Natural reading is rather like a juggling feat, as our eyes and minds are kept on several things at the same time. Instead, reading texts developed by researchers (so-called “textoids”; Graesser, Millis, & Zwaan, 1997) may be fairly simple, since this facilitates an experimental investigation. It thus provides the chance for clear statements regarding the effect of predefined variables. Likewise, most empirical studies focused only a few selected features while ignoring the great diversity of possibly important others (e.g., Rayner et al., 2001; Reichle, Rayner, & Pollatsek, 2003; Rayner & Pollatsek, 2006; Engbert et al., 2005; Rayner, 2009). However, it is not possible to directly transfer the results generated from textoids to natural reading due to the identification of more than 100 features on different hierarchical levels, which may influence processing a natural text (Graf, Nagler, & Jacobs, 2005; Jacobs, 2015a, b; Jacobs et al., 2017). The present dissertation differed from past research in that it used a literary text, i.e., Shakespeare’s sonnets, instead of texts constructed by the experimenter. The goal of the present dissertation was to investigate how psycholinguistic features may influence the reading behavior during poem perception. To this end, two problems need to be handled: Firstly, complex natural texts need to be broken up into measurable and testable features by “turning words into numbers” (Franzosi, 2010) for the sake of statistical analysis. Secondly, statistical ways were sought to deal with the non-linear webs of correlations among different features, which has long been a concern of Jacob’s working group (e.g., Willems, 2015; Willems & Jacobs, 2016; Jacobs & Willems, 2018). A quantitative narrative analysis (QNA) based predictive modeling approach was suggested to solve the above problems (e.g., Jacobs et al., 2017; Jacobs, 2017, 2018a, b). Since it is impossible to identify all relevant features of a natural text [e.g., over 50 features mentioned for single word recognition (Graf et al., 2005) or over 100 features computed for the corpus of Shakespeare sonnets (Jacobs et al., 2017)] and including more inter/supra-lexical features also requires extending sample sizes (i.e., more/longer texts and more participants), my dissertation focuses on lexical features. Seven of these are surface features (word length, word frequency, orthographic neighborhood density, higher frequency neighbors, orthographic dissimilarity index, consonant vowel quotient, and the sonority score) and two are affective-semantic features (valence and arousal). By applying the QNA-based predictive modeling approach, I conducted three eye tracking studies: study 1 (Chapter 5) asked English native speakers to read three of Shakespeare’s sonnets (sonnet 27, 60, and 66), aiming to investigate the role of seven surface psycholinguistic features in sonnets reading. Study 2 (Chapter 6) used a rereading paradigm and let another group of English natives read two of the three sonnets (sonnet 27 and 66), to find out whether the roles of the surface psycholinguistic features may be changed in rereading. In study 3 (Chapter 7), I reanalyzed the data of study 2, in which beyond the surface features I started to pay attention to the affective-semantic features, hoping to examine whether the roles of surface and affective-semantic features may be different throughout reading sessions. The three studies show highly reliable data for high feature importance of surface variables, and in rereading an increasing impact of affective-semantic features in reading Shakespeare’s sonnets. From a methodological viewpoint, all three studies show a much better sufficiency of neural net approach than the classical general linear model approach in psycholinguistic eye tracking research. For the rereading studies, in general, compared to the first reading, rereading improved the fluency of reading on poem level (shorter total reading times, shorter regression times, and lower fixation probability) and the depth of comprehension (e.g., Hakemulder, 2004; Kuijpers & Hakemulder, 2018). Contrary to the other rereading studies using literary texts (e.g., Dixon et al., 1993; Millis, 1995; Kuijpers & Hakemulder, 2018), no increase in appreciation was apparent. In summary, this dissertation can show that the application of predictive modeling to investigate poetry might be far more suitable to capture the highly interactive, non-linear composition of linguistic features in natural texts that guide reading behavior and reception. Besides, surface features seem to influence reading during all reading sessions, while affective-semantic features seem to increase their importance in line with processing depth as indicated by higher influence during rereading. The results seem to be stable and valid as I could replicate these novel findings using machine learning algorithms within my dissertation project. My dissertation project is a first step towards a more differentiated picture of the guiding factors of poetry reception and a poetry specific reading model.