Researchers proposed POSPAN, a framework for position-constrained span masking in language model pre-training, enhancing span-level masking methods by incorporating position constraint distributions. This improvement was validated through experiments on NLU benchmarks, outperforming existing methods.