22.5.2 Gene Normalization

Having identified all the mentions of biological entities in a text, the next step is to map them to unique identifiers in databases or ontologies. This task has been most heavily studied for genes, where it is known as gene normalization. Some of the complexities of the problem come from high degrees of variability in the realization of the names of specific entities in naturally-occurring text; the nature of the problem was first delineated by Cohen et al. (2002). In that work a standard discovery procedure from descriptive linguistics was used to determine what sorts of variability in gene names can be ignored, and what sorts must not be ignored. More recently, Morgan et al. (2007) have shown how linguistic characteristics of community-specific gene-naming conventions affect the complexity of this task when the normalization of genes from varying species is attempted. Gene normalization can be considered a type of word sense disambiguation task, midway between a targeted WSD task and an all-words WSD task.

An important thread of work on this problem involves mapping named entities to biomedical ontologies, especially the Gene Ontology (Ashburner et al., 2000). This has proven considerably more challenging; terms in the Gene Ontology tend to be long, to have many possible lexical and syntactic forms, and to sometimes require significant amounts of inference. ? (?) introduce this ontology from the perspective of computational lexical semantics and review much of the named entity recognition work that has involved it.