Semantic Similarity in a Taxonomy: An Information-Based Measure and its Application to Problems of Ambiguity in Natural Language

Analysis

Why This Paper Matters

Resnik's 1999 paper introduced a principled, information-theoretic approach to measuring semantic similarity within a taxonomy, moving beyond the simplistic edge-counting methods that dominated earlier work. By grounding similarity in the shared information content of concepts—quantified as the negative log probability of their least common subsumer—the measure aligns more closely with human intuition and empirical judgments. This shift from structural to probabilistic similarity was pivotal because it leveraged corpus statistics to capture the actual usage and informativeness of concepts, making the metric both theoretically sound and practically effective.

The paper's significance is amplified by its direct application to ambiguity resolution, a core challenge in natural language processing. Resnik demonstrated that taxonomic similarity could be harnessed to disambiguate both syntactic (e.g., prepositional phrase attachment) and semantic (e.g., word sense) ambiguities, providing a unified framework. This work bridged lexical semantics and statistical NLP, influencing later developments in word sense disambiguation, ontology alignment, and semantic relatedness measures.

Technical Contributions

• Information-based similarity measure: Defines similarity between two concepts as the information content of their lowest common subsumer in a taxonomy, where information content is -log(p(c)) and p(c) is estimated from corpus frequencies.
• Empirical validation: The measure is evaluated against a benchmark set of human similarity judgments (e.g., Miller-Charles dataset), showing higher correlation than edge-counting baselines.
• Ambiguity resolution algorithms: Two algorithms are presented—one for syntactic ambiguity (PP attachment) and one for semantic ambiguity (word sense disambiguation)—both using taxonomic similarity to select the most coherent interpretation.
• Corpus-based probability estimation: Introduces a method to compute concept probabilities by aggregating word frequencies from a corpus, enabling the information content to reflect real-world usage.

Results

• The information-based similarity measure achieved a correlation of 0.79 with human judgments on the Miller-Charles dataset, compared to 0.66 for edge-counting, demonstrating a clear improvement.
• In syntactic ambiguity resolution (PP attachment), the algorithm achieved 84.5% accuracy on a standard test set, outperforming prior statistical methods.
• For semantic ambiguity (word sense disambiguation), the approach showed consistent gains over baseline methods, particularly for nouns in the taxonomy.

Significance

Resnik's work established information-theoretic semantic similarity as a cornerstone of lexical semantics and NLP. It provided a rigorous, data-driven alternative to hand-crafted semantic networks and inspired a generation of similarity measures (e.g., Lin, Jiang-Conrath). The paper's influence extends beyond NLP to fields like bioinformatics, where taxonomic similarity is used for gene ontology comparisons, and to information retrieval, where it enhances query expansion and document clustering. By demonstrating that a simple, theoretically motivated metric could outperform more complex alternatives, Resnik set a standard for principled similarity computation that remains relevant today.