A multi-aspect comparison study of supervised word sense disambiguation

Area of Science:

• Computational linguistics
• Natural Language Processing
• Machine Learning

Background:

• Supervised word sense disambiguation (WSD) is crucial for accurately interpreting word meanings in context.
• Comparing WSD performance across different domains, such as biomedical and general English, is essential for developing robust algorithms.

Purpose of the Study:

• To investigate the interdependencies of various factors in supervised WSD.
• To compare the effectiveness of supervised WSD in the biomedical domain versus the general English domain.

Main Methods:

• Utilized three datasets: biomedical abbreviations, general biomedical terms, and general English.
• Implemented and compared three machine learning algorithms: Naïve Bayes (NBL), traditional decision lists (TDLL), adapted decision lists (ODLL), and mixed supervised learning (MSL).
• Evaluated six feature representations and five window sizes (2-10 words).

Main Results:

• Sufficient sense-tagged instances (dozens per sense) are necessary for effective supervised WSD.
• Collocations and neighboring words are effective contextual features.
• Optimal window sizes differ: larger for biomedical terms with unrelated senses, moderate (4-10) for general English.
• Mixed supervised learning (MSL) demonstrated stable and superior performance across all tested datasets.

Conclusions:

• Different components of supervised WSD are interconnected and influence overall performance.
• The experimental framework can guide the selection of optimal WSD classifiers for specific ambiguous terms.