Automated Calculation of Term Relatedness Weights for Semantic Searches

Abstract

Information retrieval - finding and retrieving relevant sources of data, such as documents or geospatially located records - is a bottleneck in the process of accessing online data. Metadata describing data sources is variable in quality and quantity, textual descriptions are defined by data providers and the terminology they use will not always match search terms, particularly in fields with specialised terminology, such as health. Augmenting the original query with related terms increases the likelihood of matching to relevant metadata. Related terms can be extracted from thesaurus and term definition resources or from the Semantic Web, which defines resources and relationships between them. However, relationships between terms are complicated by multiple interpretations, often dependent upon context (for example, 'sign' may mean a 'road sign' or a 'medical sign', such as fever). Including the strength and/or context of a relationship in a semantic link could help narrow down extra terms to those most relevant to the query. In this paper, methods for automatically calculating the relative strength of relationships between terms were investigated and compared for general and domain-specific terms. Calculations were based on a variety of textual resources including public, crowd-sourced online sources Wikipedia and Google search engine. Measures for term relatedness in a specialist domain were tested using health as a case study. Results show promise for automatic calculation of weights between terms, which can be used to develop weighted graphs for use in semantic searches.