Modeling preference in fictional texts based on structural features

Computational textual aesthetics is an emerging field that aims to investigate observable differences between aesthetic categories of text. In this study, we explored structural differences between preferred and non-preferred fictional texts. To put our results into perspective, we also analyzed non-fictional texts and compared them with fictional texts. Canonization was used to operationalize preference for texts from the 19th and early 20th centuries, while for contemporary texts, sales figures were regarded as a proxy for readers' preference. Looking for the distinctive structural characteristics of text categories, we represented texts as sequences (series) of text properties and analyzed them using three main approaches: variability, fractality, and predictability analysis. Our findings revealed that canonical fiction exhibits more variability compared to non-canonical fiction. Fractality analysis showed that long-range correlation patterns are more similar in canonical and non-canonical texts, suggesting that fractality is a universal feature of text, slightly more pronounced in non-fictional texts. Predictability analysis focuses on (ir)regularities and uncertainty within texts. We analyzed different aesthetic categories by applying Approximate Entropy as a measure of surprise in local structures of texts, and Shannon Entropy as a global measure of unpredictability. Our findings demonstrated that preferred texts are less predictable than non-preferred texts, and predictability analysis can reveal structural differences between various categories of text. We further investigated whether structural properties of text, which are potential textual correlates of preference, vary across different time periods. Our findings confirm that design features of text change over time, and they can be utilized to distinguish text categories with varying degrees of preference.