A rule-extraction-based optimization method for feature selection in word sense disambiguation

Feature selection is an important process in classification and pattern recognition and it has a direct influence on the accuracy of classifier. In this study, a new optimization method of feature selection by means of rule extraction is proposed for word sense disambiguation (WSD) of English modal verb “must”. A WSD model with all candidate features for “must” is constructed first with the approach of structural partialordered attribute diagram (SPOAD) and the accuracy of WSD is tested to be 94.5%. Then based on the WSD model and the rule-extraction algorithm, rules for the two senses of “must” are extracted, and accordingly the optimized feature set with only 6 attributes is obtained. The WSD model with the optimized feature set yields a classification accuracy of 97.5%, which is 3% higher than that of the original model. Therefore, it is concluded that the proposed method can optimize the feature set and is effective in dealing with binary classification problems in WSD. It can also be applied to other binary-classifier research and provides valuable reference for feature selection in machine learning and natural language processing. © 2016 ISSN.