Paper

[MPNet: Masked and Permuted Pre-training for Language Understanding]

1. 핵심 요약

MLM(e.g. BERT)

PLM(e.g. XLNet)

MPNet = MLM(BERT) + PLM (XLNet)

2. MPNet

2.1 Background

MLM in BERT

Given Sentence : $x= [x_1, ..., x_n]$

$\kappa$ : A set of masked poisition

A set of masked tokens : $x_{\kappa}$