Assessing the relative reading level of sentence pairs for text simplification

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 288–297, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Assessing the relative reading level of sentence pairs for text simplification Sowmya Vajjala and Detmar Meurers LEAD Graduate School, Seminar f¨ur Sprachwissenschaft Universit¨at T¨ubingen {sowmya,dm}@sfs.uni-tuebingen.de Abstract While the automatic analysis of the read- ability of texts has a long history, the use of readability assessment for text simplifi- cation has received only little attention so far. In this paper, we explore readability models for identifying differences in the reading levels of simplified and unsimpli- fied versions of sentences. Our experiments show that a relative rank- ing is preferable to an absolute binary one and that the accuracy of identifying rel- ative simplification depends on the ini- tial reading level of the unsimplified ver- sion. The approach is particularly success- ful in classifying the relative reading level of harder sentences. In terms of practical relevance, the ap- proach promises to be useful for identi- fying particularly relevant targets for sim- plification and to evaluate simplifications given specific readability constraints. 1 Introduction Text simplification essentially is the process of rewriting a given text to make it easier to process for a given audience. The target audience can ei- ther be human users trying to understand a text or machine applications, such as a parser analyzing text. Text simplification has been used in a vari- ety of application scenarios, from providing sim- plified newspaper texts for aphasic readers (Can- ning and Tait, 1999) to supporting the extraction of protein-protein interactions in the biomedical do- main (Jonnalagadda and Gonzalez, 2009). A related field of research is automatic readabil- ity assessment, which can be useful for evaluating text simplification. It can also be relevant for in- termediate simplification steps, such as the identi- fication of target sentences for simplification. Yet, so far there has only been little research connect- ing the two subfields, possibly because readability research typically analyzes documents, whereas simplification approaches generally targeted lex- ical and syntactic aspects at the sentence level. In this paper, we attempt to bridge this gap between readability and simplification by studying read- ability at a sentence level and exploring how well can a readability model identify the differences be- tween unsimplified and simplified sentences. Our main research questions in this paper are: 1. Can the readability features that worked at the document level successfully be used at the sen- tence level? 2. How accurately can we identify the differences in the sentential reading level before and after simplification? To pursue these ques- tions, we started with constructing a document- level readability model. We then applied it to nor- mal and simplified versions of sentences drawn from Wikipedia and Simple Wikipedia. As context of our work, we first discuss rel- evant related research. Section 2 then describes the corpora and the features we used to construct our readability model. Section 3 discusses the performance of our readability model in compari- son …

Content truncated to 500 words