Pre-trained Models Perform the Best When Token Distributions Follow Zipf's Law.

Yanjin He, Qingkai Zeng, Meng Jiang

Browse the full EMNLP paper archive.