Exploiting global contextual information for document-level named entity recognition

Named entity recognition (NER, also known as entity chunking/extraction) is a fundamental sub-task of information extraction, which aims at identifying named entities from an unstructured text into pre-defined classes. Most of the existing works mainly focus on modeling local-context dependencies in a single sentence for entity type prediction. However, they may neglect the clues derived from other sentences within a document, and thus suffer from the sentence-level inherent ambiguity issue, which may make their performance drop to some extent. To this end, we propose a Global Context enhanced Document-level NER (GCDoc) model for NER to fully exploit the global contextual information of a document in different levels, i.e., word-level and sentence-level. Specifically, GCDoc constructs a document graph to capture the global dependencies of words for enriching the representations of each word in word-level. Then, it encodes the adjacent sentences for exploring the contexts across sentences to enhance the representation of the current sentence via the specially devised attention mechanism. Extensive experiments on two benchmark NER datasets (i.e., CoNLL 2003 and Onenotes 5.0 English dataset) demonstrate the effectiveness of our proposed model, as compared to the competitive baselines.