ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators

TL;DR: This paper proposed a more sample-efficient pre-training task called replaced token detection, which corrupts the input by replacing some input tokens with plausible alternatives sampled from a small generator network and then predicts whether each token in the corrupted input was replaced by a generator sample or not.

Abstract: While masked language modeling (MLM) pre-training methods such as BERT produce excellent results on downstream NLP tasks, they require large amounts of compute to be effective. These approaches corrupt the input by replacing some tokens with [MASK] and then train a model to reconstruct the original tokens. As an alternative, we propose a more sample-efficient pre-training task called replaced token detection. Instead of masking the input, our approach corrupts it by replacing some input tokens with plausible alternatives sampled from a small generator network. Then, instead of training a model that predicts the original identities of the corrupted tokens, we train a discriminative model that predicts whether each token in the corrupted input was replaced by a generator sample or not. Thorough experiments demonstrate this new pre-training task is more efficient than MLM because the model learns from all input tokens rather than just the small subset that was masked out. As a result, the contextual representations learned by our approach substantially outperform the ones learned by methods such as BERT and XLNet given the same model size, data, and compute. The gains are particularly strong for small models; for example, we train a model on one GPU for 4 days that outperforms GPT (trained using 30x more compute) on the GLUE natural language understanding benchmark. Our approach also works well at scale, where we match the performance of RoBERTa, the current state-of-the-art pre-trained transformer, while using less than 1/4 of the compute.