An Incrementally Trainable Statistical Approach to Information Extraction Based on Token Classification and Rich Context Models