An Imitation Learning Curriculum for Text Editing with Non-Autoregressive Models

We propose a framework for training non-autoregressive sequence-to-sequence\nmodels for editing tasks, where the original input sequence is iteratively\nedited to produce the output. We show that the imitation learning algorithms\ndesigned to train such models for machine translation introduces mismatches\nbetween training and inference that lead to undertraining and poor\ngeneralization in editing scenarios. We address this issue with two\ncomplementary strategies: 1) a roll-in policy that exposes the model to\nintermediate training sequences that it is more likely to encounter during\ninference, 2) a curriculum that presents easy-to-learn edit operations first,\ngradually increasing the difficulty of training samples as the model becomes\ncompetent. We show the efficacy of these strategies on two challenging English\nediting tasks: controllable text simplification and abstractive summarization.\nOur approach significantly improves output quality on both tasks and controls\noutput complexity better on the simplification task.\n