Named entity recognition (NER) is an essential task in information extraction and knowledge graph construction. Numerous traditional methods for NER are sentence-level, which only utilize all the information in one sentence and predict entities inconsistently sometimes. Recently, researchers observe that the relationship between sentences in a document is helpful for the NER task, therefore they consider to conduct documentlevel approaches which capture context information within a document. Though these methods are effective and have been widely used, they ignore correlations between sentences in different documents. To make full use of the cross-document context information, we design an attention mechanism to model the semantic association between occurrences of the same word in different documents. In addition, we find that it is extremely helpful to add a multi-classification auxiliary task, which focuses on coarse-grained entity information. The multi-objective optimization is implemented by weighted summation and an autonomous approach using the homoscedastic uncertainty of each task. Extensive experiments in different datasets show that our approach is effective, and our models achieve better results than the sentence-level and document-level NER approaches.