Named Entity Recognition Approaches Applied to Legal Document Segmentation

Document Segmentation is a method of dividing a document into smaller parts, known as segments, which share similarities that allow machines to distinguish between them. It might be useful to classify these segments, making it a problem with two steps: (I) the extraction of the segments; and (II) the annotation of these segments. The Named Entity Recognition problem's goal is to identify and classify entities within a text, having also to deal with those two questions: extraction and classification. In this study, we tackle the problem of Document Segmentation and the annotation of these segments through NER approaches, using CRF, CNN-CNN-LSTM and CNN-biLSTM-CRF models. The study is focused on Brazilian legal documents, proposing a data set of 127 annotated Portuguese texts from the Official Gazette of the Federal District, published between 2001 and 2015. The experiments were made using word-based and sentence-based models, with CRF sentence-based model showing the best results.