Conceptual Blending Approach to Concept Generation

Do Thanh Mai

Abstract

This research formalizes Conceptual Blending framework to create a support for concept generation process. We tie together two research areas: Conceptual Blending introduced by Fauconnier and Turner in the early 1990s and Conceptual Graph introduced by John Sowa in the late 1980s. Conceptual Blending framework is credited for its simplicity and ubiquity in explaining events happening in the creative mind. Conceptual Graphs possesses graph-based reasoning mechanism and good quality of visualization. By formalizing Conceptual Blending based on Conceptual Graph, we would like to integrate their advantages in the field of computer-aid innovation and research to generate new concepts based on existing knowledge.

Firstly, this research proposes several extensions of Conceptual Graphs into Extended Simple Graph (ESG) to represent input spaces. Namely, we elaborate 1-to-n relation between entity and concept by co-reference, synonym and multi-perspective link. These links allow different concept nodes to represent one entity, which makes input process more flexible in multi-user environment. The links also enable merging and splitting these nodes, which reduces complexity of domain knowledge and facilitate graphical display of information within a computer screen. Other changes include the representation of markers as a specified concept instead of embedded list in concept nodes, which simplifies concept representation and data input procedure. Secondly, we employ ESG in a knowledge representation, named Flexi-rep, to formulize Conceptual Blending Framework. We also point out a set of elementary operations that perform conceptual blending and four characteristics of an input space namely dynamical modifiability, partial order, graph control and variety by perspectives. Finally, we propose Birdcage integration network to model two common problems in engineering namely Substitution Problem and Usage Context Problem.

In conclusion, we have proven that Conceptual Graph is a suitable representation of input space in Conceptual Blending, and that a support for human concept generation is feasible. We have also shown how to localize Conceptual Blending to produce falsifiable theory. The current work is evaluated theoretically without computational implementation due to lack of Conceptual Blending and Conceptual Graph benchmarks. However, this thesis is a kick off for computer-aid concept generation future research. Making a bold step towards the exciting domain which is intersection of cognitive modelling, psychology, creative theories and artificial intelligence, our vision is when computer can compute information like it does with numbers, and blending is just an operator among many to be defined.