Learning with joint cross-document information via multi-task learning for named entity recognition

In information extraction, named entity recognition (NER) aims to locate named entities in unstructured text and classify them into predefined categories. Most existing methods for NER are sentence-level approaches, which only leverage the context information within a sentence and may cause inconsistent entity prediction. Several researchers have identified latent relationships between sentences in a document and therefore present numerous document-level frameworks that utilize the context information in a document. However, these frameworks cannot establish the correlation between sentences in different documents. To address this problem, we present a cross-document NER model that builds internal relationships for each token among all its multiple occurrences in different documents. A cross-document attention module is designed to calculate the cross-document representations. Furthermore, we add a multiclassification auxiliary task to utilize the coarse-grained entity information. The multi-objective optimization is employed through weighted summation and an autonomous approach utilizing the homoscedastic uncertainty of each task. Extensive experimental results on various data sets clearly demonstrate the effectiveness of our proposed approach, and our model achieves better results than the sentence-level and document-level NER models.