NotaBene е електронно списание за философски и политически науки. Повече за нас

Towards Uniform Semantic Interpretation


Dimitar Hrisimirov Popov

 

The aim of this paper is to create a theoretical outlier of hybris artificial intelligence system, which uses both symbolic logic and sub-symbolic methodologies for learning to create unified approach for discovery of semantic correlation of entities based on similarity measures and their relationship rules that governs them. Since this is extremely broad formalization, we will concentrate our effort on a particular data structure, which has powerful enough representation and could be used to transform comfortably enough, other representations like text, images, tables, namely graphs.

Why graphs? Graphs are non-linear data structures and as such has expressive power to capture almost every hidden relation from another type of data. For example, in natural language process a textual representation of character string (e.g. “Dimitar lives in Ravda”, (1)) could be transformed as graph that contains set of related entities and their relation. If we consider the example in (1) a graph representation would look like the figure 1 below. In the simple schematics entities {Dimitar, Ravda} are members of entity set containing all the entities for given graph. In graph theory these are called nodes of edges of a graph. The link between the two nodes is the semantic relation between two nodes, in the simple example on Figure 1, the relation is named lives_in . In a graph, relations are called vertexes or links, vertexes could be directed, if they are depicted with an arrow.

 

Figure 1. Graph with two nodes representing semantic relation: lives_in


This depicting arrow carries information about the orientation of the edges. There are various graph relationship mining methods, which formalize mathematical model to discover relationships in a complex graph structure. Based on those methods and the represented graphs, one may wish to discover what are the relation to yet unidentified link in the graph or classified a node label altogether. In figure 2 we could see a graph, that has one unlabeled node connected with one unlabeled vertex to the node Ravda.

 

Read the full article here