Contextes multilingues alignés pour la désambiguïsation sémantique : une étude expérimentale

Boxing Chen, Meriam Haddara, Olivier Kraif, Grégoire Moreau de Montcheuil, Marc El-Bèze

Presented at TALN 2005, this research explores the utility of parallel corpora—texts translated across multiple languages—to address Word Sense Disambiguation (WSD). By "triangulating" meanings across English, French, and Spanish, the authors demonstrate significant improvements in a system's ability to select the correct sense of a word in context.

The Challenge of Semantic Ambiguity

Word Sense Disambiguation (WSD) involves determining which specific meaning of a word is intended in a given context (e.g., distinguishing "bank" as a financial institution from a riverbank). While monolingual context provides essential clues, it is often insufficient for high-precision tasks.

However, translation inherently involves disambiguation. If the English word "interest" is translated as "intérêt" (financial) in one instance and "curiosité" (hobby) in another, the translation serves as a semantic label. The core challenge lies in effectively utilizing these "noisy" translation signals, particularly when alignments are imperfect or translations are idiomatic.

Methodology

The study evaluates three primary approaches:

  1. Unsupervised Disambiguation: Utilizing semantic networks (EuroWordNet), the system compares possible senses of an English word and its aligned French translation. A sense is selected if both words point to the same "Interlingual Index" (ILI).
  2. Supervised Disambiguation (KNN): A K-Nearest Neighbors algorithm is adapted to include "projected" contexts. When an English word is aligned with a French word, the French context provides additional evidence for the similarity function.
  3. The "Triangulation" Criterion: To mitigate errors from poor alignments, the authors introduce a third language (Spanish) as a pivot. A Dice coefficient measures the overlap of Spanish translations for both English and French words. High overlap indicates a "semantically reliable" alignment, warranting greater weight for multilingual evidence.

Experimental Setup

The researchers utilized a multilingual corpus including Flaubert’s Madame Bovary and Darwin’s The Voyage of the Beagle (approx. 200,000 words per language). Evaluation was conducted on 15 English lemmas (6 verbs, 5 nouns, 4 adjectives) across 2,886 manually disambiguated contexts, using EuroWordNet as the primary lexical resource.

Key Results

Sense Reduction

In unsupervised testing, the multilingual approach proved highly effective at narrowing down semantic possibilities. The system successfully eliminated a significant portion of incorrect senses, as detailed in the table below.

Average proportion of senses eliminated and percentage of units fully disambiguated.

Accuracy Boost

The integration of multilingual evidence, particularly when filtered through the triangulation criterion, yielded measurable improvements over monolingual baselines.

Comparative accuracy results for monolingual vs. multilingual WSD approaches.

As shown in the results, the monolingual English baseline achieved 62.2% accuracy. While a systematic bilingual approach (English + French) provided only a negligible gain (62.4%), the fine-grained selection using the Spanish pivot (Opt.Fin) reached 66.0% accuracy. This confirms that the selective application of multilingual data is superior to the indiscriminate use of all available alignments.

Conclusion

The study highlights "reciprocal revelation"—the way different languages partition semantic space—as a powerful tool for WSD. While bilingual clues offer high precision, their coverage is often limited. The most effective strategy is a hybrid approach that relies on monolingual context by default but leverages multilingual evidence when triangulation confirms alignment reliability. Future work may involve using these techniques to automatically generate high-quality training data for languages lacking extensive manual annotations.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{chen2005contextes,
  author = {Boxing Chen and Meriam Haddara and Olivier Kraif and Grégoire Moreau de Montcheuil and Marc El-Bèze},
  title = {{Contextes multilingues alignés pour la désambiguïsation sémantique : une étude expérimentale}},
  booktitle = {TALN (Articles Courts)},
  year = {2005},
  url = {https://openreview.net/forum?id=qEeDjYfjbDW},
}
Figure