Automatic recognition of structures in obituaries

Abstract

Obituaries are a less common text type in research that contains a lot of information about people, events in history and culture. The information that can be obtained by zoning such obituaries enables new research, e.g., in social studies. Our work focuses on the question if the structuring of obituaries is possible and viable. Therefore we created a corpus for this work containing 20058 obituaries of which 1008 were annotated manually by us. We implemented four models, a CNN text classifier and three variations of a Bi-LSTM sequence labeler, to see if the zoning procedure is possible and which among the models performs best for this task. The CNN text classifier showed the most promising results together with the variant of the Bi-LSTM model using a Bag-of-Word model.