Improving Statistical Machine Translation Using Domain Bilingual Multiword Expressions

Zhixiang Ren, Yajuan Lü, Jie Cao, Qun Liu, Yun HuangMWE@IJCNLP 2009 · OpenReview

Abstract

Zhixiang Ren, Yajuan Lü, Jie Cao, Qun Liu, Yun Huang. Proceedings of the Workshop on Multiword Expressions: Identification, Interpretation, Disambiguation and Applications (MWE 2009).

2009.

This paper introduces a novel approach to boost Statistical Machine Translation (SMT) performance by automatically extracting and integrating domain-specific bilingual multiword expressions (BiMWEs). By using a hierarchical reducing algorithm and testing three different integration strategies in the Moses SMT system, the authors achieved significant improvements in translation quality for specialized fields like traditional medicine and chemical engineering.

Introduction

Phrase-based SMT models often struggle when moving from general language to specialized domains. Technical terminology, idioms, and fixed expressions—collectively known as Multiword Expressions (MWEs)—frequently have "idiosyncratic" meanings that cannot be predicted by looking at their individual component words. For example, a medical term might be translated literally by a general-purpose system, resulting in "word salad" rather than the correct technical term.

While researchers know that identifying these expressions is crucial, integrating them into existing SMT workflows is difficult. Manual dictionaries are too slow to build, and automatic extraction often introduces "noise" (incorrect pairings) that can actually degrade translation performance if not handled carefully.

Methods

The authors proposed a two-stage pipeline: extraction and integration.

Extraction: The Hierarchical Reducing Algorithm (HRA)

To find MWEs in Chinese text, the authors developed the LLR-based Hierarchical Reducing Algorithm. Unlike standard statistical measures that only look at word pairs (bigrams), the HRA allows for the extraction of MWEs of arbitrary length by iteratively building a tree-like structure.

Hierarchical structure of a Chinese MWE

The algorithm starts with individual words and calculates the Log Likelihood Ratio (LLR) between adjacent units. It iteratively merges the pair with the highest association score into a single unit until a threshold is reached or only one unit remains. Once Chinese MWEs are found, the system looks for their English counterparts in a parallel corpus using GIZA++ alignment. A perceptron-based classifier then filters these candidates to ensure high-quality bilingual pairs.

Integration Strategies

The authors tested three ways to "teach" the SMT system (Moses) these new BiMWEs:

  • Model Retraining (Baseline+BiMWE): They added the extracted BiMWEs directly into the training corpus as if they were short sentences and retrained the entire model.
  • New Feature (Baseline+Feat): They added a binary "indicator" feature to the SMT phrase table. This feature marks whether a phrase pair contains a known BiMWE, allowing the decoder to give those pairs more "weight" during translation.
  • Additional Phrase Table (Baseline+NewBP): They created a secondary phrase table exclusively for BiMWEs and forced the decoder to consult both the original and the new table.

Datasets

The study utilized two domain-specific Chinese-to-English patent corpora:

  • Traditional Medicine Domain: ~120,000 training sentences, yielding 80,287 bilingual MWEs.
  • Chemical Industry Domain: ~120,000 training sentences, yielding 59,466 bilingual MWEs.

Key Results and Insights

The methods yielded measurable improvements in BLEU scores across both domains. In the Chemical Industry domain, the baseline score of 0.1882 rose to 0.1928 using the retraining method. In Traditional Medicine, the baseline of 0.2658 rose to 0.2734 using the "New Feature" method.

BLEU scores for Traditional Medicine with varying MWE counts

A key finding was the relationship between quantity and quality. Performance peaked at around 60,000 MWEs; adding more introduced enough noise to diminish the gains. Simply adding a new feature (the "Baseline+Feat" method) proved to be one of the most robust strategies. It allows the system to benefit from MWE knowledge without the risk of "over-correcting" the entire translation model.

Qualitatively, the MWE-enhanced system showed significant improvement. In one example, the baseline system incorrectly translated a Chinese term as "promoting salivation and water." The MWE-enhanced system correctly identified the technical term as "dredging meridians."

Conclusion

The authors acknowledge that their current application strategies are relatively straightforward. The primary limitation is the noise inherent in automatic extraction; even with filtering, incorrect MWE pairs can mislead the translator. Future directions include developing more sophisticated integration strategies beyond binary features and testing the approach on different MT architectures, such as hierarchical phrase-based or syntax-based models.

Summary updated

BibTeX
@inproceedings{ren2009improving,
  author = {Zhixiang Ren and Yajuan Lü and Jie Cao and Qun Liu and Yun Huang},
  title = {{Improving Statistical Machine Translation Using Domain Bilingual Multiword Expressions}},
  booktitle = {MWE@IJCNLP},
  year = {2009},
  url = {https://openreview.net/forum?id=rCgZ9SVx2Gq},
}
Figure