SpanBERT is a pre-training method designed to improve the representation and prediction of spans of text. It extends BERT by masking contiguous random spans rather than individual tokens and training the span boundary representations to predict the entire content of the masked span. This approach significantly outperforms BERT and its tuned baselines on span selection tasks such as question answering and coreference resolution. SpanBERT achieves 94.6% and 88.7% F1 scores on SQuAD 1.1 and 2.0, respectively, and a new state-of-the-art 79.6% F1 on the OntoNotes coreference resolution task. It also demonstrates strong performance on the TACRED relation extraction benchmark and gains on the GLUE benchmark. The key contributions of SpanBERT include the span masking scheme and the novel span-boundary objective (SBO), which encourages the model to store span-level information at the boundary tokens.SpanBERT is a pre-training method designed to improve the representation and prediction of spans of text. It extends BERT by masking contiguous random spans rather than individual tokens and training the span boundary representations to predict the entire content of the masked span. This approach significantly outperforms BERT and its tuned baselines on span selection tasks such as question answering and coreference resolution. SpanBERT achieves 94.6% and 88.7% F1 scores on SQuAD 1.1 and 2.0, respectively, and a new state-of-the-art 79.6% F1 on the OntoNotes coreference resolution task. It also demonstrates strong performance on the TACRED relation extraction benchmark and gains on the GLUE benchmark. The key contributions of SpanBERT include the span masking scheme and the novel span-boundary objective (SBO), which encourages the model to store span-level information at the boundary tokens.