Does EFL Readers' Lexical and Grammatical Knowledge Predict Their Reading Ability? Insights From a Perceptron Artificial Neural Network Study

Abstract

This study aims to examine the relationship between reading comprehension and lexical and grammatical knowledge among English as a foreign language students by using an Artificial Neural Network (ANN). There were 825 test takers administered both a second-language reading test and a set of psychometrically validated grammar and vocabulary tests. Next, their reading, grammar, and vocabulary abilities were estimated by the Rasch model. A multilayer ANN was used to classify low- and high-ability readers based on their grammar and vocabulary measures. ANN accurately classified approximately 78% of readers with reference to their vocabulary and grammar knowledge. This finding is consistent with the cognitive theories of reading that treat the lexical and grammatical knowledge of learners as a major factor in distinguishing poor from competent readers. The study also confirmed previous research in finding that vocabulary knowledge was associated with reading comprehension more strongly than grammatical knowledge.

Acknowledgments

We thank the three anonymous reviewers of Educational Assessment for their careful reading and insightful comments on the earlier drafts of this article. The findings may not be taken as evidence for the validity argument of UTEPT, as the study has not been designed to validate the uses and interpretations of the UTEPT scores.

Notes

¹ It is also important to interpret Laufer's (1992) findings with some caution: Laufer used an arbitrary score on a nonstandardized test of reading comprehension as the criterion for arguing that learners can achieve text comprehension.

² Enthusiastic readers are referred to the Mathtutor website for further information and tutorials on mathematical functions: http://www.mathtutor.ac.uk/functions/inroductiontofunctions. See also Appendix A.

³ Some computer packages, such as AMOS, offer different parameter estimation methods, such as maximum likelihood or the least squares method, although these techniques still assume linearity. For further information concerning (non)linearity in SEM, see Kelava and Brandt (2009).

⁴ The cohort of applicants who took the test in 2010 is treated as the observed population—a representative sample from the “superpopulation” or all possible applicants taking the test over the years (the concept of superpopulation was introduced by Deming & Stephan, 1941). The selected test takers from the observed population (n = 856) are treated as a sample representative of the observed population and, accordingly, the superpopulation itself.

⁵ The ROC area is equivalent to the Mann–Whitney U test, a nonparametric test that is used in non-normal data in lieu of t tests (Mason & Graham, 2002).