Projection in conceptual graphs using neural networks

Conceptual graphs are a symbolic representation of knowledge, using the notions of concepts and concept relations (of relations). Each concept in a conceptual graph represents a single idea: an object, attribute, action, etc. Conceptual graphs have proven useful in linguistic processing and as an organizing principle for databases and intelligent systems. One possible scenario of use is to construct a database of conceptual graphs representing discrete units of knowledge. These conceptual graphs can then be retrieved by the use of a query template conceptual graph, using a matching process called projection. Because of the cost of matching conceptual graphs using serial processing, it is interesting to consider the use of highly parallel computing approaches such as neural networks as part of the query process. The query graph and a fact graph from the database can be matched with respect to relations in a single step; however, because of the need to compare concept supertypes, the concept matching is an iterative process. This paper explores a technique for training a neural network to do supertype matching in a single step, using the transitive supertype graph