This article describes and reports on the performance of six related artificial neural networks that have been developed for the purpose of readability analysis. Two networks employ counts of linguistic variables that simulate a traditional regression-based approach to readability. The remaining networks determine readability from “visual snapshots” of text. Input text is transformed into a visual pattern representing activation levels for input level nodes and then “blurred” slightly in an effort to promote generalization. Each network included one hidden layer of nodes in addition to input and an output layers. Of the four snapshot readability systems, two are trained to produce grade equivalent output and two depict readability as a distribution of activation values across several grade levels. Results of preliminary trials indicate that the correlation between visual input systems and judgements by experts is low although, in at least one case, comparable to previous correlations reported between readability formulas and teacher judgement. A system using linguistic variables and numerical output correlated perfectly with a regression-based formula within the error tolerance established prior to training. The networks which produce output in the form of a readability distribution suggest a new way of reporting readability that may do greater justice to the concept of readability than traditional grade equivalent scores while, at the same time, addressing concerns that have been voiced about the illusory precision of readability formulas.
Fault and Error Tolerance in Neural Networks: A Review
