Parameter learning for a readability checking tool

This paper describes the application of machine learning methods to determine parameters for DeLite, a readability checking tool. DeLite pinpoints text segments that are difficult to understand and computes for a given text a global readability score, which is a weighted sum of normalized indicator values. Indicator values are numeric properties derived from linguistic units in the text, such as the distance between a verb and its complements or the number of possible antecedents for a pronoun. Indicators are normalized by means of a derivation of the Fermi function with two parameters. DeLite requires individual parameters for this normalization function and a weight for each indicator to compute the global readability score.

Several experiments to determine these parameters were conducted, using different machine learning approaches. The training data consists of more than 300 user ratings of texts from the municipality domain. The weights for the indicators are learned using two approaches: i) robust regression with linear optimization and ii) an approximative iterative linear regression algorithm. For evaluation, the computed readability scores are compared to user ratings. The evaluation showed that iterative linear regression yields a smaller square error than robust regression although this method is only approximative. Both methods yield results outperforming a first manual setting, and for both methods, basically the same set of non-zero weights remain.