Named Entity Recognition is the basis for many Natural Language Processing technology applications. The traditional sequence labeling-based approach fails to recognize nested entities due to its decoding method of continuous label prediction. Currently, scholars have proposed a span-based named entity recognition method, which extracts bidirectional semantics through character embedding combined with BiLSTM (Bi-directional Long Short-Term Memory) networks and predicts entity boundaries, which realizes the work of recognizing multiple types of entities. However, this character-level approach is limited by the Chinese word-segmentation problem and the BiLSTM network cannot be parallelized. To address the above problems, a Chinese Named Entity Recognition model based on bidirectional self-attention and span is proposed. The model uses a positional embedding-optimized Transformer to construct a bidirectional semantic extraction structure and combines the bidirectional self-attention mechanism to fuse character embedding and external lexicon information to solve the word segmentation problem. Finally, the convolutional module and biaffine module are used for joint decoding. The proposed model experiments on ordinary entity datasets Weibo, Resume, and nested entity dataset Cnerta with F1 values of 73.57, 96.56, and 85.41%.