Analyzing Methods for Improving Precision of Pivot Based Bilingual Dictionaries

Xabier Saralegi, Iker Manterola, Iñaki San VicenteEMNLP 2011

Abstract

Building bilingual dictionaries for low-resource languages often relies on a "pivot" language (e.g., using English to link Basque and Spanish). However, polysemy—words with multiple meanings—creates significant "noise" in these inferred dictionaries. This study demonstrates that combining structural dictionary analysis (Inverse Consultation) with statistical context analysis (Distributional Similarity) can automatically prune these errors. This hybrid approach significantly improves dictionary precision and recall without requiring rare parallel corpora, achieving an average F-score of 0.52.

Introduction: The Polysemy Problem in Pivot Dictionaries

Most language pairs lack direct bilingual dictionaries. The standard workaround is the pivot method: if dictionaries exist for language pairs A-B and B-C, an A-C dictionary can be inferred using B as a bridge. However, language is not transitive. For example, a Basque word meaning "water faucet" might map to the English word "tap," which in turn maps to a Spanish word for a "light blow."

In a raw merge of Basque-English and English-Spanish dictionaries, researchers found that over 80% of ambiguous entries contained incorrect translations. The challenge lies in automatically filtering these "faucet-vs-tap" errors, especially when large-scale parallel corpora are unavailable.

Methodology: Structural and Statistical Pruning

The researchers evaluated two distinct paradigms for pruning incorrect translations:

Inverse Consultation (IC)

IC is a structural approach based on the assumption that correct translations should have multiple paths back to the source. If a source word leads to multiple synonyms in the pivot language, and those synonyms converge on the same target word, confidence in that translation increases.

Figure 1: Structural representation of the pivot method, illustrating how source words (s) map to target words (t) through various pivot concepts (c) and words (p).

Distributional Similarity (DS)

DS is a statistical approach using comparable corpora (e.g., non-translated news articles in both languages from the same period). It builds "context vectors" for words based on their surrounding vocabulary. If a Basque word and a Spanish word appear in similar contexts (e.g., both near "water" and "sink"), they are likely translations.

Hybrid Approaches

The study tested a Linear Combination of IC and DS scores to leverage their complementary strengths. While IC provides high precision but low recall (failing when dictionaries are sparse), DS is more robust but struggles with highly frequent, polysemous words.

Performance Evaluation

The methods were tested using Basque-English and English-Spanish dictionaries, with a high-quality Basque-Spanish reference dictionary serving as the gold standard. News articles from 2006–2010 provided the comparable corpora for DS context vectors.

Figure 4: F-score performance across word frequencies for various Distributional Similarity (DS) configurations and Inverse Consultation (IC).

As shown in Figure 4, the performance of these methods varies significantly with word frequency. The hybrid approach (Linear Combination) achieved an average F-score of 0.52, outperforming both IC (0.34) and DS (0.47) alone. The methods were particularly effective for mid-frequency words, where DS has sufficient data to form clear context vectors without being muddied by extreme polysemy.

Conclusion and Future Directions

Combining structural and statistical analysis provides a powerful tool for building accurate bilingual dictionaries for low-resource languages. While IC remains a strong choice for high-precision applications, the hybrid approach is superior for large-scale dictionary construction.

Future work may address the "Missing Link" problem—where pivot words are absent from source dictionaries—and incorporate Part-of-Speech (POS) tagging to reduce grammatical noise. The reliance on standard dictionaries and news archives makes this methodology highly portable to other underserved language pairs.

Summary updated

BibTeX
@inproceedings{saralegi2011analyzing,
  author = {Xabier Saralegi and Iker Manterola and Iñaki San Vicente},
  title = {{Analyzing Methods for Improving Precision of Pivot Based Bilingual Dictionaries}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2011},
}
Figure