
PRML demonstrated significant improvements over the MEBTG baseline, showing an increase of up to 5.6% in BLEU scores on the NIST06 dataset. Crucially, it achieved translation quality comparable to or slightly exceeding the full HPTM model across all evaluation sets.
Efficiency Gains
The most striking result was the reduction in decoding time. Compared to the standard HPTM model, PRML reduced decoding time by 54.8% for NIST05, 75.6% for NIST06, and 65.2% for NIST08. While 93.6% of the processing time is still spent within the HPTM (chunk) layer, the absolute time is drastically reduced because the algorithm only operates on small, segmented chunks rather than the entire sentence.
Conclusion
The PRML model successfully bridges the gap between accuracy and efficiency in SMT. By leveraging syntactic anchors for segmentation and applying a multi-layered reordering strategy, it bypasses the search space explosion typical of hierarchical models. Future work will focus on refining the segmentation module and exploring the model's applicability to language pairs with more radical global reordering patterns.
Summary updated