A Unified Multilingual Semantic Representation of Concepts

José Camacho-Collados, Mohammad Taher Pilehvar, Roberto NavigliACL 2015 · DOI

Abstract

Traditional word embeddings often conflate multiple meanings of a word into a single vector, leading to ambiguity in Natural Language Processing (NLP) tasks. This paper presents MUFFIN (Multilingual, UniFied and Flexible INterpretation), a method for representing individual word senses in a unified, language-independent semantic space. By combining structural knowledge from the BabelNet semantic network with distributional statistics from Wikipedia, MUFFIN achieves state-of-the-art performance in cross-lingual semantic similarity and Word Sense Disambiguation (WSD), effectively bridging the gap between different languages.

Introduction

In NLP, teaching machines to distinguish between the various senses of a word—such as "bank" as a financial institution versus a river side—remains a foundational challenge. Most traditional Vector Space Models (VSMs) represent words rather than specific senses, resulting in "meaning conflation." Furthermore, high-performing models are often English-centric, making cross-lingual tasks inefficient. MUFFIN addresses these issues by creating a "unified" space where concepts are represented identically regardless of the source language (e.g., "apple," "pomme," and "Apfel" all map to the same fruit concept).

Methodology

The MUFFIN framework leverages BabelNet, a multilingual semantic network linking WordNet and Wikipedia, to construct sense-specific representations through a two-step process:

  1. Sub-corpus Extraction: For a given concept (e.g., the bird "crane"), a specialized mini-corpus is built using its Wikipedia page, linked pages, and immediate semantic relatives (hypernyms and hyponyms) within the BabelNet graph.
  2. Lexical Specificity: Instead of simple frequency counts, the model employs lexical specificity—a statistical measure based on the hypergeometric distribution—to identify words that appear in the sub-corpus significantly more often than in the general Wikipedia corpus.
The MUFFIN architecture for generating vector representations from BabelNet synsets and Wikipedia sub-corpora.

The breakthrough of MUFFIN lies in its Unified representation. While the Lexical version uses words as dimensions, the Unified version uses concepts (BabelNet synsets). By clustering related words under common parent concepts, MUFFIN creates a representation that is entirely independent of the source language.

Evaluation and Results

The researchers evaluated MUFFIN across several benchmarks, including monolingual similarity (RG-65), cross-lingual similarity, and Word Sense Disambiguation (SemEval-2013 and SemEval-2007).

Cross-Lingual Performance

MUFFIN demonstrated exceptional performance in cross-lingual tasks. In the cross-lingual RG-65 tasks, it achieved high Pearson correlations, such as 0.83 for French-English, significantly outperforming "pivot" systems that rely on machine translation.

Comparison of MUFFIN against various pivot-based measures in cross-lingual semantic similarity tasks.

Word Sense Disambiguation (WSD)

In the SemEval-2013 multilingual WSD task, MUFFIN proved highly competitive, outperforming the Babelfy system on Spanish and German datasets. When integrated with the supervised IMS system, it established new performance benchmarks, reaching an F1-score of 68.5 on the SemEval-2007 dataset.

Conclusion

MUFFIN successfully integrates structural and distributional knowledge to provide a flexible, multilingual representation of word senses. By using concepts as dimensions, it eliminates the need for language-specific models or translation steps in cross-lingual comparisons. While dependent on the coverage of BabelNet and Wikipedia, MUFFIN's performance across diverse tasks highlights its potential as a robust foundation for multilingual NLP.

Summary updated

BibTeX
@inproceedings{camachocollados2015unified,
  author = {José Camacho-Collados and Mohammad Taher Pilehvar and Roberto Navigli},
  title = {{A Unified Multilingual Semantic Representation of Concepts}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2015},
  doi = {10.3115/v1/p15-1072},
}
Figure