Base Noun Phrase Translation Using Web Data and the EM Algorithm

Yunbo Cao, Hang LiCOLING 2002 · OpenReview

Abstract

Yunbo Cao, Hang Li. COLING 2002: The 19th International Conference on Computational Linguistics.

2002.

Introduction

Translating Base Noun Phrases (Base NPs)—simple, non-recursive phrases like "information age" or "calcium ion"—is a critical task in machine translation and cross-language information retrieval. These phrases often represent holistic concepts that cannot be accurately translated by simply concatenating dictionary definitions of individual words.

While parallel corpora (direct translations) are the gold standard for training translation models, they are scarce. Conversely, non-parallel corpora are abundant but present a "vector transformation" challenge: comparing the context of an English word with a Chinese word is difficult when words have multiple, ambiguous translations. This paper proposes a robust framework to bridge this gap using web data and the Expectation-Maximization (EM) algorithm.

Methodology

1. Translation Candidate Collection

The first step involves generating potential translations for an English Base NP using a bilingual dictionary. To filter these candidates, the system queries a web search engine to determine their document frequencies in Chinese web pages. Candidates with zero frequency are discarded, ensuring the model only considers phrases that actually appear in natural language use.

The process of generating and filtering translation candidates for 'information age' using web document frequencies.

2. Translation Selection via EM

To select the optimal translation from the filtered candidates, the authors developed two EM-based approaches:

  • EM-NBC-Ensemble: This method uses an ensemble of Naïve Bayes Classifiers. The EM algorithm estimates the probability of an English context word being a translation of a Chinese word. By employing multiple "window sizes" (from $\pm1$ to $\pm11$ words), the model captures both local syntactic cues and broader semantic context.
  • EM-TF-IDF: This approach calculates similarity between context vectors. Instead of a simple one-to-one mapping, the EM algorithm "splits" the frequency of an English word across its possible Chinese translations in a theoretically sound manner.
Illustration of context vector transformation and word mapping using the EM algorithm to handle translation ambiguity.

The system also utilizes a Back-off Strategy, prioritizing high-precision parenthetical translations (e.g., "信息时代 (information age)") found on the web before falling back to the EM-based selection.

Experimental Setup

The study utilized 3,000 English noun-noun pairs from the Encarta corpus. Web data was sourced via Google, and the HIT English-Chinese dictionary (containing ~76k Chinese and ~60k English words) served as the translation base. For comparison, static corpora (Wall Street Journal and People’s Daily) were used as baselines.

Key Results and Insights

  • The Power of the Web: The vastness of the web significantly improves coverage. Many valid translations absent in traditional corpora are readily found online.
  • EM Effectiveness: The EM-based transformation proved superior to "Major Translation" (picking the most frequent dictionary entry). It allowed the model to distribute weights across all possible translations based on observed data, resulting in context vectors that closely match the target language distribution.
  • Performance: The combined strategy achieved 91.4% coverage and a top-3 accuracy of 79.8%. Switching from web data to static corpora caused coverage to drop to 79.3% and Top-1 accuracy to fall from 62.9% to 56.9%.

Conclusion

The proposed framework demonstrates that leveraging web data and probabilistic modeling via EM can overcome the limitations of scarce parallel corpora for Base NP translation. While ranking errors remain a challenge (accounting for 68% of errors), the approach provides a scalable and accurate solution for cross-language concept mapping. Future work will explore extending this method to more complex phrase structures and additional language pairs.

Summary updated

Abstract

Yunbo Cao, Hang Li. COLING 2002: The 19th International Conference on Computational Linguistics. 2002.
BibTeX
@inproceedings{cao2002base,
  author = {Yunbo Cao and Hang Li},
  title = {{Base Noun Phrase Translation Using Web Data and the EM Algorithm}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
  url = {https://openreview.net/forum?id=IOwaRvA0yBr},
}
Figure