First-order Frequent Patterns in Text Mining

In this paper a universal framework for mining long first-order frequent patterns in text data is presented. It consists of RAP, an ILP system for mining maximal first-order frequent patterns, and two types of predefined background knowledge. Two methods of using generated patterns for solving text mining tasks are described: propositionalization and CBA (class based association). A new variant of the CBA rule based classifier is proposed. The framework is used for solving three text mining tasks: information extraction from biomedical texts, context-sensitive text correction of English and morphological disambiguation of Czech. The distributed mining of frequent patterns is described and its influence on mining in text is discussed. It is shown that frequent patterns as new features for propositionalization usually provide better results than CBA