A Neural Network model for the Evaluation of Text Complexity in Italian Language: a Representation Point of View

The goal of a text simplification system (TS) is to create a new text suited to the characteristics of a reader, with the final goal of making it more understandable.The building of an Automatic Text Simplification System (ATS) cannot be separated from a correct evaluation of the text complexity. In fact the ATS must be capable of understanding if a text should be simplified for the target reader or not. In a previous work we have presented a model capable of classifying Italian sentences based on their complexity level. Our model is a Long Short Term Memory (LSTM) Neural Network capable of learning the features of easy-to-read and complex-to-read sentences autonomously from a annotated corpus created specifically for text simplification. In this paper we further investigate on the role of the text representation, i.e. how different ways of representing the input text can affect the accuracy of the proposed system. In detail, we will use our Neural Network model for evaluating the sentence complexity using different kind of representations such as GloVe, Word2vec, FastTex and a new one based on a representation learning scheme.