Event extraction on Indonesian news article using multiclass categorization

Event extraction identifies who did what, when, where, why, and how, which is known as 5W1H. We aim to investigate event extraction on Indonesian news articles as multiclass-categorization problem, and apply statistical learning-based approach that treats event extraction as a sequence labeling problem under BIO (Begin Inside Outside) labeling scheme. Each token of input text will be classified into one of 13 predefined classes. Our contributions are providing 5W1H corpus, and the best technique to build model of event extraction. Our experiments show that C4.5 is better than AdaboostM1 although Adaboost can identify minority labels better than C4.5. In addition, C4.5 with all features gave the best Fmeasure of 0.666.