Sampling-based Alignment and Hierarchical Sub-sentential Alignment in Chinese-Japanese Translation of Patents

Wei Yang, Zhongwen Zhao, Baosong Yang, Yves LepageWAT 2015

Introduction

In the field of Phrase-based Statistical Machine Translation (PB-SMT), the alignment process—identifying correspondences between words or phrases in parallel sentences—remains a significant computational bottleneck. Traditional tools like GIZA++ utilize complex probabilistic models that seek a global optimum across the entire dataset. While accurate, these methods are computationally expensive and slow, particularly when applied to large-scale datasets like patent corpora. Researchers from Waseda University have addressed this challenge by proposing an associative approach that leverages local word associations to achieve comparable translation quality with a fraction of the training time.

Methodology

The proposed system combines two distinct associative alignment techniques, optimized for speed and parallel processing:

1. Sampling-based Alignment (Anymalign)

Anymalign calculates translation probabilities by randomly sampling sub-corpora and observing the co-occurrence of words. Unlike global optimization models, Anymalign can be terminated at any time (a "timeout"), providing a "good enough" result rapidly. This method is particularly effective for technical domains like patents, where terminology is highly consistent.

Word association matrix generated by Anymalign for a Chinese-Japanese sentence pair, showing associative strengths between tokens.

2. Hierarchical Sub-sentential Alignment (Cutnalign)

Using the association strengths derived from Anymalign, Cutnalign treats a sentence pair as a matrix. It employs recursive binary segmentation to find the optimal alignment path between sub-sentence units. To enhance performance, the authors re-implemented the core of Cutnalign in C, introduced multi-processing, and utilized mathematical approximations to skip redundant calculations during matrix splitting.

The recursive hierarchical segmentation process used by Cutnalign to partition the association matrix into sub-sentential alignment blocks.

Experimental Setup

The study utilized the JPO Patent Corpus (JPC) from the 2015 Workshop on Asian Translation (WAT), consisting of 1 million Chinese–Japanese parallel sentence pairs across four domains: Chemistry, Electricity, Mechanical Engineering, and Physics. Text was preprocessed using the Stanford Segmenter for Chinese and Juman for Japanese.

Key Results and Discussion

The optimized pipeline demonstrated a significant reduction in training time compared to standard baselines:

  • Training Speed: The fastest configuration completed training in just 57 minutes, representing an 80% reduction compared to the 4 hours and 43 minutes required by GIZA++ and the 5 hours and 34 minutes required by the parallelized MGIZA.
  • Translation Quality: The proposed method achieved a BLEU score of 36.11, which is highly competitive with the GIZA++ baseline of 37.46.

The researchers found that associative methods are "embarrassingly parallel," allowing them to scale efficiently on modern multi-core processors. Furthermore, the structured nature of patent text makes it an ideal candidate for sampling-based methods, as recurring patterns are easily identified even with limited sampling.

Conclusion

By combining sampling-based alignment with optimized hierarchical segmentation, the researchers achieved a 5x speedup in training for Chinese-Japanese patent translation with only a marginal impact on BLEU scores. While the industry has largely transitioned toward Neural Machine Translation (NMT), these fast alignment principles remain highly relevant for data filtering and the generation of synthetic parallel corpora in modern translation workflows.

Summary updated

BibTeX
@inproceedings{yang2015sampling,
  author = {Wei Yang and Zhongwen Zhao and Baosong Yang and Yves Lepage},
  title = {{Sampling-based Alignment and Hierarchical Sub-sentential Alignment in Chinese-Japanese Translation of Patents}},
  booktitle = {WAT},
  year = {2015},
}
Figure