CATS: A Tool for Customized Alignment of Text Simplification Corpora

CATS: A Tool for Customized Alignment of Text Simplification Corpora Sanja ˇStajner1, Marc Franco-Salvador2, Paolo Rosso3, Simone Paolo Ponzetto1 1 DWS Research Group, University of Mannheim, Germany 2 Symanto Research, Nuremberg, Germany 3 PRHLT Research Center, Universitat Polit`ecnica de Val`encia, Spain {sanja, simone} @informatik.uni-mannheim.de, marc.franco@symanto.net, prosso@prhlt.upv.es Abstract In text simplification (TS), parallel corpora consisting of original sentences and their manually simplified counterparts are very scarce and small in size, which impedes building supervised automated TS systems with sufficient coverage. Furthermore, the existing corpora usually do not distinguish sentence pairs which present full matches (both sentences contain the same information), and those that present only partial matches (the two sentences share the meaning only partially), thus not allowing for building customized automated TS systems which would separately model different simplification transformations. In this paper, we present our freely available, language-independent tool for sentence alignment from parallel/comparable TS resources (document-aligned resources), which additionally offers the possibility for filtering sentences depending on the level of their semantic overlap. We perform in-depth human evaluation of the tool’s performance on English and Spanish corpora, and explore its capacities for classification of sentence pairs according to the simplification operation they model. Keywords: text simplification, tools and resources, sentence similarity 1. Introduction Automated text simplification (ATS) has the goal of au- tomatically transforming sentence structure and lexical choices in a way that it provides better understanding and wider accessibility to large audiences. The main obsta- cle for successful supervised ATS is the scarcity and lim- ited size of parallel TS corpora which would contain orig- inal sentences and their manual simplifications. The par- allel TS corpus for Brazilian Portuguese, compiled for the purposes of the PorSimples project (Alu´ısio et al., 2008) contains around 4,500 aligned sentences, and the paral- lel TS corpus for Spanish, compiled for the purposes of the Simplext project (Saggion et al., 2015) contains only around 1,000 aligned sentences. The largest existing TS comparable corpora is the English Wikipedia – Simple En- glish Wikipedia (EW–SEW), consisting of 170,000 sen- tence pairs (Kauchak, 2013), or 150,000 full matches and 130,000 partial matches in the newer version (Hwang et al., 2015). In both cases, the sentences were automatically aligned from comparable English Wikipedia and Simple English Wikipedia articles. However, the use of EW–SEW dataset for modeling TS has been disputed (Amancio and Specia, 2014; ˇStajner et al., 2015; Xu et al., 2015) for sev- eral reasons: (1) the simplified articles are not necessarily direct simplifications of the original articles; (2) the quality of simplifications is not checked; (3) the dataset does not cover sentence splitting which is one of the most common operations in text simplification. The Newsela corpora1 of document-aligned news texts, manually simplified at four different simplification levels have been freely available for a few years for research pur- poses. These corpora have several advantages over the EW– SEW dataset (Xu et al., 2015; ˇStajner et al., 2017): (1) simplified texts present direct simplifications of the orig- inal articles; (2) simplification was performed by …

Content truncated to 500 words