Learning Named Entity Classifiers Using Support Vector Machines

Traditional methods for named entity classification are based on hand-coded grammars, lists of trigger words and gazetteers. While these methods have acceptable accuracies they present a serious draw-back: if we need a wider coverage of named entities, or a more domain specific coverage we will probably need a lot of human effort to redesign our grammars and revise the lists of trigger words or gazetteers. We present here a method for improving the accuracy of a traditionally-built named entity extractor. Support vector machines are used to train a classifier based on the output of an existing extractor system. Experimental results show that this approach can be a very practical solution, increasing precision by up to 11.94% and recall by up to 27.83% without considerable human effort.