Bilingual Lexicon Construction from Comparable Corpora via Dependency Mapping

Longhua Qian, Hongling Wang, Guodong Zhou, Qiaoming ZhuCOLING 2012 · OpenReview

Introduction

Building bilingual lexicons is a foundational task for machine translation and cross-language information retrieval. While parallel corpora—texts that are direct translations—provide a gold standard for this task, they are often scarce for many language pairs. Consequently, researchers have turned to comparable corpora, such as English and Chinese Wikipedia, which cover similar topics without being direct translations.

Traditional approaches to building lexicons from comparable corpora rely on the "distributional hypothesis": words that are translations of each other tend to appear in similar contexts. However, simple word-proximity models often suffer from the "noisy neighbor" problem. For instance, the English word "profit" and the Chinese word for "enterprise" might both frequently appear near terms like "realize" or "cost," leading a model to incorrectly pair them. This paper introduces a method to mitigate this noise by incorporating syntactic structure through dependency mapping.

The Dependency Mapping Approach

The core innovation of this research is the Bilingual Dependency Mapping model. Instead of analyzing a simple window of neighboring words, the model examines dependency triples: (word, relationship, dependent). By focusing on how a word functions grammatically—such as whether it acts as a direct object or a nominal subject—the model can more precisely identify its translation equivalent.

Visual representation of dependency relationship mapping between English and Chinese.

As shown in Figure 1, the system maps specific linguistic relationships across languages. For example, an English nsubj (nominal subject) is likely to correspond to a Chinese nsubj. This structural alignment provides a much stronger signal than simple proximity, as it captures the functional role of the word within a sentence.

Automatic Mapping and Weight Learning

Rather than relying on manually defined linguistic rules, the authors developed a system to automatically discover these mappings using a small "seed" lexicon of 1,000 frequent word pairs. This process identifies which English dependency types most frequently correspond to specific Chinese ones.

Top 30 automatically discovered English-to-Chinese dependency mappings and their frequencies.

Figure 3 details the top 30 discovered mappings. Interestingly, the system found that relationships like prep_of-nn (7.3%) and nn-nn (7.0%) are highly frequent and informative. To further refine these results, the researchers employed an Averaged Perceptron algorithm to learn weights for different mappings. This allows the model to prioritize highly discriminative relationships, such as conjunctions or possessives, over more generic or noisy ones.

Key Results and Insights

The dependency-based model demonstrated a significant performance leap over traditional word-context baselines. On a test set of frequent nouns, the model achieved a 14-point improvement in both Precision and Mean Reciprocal Rank (MRR).

Several key insights emerged from the study:

  • Syntax vs. Proximity: Grammatical roles provide a more precise "fingerprint" for a word than mere proximity. This effectively solves the "enterprise vs. profit" trap where high-frequency words "steal" translations in simpler models.
  • The Adjective Hurdle: While the model excelled with nouns and verbs, it showed little improvement for adjectives. The authors attribute this to the simplicity of adjective dependencies (often just a single amod relationship), which lack the structural variety needed for the dependency model to outperform simpler window-based methods.
  • Automation Success: Automatically generated mappings performed as well as, or better than, those crafted by human experts, proving the robustness of the automated discovery process.

Conclusion

By leveraging the syntactic structure of language, this research provides a powerful new tool for bilingual lexicon construction from comparable corpora. While challenges remain—particularly regarding data sparseness for rare words and the limited structural complexity of adjectives—the 14% accuracy boost for nouns represents a major step forward in cross-lingual natural language processing.

Summary updated

BibTeX
@inproceedings{qian2012bilingual,
  author = {Longhua Qian and Hongling Wang and Guodong Zhou and Qiaoming Zhu},
  title = {{Bilingual Lexicon Construction from Comparable Corpora via Dependency Mapping}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2012},
  url = {https://openreview.net/forum?id=z6OeIWTc1H},
}
Figure