Automatic Text Difficulty Classifier - Assisting the Selection Of Adequate Reading Materials For European Portuguese Teaching

This paper describes a system to assist the selection of adequate reading materials to support European Portuguese teaching, especially as second language, while highlighting the key challenges on the selection of linguistic features for text difficulty (readability) classification. The system uses existing Natural Language Processing (NLP) tools to extract linguistic features from texts, which are then used by an automatic readability classifier. Currently, 52 features are extracted: parts-of-speech (POS), syllables, words, chunks and phrases, averages and frequencies, and some extra features. A classifier was created using these features and a corpus, previously annotated by readability level, using a five-levels language classification official standard for Portuguese as Second Language. In a five-levels (from A1 to C1) scenario, the best-performing learning algorithm (LogitBoost) achieved an accuracy of 75.11% with a root mean square error (RMSE) of 0.269. In a three-levels (A, B and C) scenario, the best-performing learning algorithm (C4.5 grafted) achieved 81.44% accuracy with a RMSE of 0.346.