Learning when to simplify sentences for natural text simplification

Abstract. This paper introduces a corpus-based approach for selecting sen-tences that require simplification in the context of Brazilian Portuguese text sim-plification system. Based on a parallel corpus of original and simplified text versions, we apply a binary classifier to decide in which circumstances a sen-tence should or not be split – which is the most important syntactic simplification operation – so that the resulting simplified text is natural and not over simpli-fied. Our classifier reaches 73.5 % precision and 73.4 % recall when selecting the sentences to be split or kept together. 1