PMI-Masking: Principled masking of correlated spans

Yoav Levine, Barak Lenz, Opher Lieber, Omri Abend, Kevin Leyton-Brown, Moshe Tennenholtz, Yoav Shoham

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review


Masking tokens uniformly at random constitutes a common flaw in the pretraining of Masked Language Models (MLMs) such as BERT. We show that such uniform masking allows an MLM to minimize its training objective by latching onto shallow local signals, leading to pretraining inefficiency and suboptimal downstream performance. To address this flaw, we propose PMI-Masking, a principled masking strategy based on the concept of Pointwise Mutual Information (PMI), which jointly masks a token n-gram if it exhibits high collocation over the corpus. PMI-Masking motivates, unifies, and improves upon prior more heuristic approaches that attempt to address the drawback of random uniform token masking, such as whole-word masking, entity/phrase masking, and random-span masking. Specifically, we show experimentally that PMI-Masking reaches the performance of prior masking approaches in half the training time, and consistently improves performance at the end of pretraining.
Original languageEnglish
Title of host publicationICLR 2021
Subtitle of host publicationInternational Conference on Learning Representations
Place of PublicationVienna, Austria
Number of pages14
StatePublished - 2021
EventInternational Conference on Learning Representations, ICLR 2021 - Vienna, Austria
Duration: 3 May 20217 May 2021
Conference number: 9


ConferenceInternational Conference on Learning Representations, ICLR 2021
Internet address


  • Language modeling
  • BERT
  • pointwise mutual information


Dive into the research topics of 'PMI-Masking: Principled masking of correlated spans'. Together they form a unique fingerprint.

Cite this