Fuzzy feature analysis for unsupervised knowledge discovery in narrative texts
P. Perrin, Frederick E. Petry · 2002
This work demonstrates an application of fuzzy sets to the feature analysis problem to enhance the ability to find interesting knowledge in textual databases. We believe that the extraction of a satisfiable set of features yields to represent texts with a sufficiently structured and interesting system in a finite space leading to the discovery of interesting knowledge. We discuss an approach to extract a set of good features to effectively represent each text and best ensure that interesting knowledge can be found by applying any data mining technique. Feature analysis consists of automatically extracting context-dependent measurements to sense events described in narrative texts and transform them into a finite set of features maximizing the cohesion of the system (minimizing the entropy) with an acceptable level of interestingness. Preliminary experiments on real databases of dictated psychiatric evaluation notes were promising. Further experiments are in progress to further demonstrate the relationships between a situation's structure and the level of guarantee to find interesting information.