Automatic WordNet Mapping Using Word Sense Disambiguation

Daniel M. BikelEMNLP 2000 · OpenReview

Abstract

This paper presents the automatic construction of a Korean WordNet from pre-existing lexical resources. A set of automatic word sense disambiguation (WSD) techniques is described for linking Korean words collected from a bilingual machine-readable dictionary (MRD) to English WordNet synsets. We demonstrate how individual linking provided by each WSD method is combined via a decision tree to produce a high-precision Korean WordNet for nouns.

Introduction

Ontologies like WordNet are essential for modern Natural Language Processing (NLP) tasks, such as information retrieval and machine translation. However, while the English WordNet is mature and comprehensive, creating similar resources for other languages is incredibly expensive and time-consuming when performed manually by experts.

The primary challenge in automating this process is semantic ambiguity. When a Korean word is looked up in a bilingual dictionary, it often yields English translations with multiple meanings (synsets) in WordNet. For example, the English word "bush" has five distinct synsets ranging from a shrub to a mechanical lining. Determining which specific English "sense" matches the original Korean word is the core "mapping problem" addressed in this research.

Methodology

The authors propose a multi-heuristic approach to disambiguate these mappings. Instead of relying on a single algorithm, the system utilizes six distinct "experts" (heuristics) to evaluate potential links:

  1. Maximum Similarity: Assumes that multiple English translations of a single Korean word should be semantically similar within the WordNet hierarchy.
  2. Prior Probability: Favors "monosemous" translations (words with only one meaning), as they are statistically less likely to be incorrect.
  3. Sense Ordering: Leverages frequency data from the SemCor corpus, prioritizing more common word senses.
  4. IS-A Relation: Validates links by checking if hierarchical relationships in the source language (e.g., "sparrow" is a "bird") are mirrored in WordNet.
  5. Word Match: Compares dictionary definitions of the Korean word with WordNet definitions of the English synset for overlapping content words.
  6. Co-occurrence: Analyzes semantic tags to determine how often certain concepts appear together in Korean definition sentences.
Mapping Korean words to WordNet synsets via English translations

To synthesize these inputs, the authors employed the C4.5 Decision Tree algorithm. The model was trained on a manual evaluation set of 3,260 senses to classify potential links as either "Link" (correct) or "Discard" (incorrect). This ensemble approach allows the system to choose the most reliable heuristic based on the specific context of the word.

Results and Discussion

The decision tree approach achieved a 93.59% precision rate, significantly outperforming simpler combination methods like summation or logistic regression. This confirms that the relationship between linguistic heuristics is non-linear and benefits from machine learning classification.

Performance comparison of different combination methods

Using this method, the authors generated a preliminary Korean WordNet containing 21,654 senses for 17,696 Korean nouns. While the precision was high, the coverage stood at 77.12%, indicating that the system prioritized accuracy by discarding ambiguous senses that could not be confidently mapped.

Limitations and Future Work

Currently, the study focuses exclusively on nouns. Mapping verbs and adjectives presents higher levels of polysemy and different hierarchical structures that require further investigation. Additionally, the quality of the resulting WordNet remains dependent on the breadth of the underlying bilingual dictionary. Future work will focus on expanding the system to other parts of speech and refining heuristics to improve coverage without sacrificing precision.

Summary updated

Abstract

This paper presents the automatic construction of a Korean WordNet from pre-existing lexical resources. A set of automatic WSD techniques is described for linking Korean words collected from a bilingual MRD to English WordNet synsets. We will show how individual linking provided by each WSD method is then combined to produce a Korean WordNet for nouns.
BibTeX
@inproceedings{bikel2000automatic,
  author = {Daniel M. Bikel},
  title = {{Automatic WordNet Mapping Using Word Sense Disambiguation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2000},
  url = {https://openreview.net/forum?id=Bk4gAZfOWH},
}
Figure