Transformation-based learning meets frequent pattern discovery

Transformation-based learning (TBL) and frequent pattern discovery (FPD) are two popular research paradigms, one from the domain of empirical natural language processing , the second from the eld of data mining. This paper describes how Eric Brill's original TBL algorithm can be improved via incorporation of FPD techniques. The algorithm B-Warmr is presented that upgrades TBL to rst-order logic and speeds up the search for transformations, also in the original propositional case. We demonstrate some scaling properties of B-Warmr and discuss how the algorithm can be tuned to generate ((rst-order) decision lists. We also propose a new method, Disjunctive Transformation-Based Learning (DTBL) that combines the advantages of TBL and decision lists.