Corpus and Baseline Model for Domain-Specific Entity Recognition in German

Abstract

Transfer Learning approaches are a promising means to analyze low-resource domain specific texts. The German SmartData corpus is the first German corpus, annotated with entities from different domains, and thus allows to investigate transfer learning approaches for Named Entity Recognition (NER) on different domains. In order to prepare such investigations, this work includes a thorough analysis of the SmartData corpus, and a revision w.r.t. annotations and the split into training and test data, considering the distribution of document and entity types. Based on that a baseline model for NER using BiLSTM-CRF neural networks including hyperparameter optimization is presented.