Assessing the readability of sentences: which corpora and features?

The paper investigates the problem of sentence readability assessment, which is modelled as a classification task, with a specific view to text simplification. In particular, it addresses two open issues connected with it, i.e. the corpora to be used for training, and the identification of the most effective features to determine sentence readability. An existing readability assessment tool developed for Italian was specialized at the level of training corpus and learning algorithm. A maximum entropy-based feature selection and ranking algorithm (grafting) was used to identify to the most relevant features: it turned out that assessing the readability of sentences is a complex task, requiring a high number of features, mainly syntactic ones.