Abstract
This thesis describes new approaches to text readability that can help in making written communications more readily understood. The research aims to undergo a deeper exploration of readability and the properties of text that make it easier to understand. The work will attempt to gain an improved understanding of the relationship between text and the reader and propose new methods for assessing readability that will move away from existing surface measures of readability. These existing readability formulas, in spite of their popularity, raise various concerns in failing to detect a number of common writing problems because their focus is largely on word and sentence length. Many users also attempt to improve their writing on the basis of the results produced by these formulas, but acting on the basis of a single numerical output for an entire text presents difficulties in locating troublesome writing. There is criticism, also, that the existing readability formulas were validated using out-dated school grades. In our research, we are taking an extended view of readability that incorporates features of text cohesion, propositional density, word familiarity and the abilities and knowledge of the reader. We have devised a framework that incorporates these features, and derived a number of new configurable measures of readability that can provide reader-specific scoring. Some of the new measures can be applied at sentence level to direct authors to specific writing problems and provide automated feedback on their resolution. To assess the success of our new approaches, we critique and assess existing methods for evaluating readability measures. We use reader preferences over short text passages, when controlled for other features of readability to evaluating our readability methods. We found close agreement between the new measures and reader preferences, demonstrating that the measures can be configured to provide agreement for different types of reader. In addition, we devised a new method for evaluating readability measures using reading times from eye-tracking data. One of our measures, for word familiarity, shows a strong correlation (0.941) with average reading time in comparison to the next best performing existing readability measure (0.906). This particular measure incorporates sentence length in characters, due to the finding of a significant correlation (0.963) between sentence-level reading time and sentence length in characters. The latter indicates that character counts might reflect sentence-level text difficulty rather more accurately than the existing readability measures but would be insufficient alone to offer targeted feedback, and our combination is advantageous even with the cost of a slightly lowered correlation. Using the word familiarity measure, we have also derived a way to estimate how much reading time can be saved by improving readability. We demonstrate our approach in several software prototypes which can provide additional automated feedback to users on how to improve their readability. One of these prototypes, for Open Office, has generated over 12,000 downloads to date.