Iterative Edit-Based Unsupervised Sentence Simplification

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7918–7928 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 7918 Iterative Edit-Based Unsupervised Sentence Simplification Dhruv Kumar,1 Lili Mou,2 Lukasz Golab,1 Olga Vechtomova1 1University of Waterloo 2Department of Computing Science, University of Alberta Alberta Machine Intelligence Institute (Amii) {d35kumar,lgolab,ovechtomova}@uwaterloo.ca doublepower.mou@gmail.com Abstract We present a novel iterative, edit-based ap- proach to unsupervised sentence simplifica- tion. Our model is guided by a scoring func- tion involving fluency, simplicity, and mean- ing preservation. Then, we iteratively per- form word and phrase-level edits on the com- plex sentence. Compared with previous ap- proaches, our model does not require a paral- lel training set, but is more controllable and interpretable. Experiments on Newsela and WikiLarge datasets show that our approach is nearly as effective as state-of-the-art super- vised approaches.1 1 Introduction Sentence simplification is the task of rewriting text to make it easier to read, while preserving its main meaning and important information. Sentence sim- plification is relevant in various real-world and downstream applications. For instance, it can bene- fit people with autism (Evans et al., 2014), dyslexia (Rello et al., 2013), and low-literacy skills (Watan- abe et al., 2009). It can also serve as a preprocess- ing step to improve parsers (Chandrasekar et al., 1996) and summarization systems (Klebanov et al., 2004). Recent efforts in sentence simplification have been influenced by the success of machine transla- tion. In fact, the simplification task is often treated as monolingual translation, where a complex sen- tence is translated to a simple one. Such simplifi- cation systems are typically trained in a supervised way by either phrase-based machine translation (PBMT, Wubben et al., 2012; Narayan and Gardent, 2014; Xu et al., 2016) or neural machine translation (NMT, Zhang and Lapata, 2017; Guo et al., 2018; Kriz et al., 2019). Recently, sequence-to-sequence 1Code is released at https://github.com/ ddhruvkr/Edit-Unsup-TS (Seq2Seq)-based NMT systems are shown to be more successful and serve as the state of the art. However, supervised Seq2Seq models have two shortcomings. First, they give little insight into the simplification operations, and provide little con- trol or adaptability to different aspects of simplifi- cation (e.g., lexical vs. syntactical simplification). Second, they require a large number of complex- simple aligned sentence pairs, which in turn require considerable human effort to obtain. In previous work, researchers have addressed some of the above issues. For example, Alva- Manchego et al. (2017) and Dong et al. (2019) explicitly model simplification operators such as word insertion and deletion. Although these ap- proaches are more controllable and interpretable than standard Seq2Seq models, they still require large volumes of aligned data to learn these oper- ations. To deal with the second issue, Surya et al. (2019) recently proposed an unsupervised neural text simplification approach based on the paradigm of style transfer. However, their model is hard to in- terpret and control, like other neural network-based models. Narayan and Gardent (2016) attempted to address both issues using a pipeline of lexical substitution, …

Content truncated to 500 words