Detecting Character References in Literary Novels using a Two Stage Contextual Deep Learning approach

Current approaches for Named Entity Recognition detect reoccuring names within longer stories again and again in every sentence by modelling the task as a sequence prediction. Character references do sometimes occur in contexts where they can easily be detected (e.g. close to communication verbs) while other contexts might not yield good information. This results in a tagged document in which entities are not detected consistently.

In this work we experimented with a two stage approach using contextual deep learning using BiLSTM-CRFs to leverage this problem.