A Review on Semantic Similarity

A Review on Semantic Similarity

Montserrat Batet (Universitat Rovira i Virgili, Spain) and David Sánchez (Universitat Rovira i Virgili, Spain)
DOI: 10.4018/978-1-4666-5888-2.ch746

Chapter Preview



According to the knowledge source used to extract semantic evidences to guide the similarity assessment, measures can be grouped in several families.

Ontology-based measures estimate the similarity of two concepts according to the structured knowledge offered by ontologies. They can be classified into:

  • 1.

    Edge-counting measures evaluate the number of semantic links separating the two concepts in the ontology (Leacock & Chodorow, 1998; Li, et al., 2003; Rada, et al., 1989; Wu & Palmer, 1994). For example, Wu and Palmer compute similarity according to the number of taxonomic links (N1 and N2) between the two concepts (a, b) and their taxonomic ancestor, and the number of links (N3) of that ancestor and the root node of the ontology, which acts as a normalization factor.


  • 2.

    Feature-based measures rely on the amount of overlapping ontological features (e.g., taxonomic ancestors, concept descriptions, etc.) between the compared concepts (Petrakis, et al., 2006; Rodríguez & Egenhofer, 2003; Sánchez, et al., 2012a). For example, Sanchez and Batet measure the similarity between two concepts a and b according to the inverse non-linear ratio between their disjoint and total taxonomic ancestors T(a) and T(b).


  • 3.

    Measures based on quantifying the amount of information (i.e., Information Content (IC)) that concepts have in common (Jiang & Conrath, 1997; Lin, 1998; Resnik, 1995). Commonalties are extracted from the common taxonomic ancestors of the compared concepts, whereas the informativeness of concepts is computed either extrinsically from the concept occurrences in a corpus (Jiang & Conrath, 1997; Lin, 1998; Resnik, 1995) or intrinsically, according to the number of taxonomical descendants and/or ancestors modeled in the ontology (Sánchez & Batet, 2012; Seco, et al., 2004). For example, Lin measures the similarity between concepts a and b according to the ratio between the informativeness of their Least Common Subsumer (LCS(a,b)) and the informativeness of each individual concept.


Key Terms in this Chapter

Semantic Dissimilarity or Distance: It is the inverse to similarity. It is computed from the disjoint semantic evidences extracted from one or several knowledge sources. Dissimilarity or distance measures can be converted to similarities through a linear transformation.

Information Content (IC): It measures the informativeness of an entity as the inverse of its probability of occurrence. Thus, general entities are assumed to provide less information than more specialized ones, since the former are more likely to appear in a discourse.

Semantic Similarity: It estimates the taxonomic resemblance of two terms, based on the evaluation of the common semantic evidences extracted from one or several knowledge sources (e.g., textual corpus, thesaurus, taxonomies/ontologies, etc.)). For example, moose and reindeer are similar because they are ruminant mammals of the family Cervidae.

Ontology: It is a structured knowledge source that explicitly and consensually represents the concepts and the semantic interrelations of a domain, with the purpose of sharing and re-using knowledge. Ontologies should, at least, contain a taxonomical backbone. Non-taxonomic relationships, properties stating concept attributes, logical axioms defining restrictions or instances representing specific objects can be also included.

Tagged Corpora: It is a set of textual resources in which words and phrases have been annotated (i.e., disambiguated) according to their conceptualizations modeled in one or several ontologies.

Semantic Relatedness: It estimates the closeness of the semantic relationship between two terms considering both taxonomic and non-taxonomic (e.g., meronymy, antonymy, functionality, cause-effect, etc.) knowledge. For example, reindeer is related to antlers or radiotherapy is related to cancer .

Least Common Subsumer (LCS): It is the most specific common ancestor of two concepts found in a given ontology. Semantically, it represents the commonality of the pair of concepts. For example, the LCS of moose and kangaroo in WordNet is mammal.

Complete Chapter List

Search this Book: