Description and Evaluation of Semantic Similarity Measures Approaches

1. INTRODUCTION

• Similarity is measured between ontology concepts or terms, while relatedness is broader and does not always imply similarity.The paper illustrates the distinction with Student and Professor as related but not similar.Source: §1 · p. 1
• Ontology-based mapping compares concepts across knowledge sources, including different ontologies, to support knowledge sharing.The introduction gives KIMP-to-MeSH mapping as an example for standardized medical terms.Source: §1 · p. 1
• The paper examines four categories of semantic similarity measures and compares and evaluates approaches within each category.Source: §1 · p. 1
• The paper proceeds from ontology examples to measure categories, evaluation, and conclusion.Source: §1 · p. 1

2. EXAMPLES OF ONTOLOGIES USED WITH SEMANTIC SIMILARTY MEASURES

• Examples include WordNet, SENSUS, and Cyc.Source: §2 · p. 2
• The listed examples also include UMLS, SNOMED, MeSH, GO, and STDS.Source: §2 · p. 2
• The following section classifies ontologies as general-purpose or domain-specific.Source: §2 · p. 2

2.1.1 Wordnet

• WordNet groups nouns, verbs, adjectives, and adverbs into synsets organized in taxonomic hierarchies.Its recognized relations include Is-A and Part-Of.Source: §2.1.1 · p. 2
• The figure presents a fragment of WordNet’s hypernym hierarchy.Source: Figure 1 · p. 2

2.1.3 Cyc KB

• Cyc is an encyclopedic repository focused primarily on commonsense knowledge, represented through terms and assertions.The knowledge base contains over 500,000 terms and seven million assertions.Source: §2.1.3 · p. 2

2.2.1 UMLS

• UMLS is a multilingual biomedical and health metathesaurus built from controlled vocabularies, code sets, classifications, and term lists.It covers over one million biomedical concepts and five million concept names; its components include the Metathesaurus and Semantic Network.Source: §2.2.1 · p. 2

2.2.2 SNOMED

• SNOMED attaches terms to concept codes organized in a directed acyclic graph (DAG) for clinical-health terminology.It provides a common language for capturing, sharing, and aggregating health data across specialties and care sites.Source: §2.2.2 · p. 3

2.2.3 MeSH

• MeSH organizes medical and biological descriptors taxonomically (26,853 in 2013), while GO has three controlled vocabularies for gene-product attributes.GO covers cellular components, biological processes, and molecular functions.Source: §2.2.3 · p. 3, §2.2.4 · p. 3

2.2.5 STDS

• STDS transfers earth-referenced spatial data between dissimilar computer systems and specifies its conceptual model, content, structure, and format.Source: §2.2.5 · p. 3

• Ontology taxonomies contribute path length, depth, and local density factors that affect semantic distance, though not significantly.Source: §3 · p. 3
• Measures may use shared characteristics, taxonomic position and links, or information content, and may differ in symmetry and value ranges.Source: §3 · p. 3
• The paper classifies semantic measures into four main categories.Source: §3 · p. 3

3.1 Structure-based measures

• Structure-based measures compute similarity from ontology hierarchy paths, with closer concepts linked by more links and closer taxonomic relations.Source: §3.1 · pp. 3–4

3.1.1 Shortest Path [15]

• Shortest Path uses the minimum number of links between concepts, normalized by the maximum path length in the taxonomy.Source: §3.1.1 · p. 4

3.1 Structure-based measures

• Weighted Links [16] replaces link counts with summed weights based on hierarchy depth, taxonomy density, and parent-child connotation strength.Source: §3.1.2 · p. 4

3.1.3 Hirst and St-Onge Measure (HSO) [17]

• HSO measures relatedness using path distance, direction changes, and whether the path is allowable without digressing from the source concept’s meaning [17].Source: §3.1.3 · p. 4

3.1.4 Wu and Palmer [19]

• Wu and Palmer [19] uses distances from each concept to their closest common ancestor and that ancestor’s depth from the root.Source: §3.1.4 · p. 4
• TBK extends Wu and Palmer to address cases where a nearby concept scores above a descendant, penalizing similarity for concepts outside the same hierarchy.Source: §3.1.5 · p. 4

3.1.6 Li et al. [21]

• Li et al. combines shortest-path length with the depth of the most specific common concept in a nonlinear function, with parameters set to 0.2 and 0.6.Source: §3.1.6 · p. 5

3.2 Information Content Measures

• Information-content measures use corpus term frequency to value concepts and use the most informative shared subsumer when concepts share multiple parents.Source: §3.2 · pp. 5–6

3.2.3 Lin et al. [24]

• Lin et al. [24] combines shared information as a ratio rather than a difference and is reported to rank similarities better than Resnik.Source: §3.2.3 · p. 6

3.2.4 Jiang & Conrath [25]

• Jiang and Conrath [25] defines distance as the sum of the two concepts’ information content minus that of their most informative subsumer.Source: §3.2.4 · p. 6

3.3 Feature-Based Measures

• Feature-based measures represent each term with properties or features and compare terms using glosses or their hierarchical relations.Source: §3.3 · p. 6

3.3.1 Tversky [26]

• Tversky increases similarity with common features and decreases it with differences, while disregarding taxonomy position and information content.Source: §3.3.1 · p. 6

3.3.2 X-Similarity [27]

• X-Similarity [27] matches words in WordNet definitions and semantic neighborhoods, computing set similarities separately by relationship type.Source: §3.3.2 · p. 6

3.3.3 Rodriguez et al. [28]

• Rodriguez et al. [28] combines similarities between entity classes’ synonym sets, distinguishing features, and semantic neighborhoods using weights that sum to one.Source: §3.3.3 · p. 7
• The paper concludes that knowledge feature-based measures exploit more semantics than edge-counting approaches.Source: §3.3.3 · p. 7

3.4 Hybrid Measures

• Hybrid measures combine structural characteristics such as path length, depth, and local density with other approaches, and are reported as more accurate than basic edge-counting measures.Basic edge-counting measures depend on empirically adjusting weights for the ontology and input terms.Source: §3.4 · p. 7

3.4.1 Knappe [29]

• Knappe’s measure uses generalization and specification information across potentially multiple paths, with costs shaped by generalization, edge traversal, and specialization.The measure assigns a generalization influence value p between 0 and 1.Source: §3.4.1 · p. 7

3.4.2 Zhou et al. [30]

• Zhou et al.’s measure combines information content and path measures, with k manually adapted and its endpoints yielding path-based or information-content-based forms.At k=1 the formula is path-based; at k=0 it is information-content-based.Source: §3.4.2 · p. 7

• The evaluation relies on benchmarks built from word pairs whose similarity was judged by people, addressing the subjectivity of similarity judgments.The paper describes the Rubenstein and Goodenough study of 65 noun pairs rated on a 0.0-to-4.0 scale.Source: §4 · p. 7
• 0.87 accuracy is reported for hybrid approach, the highest among the compared hybrid measures despite its complexity and weighting parameters.The approach exploits relative depth.Source: Table 6 · p. 8, §4 · p. 8
• 0.7373 vs. 0.70: intrinsic information-content measures based on concept hyponyms are reported as more accurate than corpus-based approaches.The paper says corpus dependency constrains the applicability of classic information-content measures.Source: Table 6 · p. 8, §4 · p. 8
• 0.59 is reported for the shortest-path measure, while Wu and Palmer and Leacock and Chodorow reach 0.74 using relative taxonomy depth.The paper attributes the latter improvement to incorporating relative depth; it also reports that Li et al. combine path length and concept depth in a weighted, nonlinear manner for an improved result.Source: Table 6 · p. 8, §4 · p. 8
• 0.71-0.74: feature-based methods resemble structure-based measures and rely on concept features or non-taxonomic relations that the paper describes as secondary in ontologies.The paper says these methods use partially modeled knowledge and need more research to outperform edge-counting approaches.Source: §4 · p. 8

• The paper examines recognized semantic similarity measures, organizing their typology and key properties under each category.Source: §5 · p. 8
• The paper presents a practical comparison of the measures using two widely used benchmarks to assess accuracy.Source: §5 · p. 8
• The analyses are intended to help researchers and practitioners choose the measure that best fits a real application’s requirements.Source: §5 · p. 8