An “AI readability” Formula for French as a Foreign Language

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 466–477, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics An “AI readability” formula for French as a foreign language Thomas Franc¸ois IRCS, University of Pennsylvania 3401 Walnut Street Suite 400A Room 423 Philadelphia, PA 19104, USA frthomas@sas.upenn.edu C´edrick Fairon CENTAL, UCLouvain Place Blaise Pascal, 1 1348 Louvain-la-Neuve, Belgium Cedrick.Fairon@uclouvain.be Abstract This paper present a new readability formula for French as a foreign language (FFL), which relies on 46 textual features representative of the lexical, syntactic, and semantic levels as well as some of the specificities of the FFL context. We report comparisons between sev- eral techniques for feature selection and var- ious learning algorithms. Our best model, based on support vector machines (SVM), sig- nificantly outperforms previous FFL formulas. We also found that semantic features behave poorly in our case, in contrast with some pre- vious readability studies on English as a first language. 1 Introduction Whether in a first language (L1) or a second and for- eign language (L2), learning to read has been and re- mains one of the major concerns of education. When a teacher wants to improve his/her students’ reading skills, he/she uses reading exercises, whether there are guided or independent. For this practice to be efficient, it is necessary that the texts suit the level of students (O’Connor et al., 2002). This condition is sometimes difficult to meet for teachers wishing to get off the beaten tracks by not using texts from levelled textbooks or readers. In this context, readability formulas have long been used to help teachers faster select texts for their students. These formulas are reproducible meth- ods that aim at matching readers and texts relative to their reading difficulty level. The Flesch (1948) and Dale and Chall (1948) formulas are probably the best-known examples of those. They are typical of classic formulas, the first major methodological paradigm developed in the field during the 40’s and 50’s. They were kept as parsimonious as possible, using linear regression to combined two, or some- times, three surface features, such as word mean length, sentence mean length, or proportion of out- of-simple-vocabulary words. Later, some scholars (Kintsch and Vipond, 1979; Redish and Selzer, 1985) argued that the classic for- mulas suffer from several shortcomings. These for- mulas only take into account superficial features, ig- noring other important aspects contributing to text difficulty, such as coherence, content density, infer- ence load, etc. They also omit the interactive as- pect of the reading process. In the 80’s, a second paradigm, inspired by structuro-cognitivist theories, intended to overcome these issues. It focused on higher textual dimensions, such as inference load (Kintsch and Vipond, 1979; Kemper, 1983), den- sity of concepts (Kintsch and Vipond, 1979), or macrostructure (Meyer, 1982). However, these at- tempts did not achieve better results than the clas- sic approach, even though they used more principled and more complex features. Recently, a third paradigm, referred …

Content truncated to 500 words