Textual CBR Authoring using Case Cohesion

During the design phase of a textual case-based reasoning (TCBR) system, decisions have to be made regarding the internal representation of the cases and the similarity metrics. Such decisions have a significant impact on the performance of the resulting TCBR system. We believe that guidance should be provided to the designer during the authoring process. Unfortunately most of the metrics proposed in the CBR literature either require subjective human evaluation or do not apply when both problem and solution parts of the cases are textual in nature. In this paper, we propose a metric, case cohesion that can be used to provide guidance in the configuration of a case base and in the selection of appropriate similarity schemes. Case cohesion is defined from the neighborhood of textual cases and allows to estimate whether both problem and solution descriptions are well aligned. Our experimental results indicate that case cohesion can discriminate among the performances of different retrieval schemes and can also help in the unsupervised selection of the parameters required for these retrieval metrics.