Rule-based approach to computational stylistics

Decision algorithms correspond to the rule-based approach to classification and pattern recognition problems. While to shorten the processing time we need as few constituent decision rules as possible, when their number is too low it may lead to a poor performance of the classifier. The decision rules can be found by providing the minimal cover of the training samples, by calculating rules with some genetic algorithms, by the exhaustive search for all rules. This last option offers the widest choice of rules, which enables tailoring the final algorithm to the task at hand, yet this is achieved by the additional cost of rule selection process. Usually there are assumed some measures indicating the quality of individual decision rules. The paper presents a different procedure, which is closer to feature reduction. In the first step there are selected condition attributes that are discarded, then the rules that contain conditions on these attributes are removed from the algorithm. The classifier performance is observed in the domain of computational stylistics, which is a study on characteristics of writing styles.