Exploring multiple profiles of highly rated learner compositions

Abstract

Recent research has come a long way in describing the linguistic features of large samples of written texts, although a satisfactory description of L2 writing remains problematic. Even when variables such as proficiency, language background, topic, and audience have been controlled, straightforward predictive relationships between linguistic variables and quality ratings have remained elusive, and perhaps they always will. We propose a different approach. Rather than assuming a linear relationship between linguistic features and quality ratings, we explore multiple profiles of highly rated timed compositions and describe how they compare in terms of their lexical, grammatical, and discourse features. To this end, we performed a cluster analysis on two sets of timed compositions to examine their patterns of use of several linguistic features. The purpose of the analysis was to investigate whether multiple profiles (or clusters) would emerge among the highly rated compositions in each data set. This did indeed occur. Within each data set, the profiles of highly rated texts differed significantly. Some profiles exhibited above-average levels for several linguistic features, whereas others showed below-average levels. We interpret the results as confirming that highly rated texts are not at all isometric, even though there do appear to be some identifiable constraints on the ways in which highly rated timed compositions may vary.