Abstract
This paper reports on an experiment in which 11 subjects were asked to highlight important passages in an 1111-word text. These results were cross-referenced with a range of word attributes in order to test hypotheses about the principles underlying highlighting decisions. With this data a combination of selection criteria was proposed that was able to predict probability of highlighting with a correlation of approximately 0.56, compared with an average correlation of 0.47 amongst the test subjects, and a figure of 0.30 for Word97's highlighting feature. The paper argues that the common factor behind the most successful hypotheses was that they are all signals denoting "new" as opposed to "given" information at the discourse level. Although based on a very limited sample this observation seems clear enough to make detecting such signals a promising candidate for further research.