Edit Aware Representation Learning via Levenshtein Prediction

Abstract

We propose a novel approach that employs token-level Levenshtein operations to learn a continuous latent space of vector representations to capture the underlying semantic information with regard to the document editing process. Though our model outperforms strong baselines when fine-tuned on edit-centric tasks, it is unclear if these results are due to domain similarities between fine-tuning and pre-training data, suggesting that the benefits of our proposed approach over regular masked language-modelling pre-training are limited.