Abstract
Philippe Muller, Philippe Langlais. Actes de la 17e conférence sur le Traitement Automatique des Langues Naturelles. Articles courts.
2010.
Introduction
Automatically constructing a thesaurus is a foundational challenge in Natural Language Processing. While humans intuitively grasp synonymy, computational systems often struggle to distinguish true synonyms from related terms like antonyms or hypernyms. This paper evaluates two primary paradigms for synonym extraction in French: distributional analysis and the "mirror" approach based on translation.
Methodology
The researchers compared two distinct methodologies using large-scale French resources:
- Distributional Approach (vsn/voisins): Based on Lin (1998), this method analyzes syntactic triplets (governor, relation, dependent) from the French Wikipedia (approx. 200M words) to calculate similarity scores between words sharing contexts.
- Mirror Approach (miroir): Following Dyvik (2002), this method utilizes a bilingual parallel corpus (TransSearch, 8.3M sentence pairs). It posits that if word A translates to B, and B translates back to C, then A and C are likely synonyms.
- Combination: An intersection of both methods was also evaluated to see if agreement between the two increased accuracy.
Performance was measured against DicoSyn, a gold-standard reference merging seven classic French synonym dictionaries.
Results and Analysis
The experimental results consistently demonstrate that the translation-based mirror approach outperforms the distributional method for both nouns (N) and verbs (V).
As shown in the performance curves, the 'miroir' approach maintains a higher F-score across various thresholds and 'nbest' candidate selections. A critical finding is the impact of word frequency; filtering for high-frequency words (freq > 5000) significantly improves reliability.
The highest F-score of 0.273 was achieved by combining both methods for verbs with a high-frequency filter. Despite this being the state-of-the-art for these methods, the absolute scores remain relatively low, suggesting that these tools are currently best suited as "suggestion engines" for human lexicographers rather than fully autonomous systems.
Conclusion
The "mirror" hypothesis proves to be a powerful semantic filter, leveraging cross-lingual information to identify synonyms more accurately than monolingual distributional context alone. While polysemy remains a significant hurdle—as these methods aggregate all word senses into a single score—the complementarity of combining syntactic context with translation data provides the most robust path forward for automatic lexical resource generation.
Summary updated