Learning How to Simplify From Explicit Labeling of Complex-Simplified Text Pairs

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 295–305, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Learning How to Simplify From Explicit Labeling of Complex-Simplified Text Pairs Fernando Alva-Manchego1 and Joachim Bingel2 and Gustavo H. Paetzold1 Carolina Scarton1 and Lucia Specia1 1Department of Computer Science, University of Sheffield, UK 2Department of Computer Science, University of Copenhagen, Denmark {f.alva,g.h.paetzold,c.scarton,l.specia}@sheffield.ac.uk bingel@di.ku.dk Abstract Current research in text simplification has been hampered by two central problems: (i) the small amount of high-quality par- allel simplification data available, and (ii) the lack of explicit annotations of simplifi- cation operations, such as deletions or sub- stitutions, on existing data. While the re- cently introduced Newsela corpus has al- leviated the first problem, simplifications still need to be learned directly from par- allel text using black-box, end-to-end ap- proaches rather than from explicit anno- tations. These complex-simple parallel sentence pairs often differ to such a high degree that generalization becomes diffi- cult. End-to-end models also make it hard to interpret what is actually learned from data. We propose a method that decom- poses the task of TS into its sub-problems. We devise a way to automatically identify operations in a parallel corpus and intro- duce a sequence-labeling approach based on these annotations. Finally, we provide insights on the types of transformations that different approaches can model. 1 Introduction Text Simplification (TS) is the task of reducing the complexity of a text without changing its meaning. Simplification can be applied at various linguistic levels, from lexical substitution to more global op- erations such as sentence splitting, paraphrasing or the deletion or reordering of entire clauses. Existing corpora for TS generally come in one of two variants. The first focuses on very spe- cific sub-problems, such as sentence compression FA and JB contributed equally to this paper. (Bingel and Søgaard, 2016) or the identification of difficult words (Paetzold and Specia, 2016a), and typically encodes relevant simplification op- erations as discrete labels on tokens. The other variant includes more general, higher-level types of simplifications that often entail the rephras- ing or re-structuring of sentences, with content added or removed. These “natural” simplifications are often created for end-users rather than for re- search purposes. Examples of the latter simplifi- cation resources include the Newsela (Xu et al., 2015) and Simple English Wikipedia corpora (Zhu et al., 2010; Coster and Kauchak, 2011b). These resources generally encode interdependencies be- tween different types of simplification better than single-purpose resources and may thus seem fa- vorable for learning simplifications. However, the freedom given to editors and lack of explicit la- bels on the modifications performed makes gen- eralization much more difficult, especially when existing resources are relatively small in compari- son to corpora for other text-to-text problems like machine translation (MT). Nevertheless, these cor- pora have been extensively used to learn phrase- based statistical and neural models for end-to-end TS systems that bear resemblance to MT mod- els (Specia, 2010; Zhu et al., 2010; Coster and Kauchak, …

Content truncated to 500 words