An “AI readability” formula for French as a foreign language

[shlomihod]

http://www.aclweb.org/anthology/D12-1043

[maxseidler]

Summary:

• Trying to predict text difficulty according to language proficiency levels (A1 to C2)
• Many hand-crafted features (406!) given as features for an SVM, feature groups:
  1. Lexical features
  2. Syntactic features
  3. Semantic features
  4. Features specific to French
• Good paper for the other group because they compare the predictive power of different features
  • Lexical and syntactic features had biggest, semantic feature had no significant predictive power. The French-spedific features only had a minor impact.
• Best model (SVM trained with the most predictive features) had only 49% accuracy