Abstract
Word Sense Disambiguation (WSD) has long been a central challenge in Natural Language Processing, often forcing a trade-off between high-precision local methods and low-precision global methods. This paper presents a "semi-complete" disambiguation algorithm that prioritizes accuracy over total coverage. By iteratively applying seven heuristic procedures—leveraging WordNet hierarchies and the SemCor corpus—the algorithm identifies the correct senses for approximately 55% of nouns and verbs in open text with 91% precision. This approach provides a reliable semantic foundation for applications such as information retrieval and semantic indexing.
Introduction
Determining the specific meaning of a word based on its context—Word Sense Disambiguation (WSD)—is frequently described as the "holy grail" of computational linguistics. The difficulty lies in the inherent ambiguity of natural language; for instance, the word "bank" may refer to a financial institution or a geographical feature depending on the surrounding text.
Historically, WSD research has been bifurcated. One school of thought focuses on high-precision methods that only target a small, pre-selected set of words. The other attempts "all-words" disambiguation, which labels every token in a text but often suffers from significantly lower accuracy. This paper proposes a middle ground: a "semi-complete" approach. The authors argue that for many downstream tasks, such as Information Retrieval, a highly reliable partial map of a text's meaning is more valuable than a complete map riddled with errors. The goal is to label as many words as possible while maintaining a strict threshold for certainty.
Methodology
The core of the proposed system is an iterative algorithm that manages the transition of words between two states: the Set of Ambiguous Words (SAW) and the Set of Disambiguated Words (SDW). Initially, all nouns and verbs in a given text are placed in the SAW. The algorithm then cycles through seven distinct heuristic procedures, moving words to the SDW as their senses are resolved.
The Seven Heuristic Procedures
The procedures are ordered generally from highest to lowest certainty, allowing the system to establish "semantic anchors" early in the process:
- Named Entity Recognition (NER): The system identifies proper nouns such as people, places, and organizations (e.g., tagging "John Doe" as a person).
- Monosemous Words: Words that possess only a single possible sense within the WordNet dictionary are automatically moved to the SDW.
- Contextual Pattern Matching: The algorithm searches for word pairs that appear in the SemCor corpus (a manually sense-tagged dataset). If a specific word pair frequently appears with a particular sense in SemCor, that sense is assigned.
- Semantic Connectivity (Distance 0): The system identifies words in the SAW that belong to the same "synset" (synonym set) as a word already established in the SDW.
- Mutual Connectivity (Distance 0): The system looks for two ambiguous words within the same text that belong to the same synset, disambiguating them simultaneously.
- Semantic Connectivity (Distance 1): This procedure utilizes WordNet’s hierarchical structure to find direct "hypernym/hyponym" (parent/child) relationships (e.g., connecting "oak" to "tree") between a word in the SAW and one in the SDW.
- Mutual Connectivity (Distance 1): The system identifies two ambiguous words in the SAW that share a direct parent/child relationship in the WordNet hierarchy.
This iterative structure is critical. Once a word is disambiguated by an "easy" rule—such as being monosemous—it provides a reliable context that helps the algorithm resolve more complex ambiguities in subsequent passes.
Experimental Setup and Data
The researchers utilized two primary linguistic resources for both the execution and evaluation of the algorithm:
- WordNet: A comprehensive lexical database of English that organizes words into synonym sets and defines semantic relationships.
- SemCor: A subset of the Brown Corpus where words have been manually linked to their specific WordNet senses. SemCor provided the contextual patterns for Procedure 3 and served as the ground truth for evaluating the algorithm's performance.
Results and Discussion
The algorithm was evaluated using randomly selected files from the SemCor corpus. The results demonstrated a successful execution of the "precision over recall" philosophy:
- Coverage (Recall): The system successfully disambiguated approximately 55% of all nouns and verbs in the test set.
- Accuracy (Precision): For the subset of words the system chose to label, it achieved a precision of 91%.
These findings suggest that the iterative use of semantic anchors is a robust strategy. By leveraging WordNet’s hierarchy, the system understands relationships beyond simple synonymy, allowing it to build a "semantic skeleton" of a document that is highly resistant to false positives.
Conclusion and Future Work
The semi-complete disambiguation algorithm represents a significant step toward practical WSD for open text. While the current version is limited to nouns and verbs and leaves roughly 45% of those words unlabelled, the 91% precision rate offers a level of reliability previously difficult to achieve in general-purpose disambiguation.
Future iterations of this work aim to expand the algorithm’s scope to include adjectives and adverbs. Furthermore, the authors are exploring the application of this method to semantic indexing, which would enable search engines to process queries based on conceptual meaning rather than simple keyword matching, potentially transforming the efficiency of information retrieval systems.
Summary updated