DOZEN: Cross-Domain Zero Shot Named Entity Recognition with Knowledge Graph
Hoang-Van Nguyen, Francesco Gelli, Soujanya Poria
Abstract
With the new developments of natural language processing, increasing attention has been given to the task of Named Entity Recognition (NER). However, the vast majority of work focus on a small number of large-scale annotated datasets with a limited number of entities such as person, location and organization. While other datasets have been introduced with domain-specific entities, the smaller size of these largely limits the applicability of state-of-the-art deep models. Even if there are promising new approaches for performing zero-shot learning (ZSL), they are not designed for a cross-domain settings. We propose Cross Domain Zero Shot Named Entity Recognition with Knowledge Graph (DOZEN), which learns the relations between entities across different domains from an existing ontology of external knowledge and a set of analogies linking entities and domains. Experiments performed on both large scale and domain-specific datasets indicate that DOZEN is the most suitable option to extracts unseen entities in a target dataset from a different domain.