A Route-aware Model for Entity Recognition with Diverse Structures

Summary: Discontinuous entities are widely existing in actual texts. However, most named entity recognition (NER) systems deal only with the flat entities and ignore others. Some NER models claim to recognize these entities but suffer from some level of ambiguity. To address this issue, we propose a novel route-aware model to unambiguously extract entities with all kinds of structures. We identify the span of all the entities and the adjacent matrix of the sentence respectively and decode all the routes within the extracted span. Experiments show that our route-aware model achieves 40.2 F1 scores in CADEC dataset and 36.8 in DDI dataset on NER with complex structures and comparable performance on NER with all the entities.