Diagnosing Causes of Reading Difficulty using Bayesian Networks

International Joint Conference on Natural Language Processing, pages 1383–1391, Nagoya, Japan, 14-18 October 2013. Diagnosing Causes of Reading Difficulty using Bayesian Networks Pascual Mart´ınez-G´omez The University of Tokyo National Institute of Informatics pascual@nii.ac.jp Akiko Aizawa The University of Tokyo National Institute of Informatics aizawa@nii.ac.jp Abstract There is a need of matching text diffi- culty to the expected reading skill of the audience. Readability measures were de- veloped with this objective in mind, first by psycholinguists, and more recently, by practitioners of natural language process- ing. A common strategy was to extract linguistic features that are good predic- tors of readability, and then train discrim- inative classification or regression models that correlate well with human judgment. But correlation does not imply causality, which is a necessary property to explain why documents are not readable. Our ob- jective is to provide mechanisms for text producers to adjust the readability of their content. We propose the use of generative models to diagnose causes of reading dif- ficulty, and bring closer the realization of automatic readability optimization. 1 Introduction Educational institutions, government agencies and some private companies have a special interest in authoring documents for a certain audience, but it is expensive to involve expert linguists to as- sess the readability of every document they pro- duce. The first psycholinguistic studies developed readability formulas for grading purposes, based on surface linguistic features. Those formulas, de- spite of their simplicity, performed well and were widely used by editors to grade reading material for young readers. However, content producers might be tempted to adapt their manuscripts by tweaking the text features present in readability formulas, without gaining (or even degrading) real readability (Davison and Kantor, 1982). Recently, the application of statistical models to linguistic problems proved successful, and am- bitious tasks in automatic document transforma- tion such as text summarization or machine trans- lation became a hot topic in computational linguis- tics. Readability optimization is one of such docu- ment transformation problems. Recent studies on readability embrace machine learning techniques to recognize readability with an even higher accu- racy. The common approach consists in extracting as many features as possible, and then training a classifier or a regression model using human an- notated texts to predict a readability score given the observation of the linguistic features. Those discriminative models correlate well with human judgment, but fail at explaining why a doc- ument is not readable. We call readability diagno- sis the automatic discovery of the causes that lead to (un)readability, and we believe it is an essential step for readability optimization. We propose the use of Bayesian causal net- works to perform readability diagnosis. That is, given a document, the objective of our Bayesian network is to recognize the specific parts of the document that are difficult to read. Bayesian networks are a type of generative model, where the joint probability distribution is constructed by making certain independence assumptions. Their main advantage is that they allow to query the model regarding any linguistic variable, general- izing …

Content truncated to 500 words