This paper investigates the problem of automated text aesthetics prediction. The availability of user generated content and ratings, e.g. Flickr, has induced research in aesthetics prediction for non-text domains, particularly for photographic images. This problem, however, has yet not been explored for the text domain. Due to the very subjective nature of text aesthetics, it is difficult to compile human annotated data by methods such as crowd sourcing with a fair degree of inter-annotator agreement. The availability of the Kindle "popular highlights" data has motivated us to compile a dataset comprised of human annotated aesthetically pleasing and interesting text passages. We then undertake a supervised classification approach to predict text aesthetics by constructing real-valued feature vectors from each text passage. In particular, the features that we use for this classification task are word length, repetitions, polarity, part-of-speech, semantic distances; and topic generality and diversity. A traditional binary classification approach is not effective in this case because non-highlighted passages surrounding the highlighted ones do not necessarily represent the other extreme of unpleasant quality text. Due to the absence of real negative class samples, we employ the MC algorithm, in which training can be initiated with instances only from the positive class. On each successive iteration the algorithm selects new strong negative samples from the unlabeled class and retrains itself. The results show that the mapping convergence (MC) algorithm with a Gaussian and a linear kernel used for the mapping and convergence phases, respectively, yields the best results, achieving satisfactory accuracy, precision and recall values of about 74%, 42% and 54% respectively.
展开▼
