Semantic Networks

Representation of Semantic Networks

Semantic networks require three constituent parts: (1) a syntax that specifies the types of nodes and edges that can be considered; (2) specification of the meaning or semantics that the nodes, links, and entire network can represent; and (3) inference rules. Data structures representing semantic networks comprise, at a minimum, nodes that are referred to as concepts and edges between the concepts. Concepts are abstract representations of the ideas, thoughts, and units of knowledge and meaning that people conceive in their minds. Concepts can also be referents of objects. When concepts have a representation in natural language, the respective word or sequence of words are used as labels for nodes in semantic networks. Examples for concepts are “social network analysis,” “collaboration,” and “community of practice.” If links in semantic networks are typed, the link type explicates the nature of the relationship between the connected nodes. Otherwise, links represent or establish some meaningful association between pairs of concepts. Link formation in semantic networks heavily depends on the type of input data and intended use of the network. Beyond these constraints, representations of semantic networks can range from strictly formalized to informal.

Across many different approaches, the smallest structured unit of a semantic network is typically a triplet that describes the source or subject, predicate, and object or target of an action. This basic structure is used, for instance, to describe who did what to whom for representing event data and who said what to whom for communication data. Such triplets can further be enhanced with background information, such as spatial and temporal data, and attributes of nodes and links.

Semantic networks result from a translation or transformation of the data from which the networks were constructed. Translations convert natural language input into isomorphic, structured representations that are used as input to inference mechanisms. This approach has been pursued in artificial intelligence technology as one attempt to understand natural language and is usable for small amounts of input. Transformations are abstraction processes that are used to preserve and reveal the entities and relations that are explicitly or implicitly represented in the input data. The goal with transformations is the reduction of the dimensionality of the input data in order to capture the [Page 767]relevant structural interrelations and to make network data available for inference.

...