Minimizing Human Labelling Effort for Annotating Named Entities in Historical Newspaper

Abstract

To accelerate the annotation of named entities (NEs) in historical newspapers like Sarawak Gazette, only two choices are possible: an automatic approach or a semi-automatic approach. This paper presents a fully automatic annotation of NEs occurring in Sarawak Gazette. At the initial stage, a subset of the historical newspapers is fed to an established rule-based named entity recognizer (NER), that is ANNIE. Then, the preannotated corpus is used as training and testing data for three supervised learning NER, which are based on Naïve Bayes, J48 decision trees, and SVM-SMO methods. These methods are not always accurate and it appears that SVM-SMO and J48 have better performance than Naïve Bayes. Thus, a thorough study on the errors done by SVM-SMO and J48 yield to the creation of ad hoc rules to correct the errors automatically. The proposed approach is promising even though it still needs more experiments to refine the rules.