SpanMRC: Query with Entity Length for MRC-Based Named Entity Recognition

The task of named entity recognition (NER) is to identify predefined types of entities from the given text. When entities overlap, this problem is referred to as nested named entity recognition (Nested NER). Recent researches have treated the NER task as a machine reading comprehension (MRC) task and proposed a lot of effective methods based on to extract flat and nested entities simultaneously. However, traditional MRC-based methods to rely on entity types in the design of questions are not justified due to neglecting the possibility of nested relationships between entities of the same type, and their efficiency is directly affected by the size of entity types. Motivated by the above problems, we apply the MRC framework to the NER task from the perspective of entity lengths but not entity types and propose a novel MRCbased method for NER called SpanMRC. Our approach can tackle flat and nested entities, and it is equally effective for nested entities of the same type. Moreover, the construction of questions is independent of the entity types, which can effectively improve the efficiency of the algorithm as the number of entity types increases. Extensive experiments demonstrate the superiority of SpanMRC in terms of entity recognition accuracy and algorithmic efficiency.