Introduction
Word Sense Disambiguation (WSD) remains a fundamental challenge in Natural Language Processing, requiring systems to distinguish between different meanings of a word based on its surroundings. While supervised models excel with hand-labeled data, they suffer from scalability issues. Unsupervised approaches typically leverage knowledge bases like WordNet, which primarily capture paradigmatic relationships (e.g., "is-a"). However, they often miss syntagmatic associations—the natural co-occurrence of words like "eat" and "food." This paper introduces a "Lexical Hub" approach that bridges this gap by integrating associative knowledge from the Edinburgh Association Thesaurus (EAT).
Methodology: The Lexical Hub Model
The proposed model departs from traditional linear lexical chains. Instead, it employs a star-shaped "Lexical Hub" topology where the target word is evaluated against all surrounding context words simultaneously.
The system utilizes two primary knowledge bases:
- WordNet: For hierarchical similarity.
- Edinburgh Association Thesaurus (EAT): To capture human-like stimulus-response associations.
The researchers evaluated six heuristics for scoring word senses. The most effective was Sense Linkage (HSL), which calculates the total number of context words that share any semantic link with a specific sense of the target word, rather than focusing on the strength of the single best link.
Experimental Setup
Evaluation was conducted using the SENSEVAL-2 English lexical sample benchmark, comprising 1,754 noun instances and 1,806 verb instances. To leverage modern similarity metrics, SENSEVAL-2 tags were mapped from WordNet 1.7 to version 2.0.
Discussion of Insights
A key finding of this research is that the quantity of semantic connections (connectivity) is more predictive of the correct sense than the intensity of individual links. As shown in Figure 1, the HSL heuristic consistently outperforms others across various context window sizes for nouns.
Furthermore, the study revealed a "Context Window Asymmetry." Nouns achieved peak accuracy with a broad context window of approximately 60 words, suggesting they are influenced by global topical themes. In contrast, verbs peaked at a much narrower window of 20 words, indicating a stronger dependence on local syntactic neighbors.
Comparative Results
The Lexical Hub system achieved significant improvements over standard baselines and other unsupervised methods. For nouns, the system reached an accuracy of 45.8%, and for verbs, 35.6%. These results are particularly impressive given the random baselines of 18.9% and 6%, respectively.
As illustrated in Figure 5, the HSL model using expanded associative context (SRORRS) outperforms several established systems, including the standard Lesk algorithm and various SENSEVAL-2 competitors, without relying on "most frequent sense" back-off strategies.
Conclusion and Future Work
By combining hierarchical and associative knowledge, the Lexical Hub model provides a robust unsupervised framework for WSD. Future work aims to extend this approach to adjectives and adverbs and integrate syntactic information—such as subject-verb-object relationships—to further refine disambiguation accuracy.
Summary updated