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

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7550 - 7563 May 22-27, 2022 c⃝2022 Association for Computational Linguistics An Imitation Learning Curriculum for Text Editing with Non-Autoregressive Models Sweta Agrawal Department of Computer Science University of Maryland sweagraw@cs.umd.edu Marine Carpuat Department of Computer Science University of Maryland marine@cs.umd.edu Abstract We propose a framework for training non- autoregressive sequence-to-sequence models for editing tasks, where the original input se- quence is iteratively edited to produce the out- put. We show that the imitation learning al- gorithms designed to train such models for machine translation introduces mismatches be- tween training and inference that lead to un- dertraining and poor generalization in editing scenarios. We address this issue with two com- plementary strategies: 1) a roll-in policy that exposes the model to intermediate training se- quences that it is more likely to encounter dur- ing inference, 2) a curriculum that presents easy-to-learn edit operations first, gradually in- creasing the difficulty of training samples as the model becomes competent. We show the efficacy of these strategies on two challenging English editing tasks: controllable text simpli- fication and abstractive summarization. Our approach significantly improves output quality on both tasks and controls output complexity better on the simplification task. 1 Introduction Neural sequence-to-sequence (seq2seq) models pri- marily developed and tested for machine translation (MT) Bahdanau et al. (2015); Vaswani et al. (2017); Gu et al. (2018) are increasingly used for other se- quence transduction tasks. This paper focuses on editing tasks, such as post-editing of MT output (Simard et al., 2007), style transfer (Jin et al., 2020), or text simplification (Chandrasekar and Srinivas, 1997; Xu et al., 2015), where systems directly edit the input sequence, instead of generating the output from scratch as in MT. As illustrated in Table 1, in these tasks, there might be substantial overlap in content between inputs and outputs, and also diverse rewrites, ranging from local substitutions to more complex restructuring. While dedicated architectures have been de- signed for these editing tasks, based on e.g., a Original: The Mauritshuis museum is staging an exhibition focusing on the 17th century self- portraits, highlighting the similarities and the differences between modern-day snapshots and historic works of art. Simplified: The Mauritshuis museum is now set to open an exhibit on the 17th century self- portraits. It shows the similarities and differ- ences between modern photos and artworks. Table 1: Text simplification is an editing task, where the output sequence overlaps with the input, while in- corporating multiple rewrite types to restructure and simplify content. multistep, tag-then-edit approach (Alva-Manchego et al., 2017; Malmi et al., 2019; Dong et al., 2019; Mallinson et al., 2020), they can also be addressed with non-autoregressive (NAR) seq2seq models which generate their output by iteratively editing intermediate sequences (Lee et al., 2018; Gu et al., 2019; Awasthi et al., 2019; Stern et al., 2019; Chan et al., 2020). NAR models hold the promise of pro- viding a more generic solution, where …

Content truncated to 500 words