Pre-Training Transformers as Energy-Based Cloze Models

TL;DR

Electric is an energy-based model for text representation that assigns likelihood scores to tokens without masking, offering efficient re-ranking and insights into ELECTRA's pre-training.

Contribution

It introduces Electric, a novel energy-based cloze model that improves likelihood scoring and provides a clearer understanding of ELECTRA's pre-training process.

Findings

Electric outperforms traditional models in re-ranking speech recognition outputs.

It is faster than masked language models for likelihood scoring.

Electric effectively transfers to downstream NLP tasks.

Abstract

We introduce Electric, an energy-based cloze model for representation learning over text. Like BERT, it is a conditional generative model of tokens given their contexts. However, Electric does not use masking or output a full distribution over tokens that could occur in a context. Instead, it assigns a scalar energy score to each input token indicating how likely it is given its context. We train Electric using an algorithm based on noise-contrastive estimation and elucidate how this learning objective is closely related to the recently proposed ELECTRA pre-training method. Electric performs well when transferred to downstream tasks and is particularly effective at producing likelihood scores for text: it re-ranks speech recognition n-best lists better than language models and much faster than masked language models. Furthermore, it offers a clearer and more principled view of what ELECTRA learns during pre-training.