Position Masking for Language Models
This is an incremental improvement for NLP practitioners, offering modest gains in efficiency and performance for language models.
The paper tackles the problem of improving masked language model pre-training by introducing position masking alongside token masking, resulting in a 0.3% performance gain on SQUAD and a 50% reduction in tokens needed for convergence on Graphcore IPU.
Masked language modeling (MLM) pre-training models such as BERT corrupt the input by replacing some tokens with [MASK] and then train a model to reconstruct the original tokens. This is an effective technique which has led to good results on all NLP benchmarks. We propose to expand upon this idea by masking the positions of some tokens along with the masked input token ids. We follow the same standard approach as BERT masking a percentage of the tokens positions and then predicting their original values using an additional fully connected classifier stage. This approach has shown good performance gains (.3\% improvement) for the SQUAD additional improvement in convergence times. For the Graphcore IPU the convergence of BERT Base with position masking requires only 50\% of the tokens from the original BERT paper.