DWIE: An entity-centric dataset for multi-task document-level information extraction

Abstract This paper presents DWIE, the ‘Deutsche Welle corpus for Information Extraction’, a newly created multi-task dataset that combines four main Extraction (IE) annotation subtasks: (i) Named Entity Recognition (NER), (ii) Coreference Resolution, (iii) Relation (RE), and (iv) Linking. DWIE is conceived as an entity-centric describes interactions properties of conceptual entities on level complete document. contrasts with currently dominant mention-driven approaches start from detection classification named entity mentions in individual sentences. Further, presented two challenges when building evaluating IE models it. First, use traditional mention-level evaluation metrics NER RE tasks DWIE dataset can result measurements dominated by predictions more frequently mentioned entities. We tackle this issue proposing new entity-driven metric takes into account number compose each predicted ground truth Second, document-level annotations require to transfer information between located different parts document, well tasks, joint learning setting. To realize this, we propose graph-based neural message passing techniques mention spans. Our experiments show improvement up 5.5 F 1 percentage points incorporating graph propagation our model. demonstrates DWIE’s potential stimulate further research networks representation IE. make publicly available at https://github.com/klimzaporojets/DWIE .