Abstract
Word Sense Disambiguation (WSD) is a fundamental challenge in Natural Language Processing. This paper presents LEXAS (LEXical Ambiguity-resolving System), a supervised learning approach that utilizes an exemplar-based algorithm to integrate diverse linguistic features. By combining part-of-speech, local collocations, and syntactic relations, LEXAS achieves significant performance gains over previous methods and demonstrates scalability to large-scale, fine-grained sense inventories like WordNet.
Introduction
Determining the correct sense of a word in context is essential for tasks ranging from machine translation to information retrieval. Historically, WSD research faced two major hurdles: feature isolation, where models relied on a single type of linguistic clue, and scalability, where systems were limited to small sets of words with coarse sense distinctions.
Hwee Tou Ng and Hian Beng Lee addressed these issues by framing WSD as a supervised learning problem. Their system, LEXAS, employs a k-nearest neighbor (exemplar-based) algorithm to classify word senses by comparing new instances to a stored memory of tagged examples.
The LEXAS System: Exemplar-Based Learning
Unlike rule-based systems, LEXAS does not derive explicit generalizations. Instead, it stores "exemplars" from a training corpus. When presented with a new word occurrence, the system calculates the distance between the new instance and stored exemplars using the Value Difference Metric (VDM). This allows the system to handle symbolic features and weigh their relative importance dynamically.
Knowledge Sources and Feature Extraction
LEXAS integrates four primary types of linguistic information:
- Part of Speech (POS) & Morphology: Captures the grammatical context within a $\pm3$ word window.
- Unordered Surrounding Words: A "bag-of-keywords" approach identifying significant words within the entire sentence.
- Local Collocations: Specific word sequences immediately adjacent to the target word.
- Verb-Object Relations: Identifies the governing verb for noun targets, providing a strong syntactic signal.
As shown in the analysis of individual features, local collocations prove to be the most predictive single source of information, achieving a mean accuracy of 80.2%. However, the core strength of LEXAS lies in the integration of all these sources, which yields higher overall accuracy than any single feature set alone.
Experimental Evaluation
The authors evaluated LEXAS on two major benchmarks. The first was the "Interest" corpus, where LEXAS achieved approximately 87.4% accuracy, a substantial improvement over previous state-of-the-art results (typically around 78-79%).
The second evaluation involved a massive, manually tagged corpus of 192,800 word occurrences using WordNet 1.5 senses. This experiment proved that supervised WSD could scale to the fine-grained distinctions of a comprehensive dictionary.
As illustrated in the test results for the Brown Corpus (BC50) and Wall Street Journal (WSJ6), LEXAS consistently outperformed the "Most Frequent Sense" heuristic. On the WSJ6 set, it reached 68.6% accuracy compared to the baseline's 63.7%.
Conclusion
The success of LEXAS highlights the importance of multi-source feature integration in WSD. By moving beyond isolated features and leveraging a large-scale tagged corpus, the authors demonstrated that high-accuracy, dictionary-scale disambiguation is achievable. While the "data bottleneck" of manual annotation remains a challenge, this work established a robust framework for supervised lexical disambiguation that influenced subsequent decades of NLP research.
Summary updated