Transformer Language Models without Positional Encodings Still Learn Positional Information.

Adi Haviv, Ori Ram, Ofir Press, Peter Izsak, Omer Levy

Browse the full EMNLP paper archive.