Unified Named Entity Recognition as Word-Word Relation Classification

So far, named entity recognition (NER) has been involved with three major types, including flat, overlapped (aka. nested), and discontinuous NER, which have mostly studied individually. Recently, a growing interest built for unified tackling the above jobs concurrently one single model. Current best-performing methods mainly include span-based sequence-to-sequence models, where unfortunately former merely focus on boundary identification latter may suffer from exposure bias. In this work, we present novel alternative by modeling NER as word-word relation classification, namely W^2NER. The architecture resolves kernel bottleneck of effectively neighboring relations between words Next-Neighboring-Word (NNW) Tail-Head-Word-* (THW-*) relations. Based W^2NER scheme develop neural framework, in is modeled 2D grid word pairs. We then propose multi-granularity convolutions better refining representations. Finally, co-predictor used to sufficiently reason perform extensive experiments 14 widely-used benchmark datasets overlapped, (8 English 6 Chinese datasets), our model beats all current top-performing baselines, pushing state-of-the-art performances NER.