Fig. 1.

nala method active learning process. Each blue box represents an iteration state of the nala method. The method and the iteration training sets are implemented in parallel. The previous iteration method (nala_t-1) is used to automatically annotate unseen documents. Selected documents with outstanding errors are reviewed manually and added to the iteration training set t. New features are evaluated in 5-fold cross validation and the method is retrained with all previous sets (nala_t). At the end, the sum of iteration training sets without IDP4 form the nala_training corpus. The final nala method is trained on nala_training (only) and evaluated against the nala_known and nala_discoveries corpora