Abstract
J. Daudé, L. Padró, G. Rigau. 1999 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora.
1999.
Introduction
Building a comprehensive Multilingual Lexical Knowledge Base (LKB) is a significant goal for natural language processing, essential for tasks like machine translation and cross-lingual search. However, manual construction is prohibitively expensive. The primary obstacle to automation is polysemy—the fact that a single word often has multiple meanings. When mapping a Spanish taxonomy to English WordNet, a single word might have over 15 candidate translations. This paper presents an automated method to link these hierarchies while preserving structural integrity.
Methodology: Relaxation Labeling
The researchers utilize Relaxation Labeling (RL), an iterative optimization technique originally developed for computer vision. RL treats the mapping task as a constraint satisfaction problem, focusing on local structural consistency to find a global solution.
The process involves:
- Variables and Labels: Each node in the Spanish taxonomy is a variable, and its potential labels are the English WordNet synsets provided by a bilingual dictionary.
- Structural Constraints: The algorithm leverages hypernym/hyponym (parent/child) relationships. It favors mappings where the hierarchical relationship in the source language is mirrored in the target language.
- Iterative Refinement: Starting with initial weights, the system iteratively adjusts the probability of each translation based on how well it satisfies the structural "family tree" constraints until the system converges.
Datasets
The study employed three main resources:
- Spanish Taxonomies: Automatically extracted from the Diccionario General Ilustrado de la Lengua Española (DGILE), specifically in the domains of
animal,food,cognition, andcommunication. - English WordNet: The target conceptual taxonomy.
- Bilingual Dictionary: An integrated resource containing approximately 125,000 translations.
Key Insights
- Structure as Context: The core finding is that while individual word translations are ambiguous, the surrounding hierarchy provides enough context to disambiguate senses without complex semantic analysis.
- Domain Sensitivity: The method performs better on concrete taxonomies (animals, food) than abstract ones (cognition), as concrete hierarchies are more stable across languages.
- Robustness: The algorithm effectively filters out "noise" from automatically generated source taxonomies and irrelevant dictionary entries.
Key Results
The relaxation labeling approach achieved high accuracy, with an average precision exceeding 80%. In concrete domains like food and animal, precision reached between 93% and 96%.
Despite an average of 15.8 possible WordNet connections per Spanish sense, the RL algorithm successfully narrowed these down to the correct synsets. Furthermore, the system demonstrated high coverage, proposing assignments for up to 88% of nodes in certain test sets.
Conclusion and Future Work
The primary limitation remains the coverage of bilingual dictionaries. Future research may explore inferring missing translations from neighboring nodes and incorporating additional relationship types, such as meronymy (part-of). Additionally, this technique could be used to map different versions of the same lexical resource, facilitating easier maintenance of databases like WordNet.
Summary updated