Wordnet creation and extension made simple: A multilingual lexicon-based approach using wiki resources

Valérie Hanoka, Benoît SagotHAL 2012

TL;DR

This paper presents a straightforward method for building and expanding multilingual wordnets by leveraging readily available lexical data from Wikipedia and Wiktionaries. By constructing a large, multilingual translation graph and using a back-translation strategy, the authors effectively enrich existing wordnets, demonstrating its utility in extending the French wordnet WOLF.

What Problem Does This Paper Address?

Creating comprehensive lexical resources like wordnets for languages other than English is a significant challenge due to the scarcity of existing digital databases. Manual creation is prohibitively expensive and time-consuming. While automatic methods exist, they often struggle with the nuances of multilingualism and the inherent conceptual differences between languages. This paper tackles the problem of efficiently creating and extending multilingual wordnets by proposing a method that capitalizes on the vast, interconnected knowledge found in wiki resources, aiming to mitigate the limitations of traditional approaches that rely on scarce bilingual lexicons.

Methods

The core of the paper's methodology lies in two main steps:

  1. Building a Multilingual Translation Graph: The authors construct a directed graph where nodes represent (word/term, language) pairs. This graph is populated by extracting translation and synonym information from Wiktionaries and Wikipedia articles across numerous languages. The use of directed edges is crucial to prevent the propagation of errors. To ensure reliability, the graph is filtered to remove redundant translations and pairs with low connectivity (nodes must have a minimum degree of 3).
  2. Filling or Extending Synsets using Back-Translation: The algorithm operates on a per-synset basis, taking as input existing synset-aligned wordnets in multiple languages. It starts with a set of "gold literals" (existing synset members with a high weight) and iteratively expands this set by:
  3. Querying the translation graph to find candidate literals (translations of existing literals).
  4. Ranking these candidates using a back-translation heuristic. This involves checking if a candidate's translation back into the original language of a synset member receives a high score.
  5. Assigning weights to candidate literals based on their back-translation quality, with bonuses for strong back-translations and penalties for weak ones.
  6. Merging the new, weighted candidate literals with the existing synset members to form an updated synset. This process can be repeated for further refinement.

Datasets

  • Wiktionaries: Used as a primary source for extracting translation and synonym pairs across 18 languages (Czech, Dutch, English, French, German, Hebrew, Indonesian, Italian, Japanese, Korean, Polish, Portuguese, Romanian, Russian, Slovak, Spanish, Swedish, and Turkish).
  • Wikipedia: Used in conjunction with Wiktionaries, leveraging inter-wiki links.
  • Princeton WordNet (PWN) 2.0: Used as the English seed wordnet.
  • BalkaNet wordnets: Used for Czech, Bulgarian, and Romanian, which are aligned to PWN 2.0.
  • WOLF (WordNet of French): Used as the target resource for extension and evaluation.

Key Insights and Ideas

  • Leveraging Wiki-Scale Data: The paper demonstrates that the sheer volume and interconnectedness of data in Wikipedia and Wiktionaries can compensate for the lack of high-quality, curated bilingual lexicons, which are often a bottleneck in multilingual NLP tasks.
  • The Power of Back-Translation: The heuristic of using back-translations to rank candidate literals is a clever way to infer the semantic equivalence of words across languages, effectively filtering out spurious translations.
  • Multilingualism as a Strength: The approach thrives on multilingualism. By considering translations across many languages, the system can find more robust connections and overcome language-specific limitations.
  • Simple but Effective: The proposed methodology is relatively simple to implement and computationally feasible, making it accessible for researchers and developers.

Key Results

  • The filtering process on the raw extracted data reduced the number of translation pairs from over 9.9 million to about 1.58 million, and synonymy pairs from over 748,000 to about 366,000.
  • The algorithm was applied twice to extend the French wordnet WOLF.
  • After applying the algorithm and retaining candidates with a score greater than 30, 10,568 (literal, synset) candidates were produced, with 6,119 (58%) not previously present in WOLF.
  • A manual evaluation of 400 candidates showed a precision of 74.1% when retaining all candidates (t=30, nmax=∞).
  • More restrictive parameter settings (e.g., t=50, nmax=1) achieved a precision of 86% over 3,353 candidates, with 1,601 new candidates having a precision of 82%.
  • The optimal parameter values identified were (t=30, nmax=3), yielding a precision of 74.8% on 10,403 candidates, suggesting that this approach can add approximately 7,781 correct candidates to WOLF.
  • The method proved particularly effective for high and medium frequency polysemous words, which are often challenging for previous wordnet extension approaches.

Limitations and Future Work

  • Reliance on Wiki Quality: The quality of the generated wordnet is inherently dependent on the quality and coverage of the source Wiktionaries and Wikipedia articles. Noisy or incomplete data can lead to errors.
  • Parameter Tuning: The performance is sensitive to the chosen thresholds (t) and maximum candidate limits (nmax), requiring careful tuning for optimal results.
  • Future Work:
    • Investigating the use of larger Wiktionaries from other languages or improving the translation graph quality for languages with smaller wiki resources.
  • Exploring more sophisticated methods for computing candidate literal scores, potentially by performing a "march" within the translation graph.
  • Upgrading the translation graph's quality, possibly by incorporating techniques from Mausam et al. (2009).

Summary updated

Abstract

International audience
BibTeX
@inproceedings{hanoka2012wordnet,
  author = {Valérie Hanoka and Benoît Sagot},
  title = {{Wordnet creation and extension made simple: A multilingual lexicon-based approach using wiki resources}},
  booktitle = {HAL},
  year = {2012},
}

Loading related content...

Figure