The text, we have seen, patterns and delimits, but it ultimately functions like a chemical element: it itself is merged in the synthesis with the other elements to produce a particular event-a poem.

We, as a universal society, have long agreed that a poem is more than a mere collection of words: it is, we might feel, a picture painted with words. Psycholinguistic and neurocognitive research has only just started to explore the underlying cognitive and affective processes which help us make sense of a poetic text (Jacobs and Kinder, 2015; Jacobs and Willems, 2018) and ultimately shape and form mental images often reported while reading poetry (Belfi et al., 2018; Papp-Zipernovszky et al., 2021).

As already assumed by proponents of the readers-response theories like Iser (1978) and Rosenblatt (1978), reading and especially reading literature can only be understood as an interaction between features of the text and characteristics of the reader, and, as recently highlighted, also by the task at hand and the reading situation (Jacobs, 2015a; Bohn-Gettler, 2019). As described in the Transactional Theory by Rosenblatt (1978), reading of a poetic text is a process of meaning construction in which the linguistic symbols, the words and the written or printed marks have to be distinguished from what a listener or reader evokes from them (Rosenblatt, 1978). Rosenblatt used the term poem only for the later one. Iser described the meaning construction process as “one of the activities through which we form the ‘gestalt' of a literary text” based on “the ‘picturing' that is done by our imagination” (Iser, 1972, p. 288). But how does this interaction between reader and text come about? Let us start our analysis by looking at the interaction between text and reader from afar: At low resolution the emphasis is on the influence of literature genre on reading behaviour. According to proponents of the Reader Response Theory, reading behaviour depends on the reader's assumptions about the text (Hanauer, 1998; Topping, 2015) and a reading mode relevant to the understanding of the genre is adopted according to the genre choice made. As choices and adaptation of reading mode take time, several empirical studies demonstrated that texts labelled as factual are read faster compared to fictional texts (Zwaan, 1994; Wolfe, 2005; Altmann et al., 2014, see Hartung et al., 2017, for contrasting results). Also memory for individual aspects of a text differs as a function of the type of text (Zwaan, 1994). In line with the general assumption that classification of a text as a poem gives rise to a reading mode characterised by heightened awareness of language patterns and hidden meanings (for a comprehensive overview see Hanauer, 1998 and Carminati et al., 2006). Osowiecka and Kolanczyk (2018) found that poetry reading stimulates divergent thinking and vice versa.

According to Graesser (2015), the above theories are sufficient to explain basic levels of comprehension, but they are insufficient to describe deeper comprehension processes assumed to play a role when reading literature. The latter requires to also consider parameters such as the influence of emotions, situations etc. both through the literal content in itself, but also through an interaction between reader and text. In other words, text-bound variables like rhythm (Kentner, 2012; Peelle and Davis, 2012), prosody (Paulmann et al., 2013), valence and arousal within the words employed (Kuchinke et al., 2005; Warriner et al., 2013; Jacobs et al., 2016; Sylvester et al., 2016) and reader-centered variables such as affective and aesthetic responses (De Jaegher and Di Paolo, 2007; Gallagher, 2012; Winkielman et al., 2014), often studied from an embodied perspective, are closely intertwined. The Neurocognitive Poetics Model (NCPM; Jacobs, 2011, 2015a,c; Jacobs and Kinder, 2017) aims at integrating data on neurocognitive and emotional (Jakobson, 1960; Panksepp, 2008) processes during literature comprehension. It is conceptualised as a dual-process model which is informed by Gestalt theory and as such distinguishes between back- and foreground elements. The understanding of a text is conceived as successful closing of a good Gestalt. Within the framework of the model a text consists of familiarising background elements which cater to the readers need for coherence and meaning construction. The model further assumes that backgrounding features are stronger associated with immersive reactions (Van Krieken et al., 2015) and a mood-empathic reading mode (Lüdtke et al., 2014; Jacobs et al., 2016). Emotional and aesthetic feelings (Fitch, 2009; Westphal-Fitch and Fitch, 2018) are created against this background through foregrounding text elements. Foregrounding elements have a defamiliarising effect (Kuiken et al., 2012; Jacobs, 2015a,b,c) thereby eliciting what Mukarovský (1964) described as a de-automatisation in the reading process, leading to reduced reading speed as a result of more and intensive active drawing of inferences. Although manipulation of words always influences both backgrounding and foregrounding elements, empirical studies on reading poetic texts primarily focussed on the manipulation of foregrounding elements. Hakemulder (2004) for example changed isolated lines of poems to manipulate the amount of foreground features and observed that foregrounding caused higher aesthetic appreciation. Obermeier et al. (2013) manipulated rhyme and metre and also demonstrated that both foregrounding features enhanced aesthetic appreciation. As a text is characterised by both background and foreground features, experimental studies on the influence of background features like word frequency are equally important for the study of poetry comprehension. Yet studies focussing on the isolated manipulation of background features are few and far between (Hanauer, 1998). This raises the question of whether background features have any influence on comprehension of poetic texts at all? The short answer is: they do. For the long answer we need to look at studies addressing these elements.

Facilitation or ease of processing, i.e., the ease of mental operations concerned with stimulus meaning and its relation to semantic knowledge structures, is known to affect perceived aesthetic value. This is recognised as the mere-exposure effect, which links repetition of a stimulus to enhanced liking for this initially neutral stimulus (Zajonc, 2000). In support of this idea, Menninghaus et al. (2017) showed that parallelistic diction, a supra-lexical background feature, enhances emotional and aesthetic appreciation as well as semantic processing. In a similar way Reber and Schwarz linked variables that facilitate the processing of a stimulus (e.g., priming, mastery and the amount of information) to positive affective reactions (Reber et al., 2002, 2004). Markedly, aesthetic appreciation, a bona fide reaction to the presence of foregrounding elements, is reciprocally influenced by and at the same time acts as a positive influence on processing fluency (Reber et al., 2004). Moreover, the absence of prototypical narrative features (which are often discussed as an example of backgrounding features Hakemulder, 2020) strongly increases processing efforts as measured by reading times (Castiglione, 2017). Manipulation of word frequency, perhaps the most prototypical backgrounding feature, is widely employed to highlight the influence of facilitating factors on processing speed and accuracy (Murray and Forster, 2004) and appears to mediate pleasentness- and familiarity ratings on single word level (Sluckin et al., 1980).

Although ease of processing was shown to influence the above, the concept is not sufficient to explain the closing of a good Gestalt sensu (Iser, 1972). While an explanation based on the ease of processing focusses on single word level cognition (e.g., higher word frequency leads to faster word recognition), a good Gestalt results from the integration of words. In the same way that cognitive theories assume that associations are the most important principle for organisation of our semantic knowledge (for an overview cf. Hofmann and Jacobs, 2014), they should also guide the closing of a good Gestalt while reading poetic texts. In this regard, we can posit an influence of what James (cited after Mangan, 2001, 2008) described as the fringe of consciousness on processing fluency-mediated construction of meaning: It represents context information in consciousness, conveys a sense of rightness and brings out the notion that more information is available.

What do we already know about the role of associations on text comprehension processes related to the closing of a good Gestalt? According to the Gestalt principle of closure, objects which are grouped together are perceived as a whole which might otherwise (i.e., if the objects were regarded as individual entities) not exist and which is more than or different from the combination of its parts. The emergence of a resulting Gestalt from an albeit sketchy figure is hinged on the presence of a unifying pattern (Warriner, 1923). Wolfgang Iser transmitted that principle onto the process of reading literary texts and therefore poetry: “By grouping together the written parts of the text, we enable them to interact, we observe the direction in which they are leading us, and we project onto them the consistency which we, as readers, require. This ‘gestalt' must inevitably be colored by our own characteristic selection process” (Iser, 1972, p. 289). A poem, composed of words that interrelate, facilitates the application of the grouping principle of closure so that a good Gestalt can emerge and potentially foster the creation of a mental image (Iser, 1972).

Yet, words have an effect on comprehension in their own right. Readers pre-activate concepts beyond the unfolding sentence to simulate the described event for predictive (online) language comprehension (for an overview see Huettig, 2015). Associations or mental connections between concepts or events are by some described as the product of contiguity, i.e., the probability of co-occurrence between words. Others conceptualise associations as the result of an event simulation heuristic which pre-activates linguistic representations (Huettig, 2015). Semantic cohesion serves as a measure for the strength of internal association between concepts and was shown to influence sentence processing on single sentence level (Hintz et al., 2020). Current machine-learning methodologies (Grainger and Jacobs, 1996; Dijkstra and van Heuven, 2002; Hofmann and Jacobs, 2014; Jacobs, 2018) as well as some association databases rely on contiguity-based approaches. Accordingly, in a recent study Jacobs and Kinder showed that latent semantic analysis (LSA, a method based on distributional semantics to predict upcoming words) is superior to surface model analysis for the prediction of ratings on difficulty and aesthetic liking of metaphors (Jacobs and Kinder, 2017). This reveals the power of associative databases which exceed purely systematic relations (such as WordNet) and instead rely on a rich network of semantic relations informed by cultural, emotional and personal experience (Netzer et al., 2009).

Poems lend themselves to the study of neurocognitive processes during reading, seen as they are well constrained literary material (Müller et al., 2017). Seen in the light of associative networks, poetry appears to be characterised by lexical associations (Netzer et al., 2009). Correspondingly, a computational analysis of poems written by professional and amateur poets revealed that the former use more concrete words to create a canvas onto which readers can project their associations. Professional poets avoid direct references to abstract and intangible concepts and instead let images instead of words convey emotions, concepts, and experiences (Kao and Jurafsky, 2012).

With the above in mind, we want to investigate the effect of associations, as an example of a backgrounding feature, on the closing of a good Gestalt and maintain that the effect exceeds a mere effect of facilitated processing. To test this proposition, we created simple word lists, which are essentially devoid of foreground features, and presented them as versions of poetic texts (Figure 1). The participants are asked to read lists of nine words with varying degree of semantic cohesion introduced as micropoems (Jacobs, 2015b). They are then asked to produce a possible one-word-title for each poem and rate the micropoems on dimensions Comprehensibility, Liking, Imageability as indication of reader's response at both the upper and lower route according to the NCPM (Jacobs, 2015a). This setup allows for an analysis of the influence of background features -more precisely the role of semantic cohesion and word frequency in the present case- on the closing of a good Gestalt independent of the influence of foregrounding figures.

As each and every word is linked to a plethora of other concepts, reading such a simple word list not only evokes simulation of the mentioned concepts but also co-activates the concepts associated with those words (Mangan, 2001; Hofmann and Jacobs, 2014). Each group of concepts associated with a single word in itself is likely to evoke its own association, seen as they are all semantically related and trace back to a common anchor word (Deese, 1959; Roediger and McDermott, 1995). The idea behind our set-up is not that each word per se leads to the generation of many associations. Instead, the words taken together are all highly semantically related (high semantic cohesion as a feature facilitating formation of a Gestalt) and thus, as a whole, facilitate the generation of associations in the reader's mind. We therefore assume a reciprocal enhancement of activation which facilitates the formation of a consistent mental image or good Gestalt as for example opposed to a serial position effect. If only the first or the last word of a poem was used for the generation of a title, we should not observe any effect of semantic cohesion on semantic relatedness. Reading poems with low internal semantic cohesion, on the other hand should merely activate concepts vaguely associated with the individual words. What is more, associations may neutralise each other through lateral inhibition at the conceptual level (for a thorough analysis see Hofmann and Jacobs, 2014) and therefore inevitably fail to close a good Gestalt.

Consequently, we should be able to test whether semantic cohesion of the words composing a poem has a significant impact on the closing of a good Gestalt, and therefore of the formation of mental imagery, a process associated with the recognition of the bigger picture conveyed by a poetic text. Hence, we expect participants to understand the micropoems with higher semantic cohesion more easily and additionally report the experience of more mental imagery. Moreover, semantic cohesion should also influence affective and aesthetic responses as described in the NCPM (Jacobs et al., 2015). Following the scheme introduced by Dixon and Bortolussi (2015) a set of explicitly surveyed rating data (mood induced, liking of the micropoem, perceived difficulty of the micropoem, perceived imagery induced) and ancillary reaction parameters (mean semantic relatedness of the response, mean reading time) as implicit control variables will be used.

In accordance with the assumptions of the NCPM, a modification of micropoems in terms of internal semantic cohesion should be reflected in the mood induced in the reader (NCPM: backgrounding, fast route) and self-reported aesthetic appreciation of the micropoems (NCPM: foregrounding, slow route) (Jacobs, 2015a). With regards to difficulty, the effect should be mirrored in the self-reported difficulty to understand the poems (explicit measure) and reading times (implicit measure). The capacity of a poem to induce a mental image in the reader should be seen in a self-reported rating of perceived mental images. Moreover, participants will be asked to put their first spontaneous association to the poems forward in form of a freely chosen title. Semantic relatedness between spontaneous associations to the poems (i.e., titles) and the words composing a poem as well as the number of different titles found, as an indicator for the closing of a good Gestalt, are used to control this effect with an implicit measure.

One might argue that the effects discussed above can be put down to facilitated processing due to semantic cohesion. Following this line of thought, if the root cause of any differences observed was indeed attributable to an ease of processing, this effect should reproduce if another factor influencing the ease of processing (e.g., supra-lexical word frequency) was modified. If supra-lexical word frequency has no or a different effect, then it is clear that the influence of semantic cohesion is based on more than a simple facilitation of word identification.

We expect self-reported mood and imagery to be stronger for micropoems with high semantic cohesion than low semantic cohesion. Micropoems with high semantic cohesion should be easier to understand (less self reported difficulty in understanding) and should be better liked. Regarding the implicit measures we expect higher semantic relatedness between the title and the reported associations (i.e., titles given) and faster reading times for micropoems with high internal semantic cohesion. Moreover the intra-subject variance of associations found should increase for micropoems with reduced internal semantic cohesion.

We attribute the effects to the evolution of more elaborate associations in the participants' minds as a function of internal semantic cohesion and therefore do not expect any effect for word frequency.

We have always started our analysis with a complex model with maximal degrees of freedom for the random effects. All initial models contained both fixed effects, i.e., semantic cohesion and supra-lexical word frequency and their interaction (cf. Table 1). We then reduced the random effects model in stepwise manner until the model was no longer overfitting (Baayen et al., 2008; Barr, 2013; Barr et al., 2013). For all dependent variables interaction or main effects with supra-lexical word frequency never became significant (all F < 0.4, all p > 0.5) and were therefore removed.

We have set out to understand the mechanisms that make us, as readers of a poem, see a bigger picture upon the inward eye. We hypothesised that even background features should have a decisive influence on reading and understanding poetic texts. More precisely, we assumed that semantic relatedness of the words making up a poem, as one example for a supra-lexical background feature, is a main factor governing the formation of the closed Gestalt. To test this hypothesis we had participants read lists of words with different amount of semantic relatedness and assayed their emotional response as well as their associations as a proxy for the mental images produced. Moreover we also manipulated supra-lexical word frequency defined as the mean word frequency of all words composing the micro poem, to test whether possible effects rely on a simple facilitation of word identification only. Taken together, we found no effect for supra-lexical word frequency but an effect for semantic cohesion on all objective variables as well as on subjective ratings liking and difficulty.

The observed differences of semantic relatedness between titles and micro-poems as a function of semantic cohesion lets us understand two things: first we can safely conclude that our manipulation was successful. Secondly, on condition that semantic relatedness of the titles is indicative of the capacity of a poem to induce a mental image, we have shown that the closing of a good Gestalt is at least partially hinged on semantic cohesion of a micropoem. Up to this point a mere effect of facilitated processing would be an equally plausible explanation for the present results. If that was the case, however, supra-lexical word frequency, which famously facilitates processing of words and texts (Lüdtke et al.)² should lead to the same results. We, however, did not find any main or interaction effect of word frequency on any variable examined. Therefore the present results need to be explained through effects beyond facilitation of processing. Our findings are in line with the results of Jolsvai et al. (2013), showing that meaningfulness of sentence fragments is a superior predictor for language processing speed than word frequency. Taken together this indicates that regarding the formation of a bigger picture facilitation of lexical access, i.e., basic comprehension processes at simple word level, is a second player to integration processes. In the following we will discuss this in more detail in reflection of current discourse on mental imagery, aesthetic appreciation, and comprehension difficulty.

Let us begin with an analysis of the results concerning mental imagery. Further to semantic relatedness, we analysed the number of different titles between items with high and low semantic cohesion, as a direct measure of mental imagery and the closing of a good Gestalt. Items with high semantic cohesion were expected to point to a target word or induce a target concept and therefore to yield a lower number of different associations. Items with low semantic cohesion on the other hand should lead to the generation of more different and less related associations. The present data confirmed this hypothesis: items with high semantic cohesion yielded fewer and less semantically related associations. While the analysis of the number of titles would indicate that a good Gestalt was indeed derived from items with high semantic cohesion, directly assayed generation of mental imagery (see above) did not. It can only serve as indicator or a trend pointing in the same direction. Although the latter finding at first glance seems to contradict our hypothesis, it is, upon closer inspection, a faithful reflection of the ongoing discussion whether mental imagery is indeed produced during reading, or is more of a literal afterthought (Jacobs and Willems, 2018). Moreover, the poems were essentially devoid of a deeper meaning or plot (Teng et al., 2016). This already by and in itself may disturb the process of closing a good Gestalt and thereby have interfered with the creation of mental images. In fulfillment of Rosenblatt's Transactional Theory (Rosenblatt, 1978), we also may not neglect the reader in this equation. Reader-based variables, particularly their response to the set task and the consequently adopted reading goal, can increase a reader's tendency to employ an inferential reading mode (Bohn-Gettler, 2019). In light of this, the task set here may have distracted the reader from adopting an inferential reading mode which otherwise would have allowed for mental images to be drawn. The readers may instead have focussed on getting it right and may therefore have lost access to any actively communicable notion of a mental image. Looking at participants' ratings for aesthetic appreciation of the items, no effects were found on dimension experience of Mood, but clear effects were obtained for Liking. This is in opposition to the predictions derived from the NCPM, according to which manipulation of background features should be reflected in upper route processes more than in lower route processes. Back- and foregrounding elements, however, cannot be thought as entities isolated from each other, seen as they mutually affect each other (Reber et al., 2004). In this sense, Obermeier et al. (2013) showed that manipulation of foreground features can influence upper route processes. Also Lüdtke et al. (2014) demonstrated cross-communication between foregrounding and backgrounding elements. These findings point in the direction already outlined by Jacobs (2011) suggesting that it is the quotient of backgrounding and foregrounding elements more than their individual contribution that facilitates enhanced processing along either route. More research is needed to shed light on this effect. In this sense, our approach lends itself to the analysis of the influence of semantic activation on both trajectories through the use of a set of micropoems with continuously increasing semantic cohesion. Moreover Belfi et al. (2018) showed that vividness of mental imagery has a crucial effect on aesthetic appreciation of poems.

On the axis Comprehensability, the subjective measure (survey data) became significant, while the objective measure (reading time) just failed to be significant on a Bonferroni corrected alpha level, but still indicated a correlation on a descriptive level. Reading time could, theoretically, be construed as a measure for ease of processing. By showing that word frequency, a factor directly related to processing speed, does not have any influence on reading times we demonstrate that semantic cohesion does not only influence basic language processing like word identification but presumably also higher order processes like integration and the closing of a good Gestalt. We argue that micropoems encouraging the formation of mental images should be easier understood and hence read faster. Magyari et al. (2020), per contra, reported prolonged fixation duration on words associated with an enactive style text, for which participants reported more mental images. On a fist glance this, again, seems to contradict our results which indicated faster reading times for items that were reported to induce more mental images. Yet, processing a prose text differs fundamentally from processing of our micropoems. Reading poetic texts is characterised by enhanced re-reading (Xue et al., 2020). Therefore the reading times measured in this study not only reflect ongoing reading processes (i.e., first pass reading and fixation duration as captured in the eyetracking data reported by Magyari et al., 2020) but also higher order comprehension processes necessary for fulfilling the task at hand. Contrary to studies on literary reading, in our case subjects had to find one word titles. We assumed that semantic cohesion fosters the closing of a good Gestalt which leads to enhanced mental imagery. This should simplify the task and thereby enhance participants' reaction times. More studies need to sheet light onto the role of mental imagery in reading prose and poetry. The method employed here is a promising tool to account for factors influencing mental imagery when reading poetic texts.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Ethik Kommission der Freien Universität Berlin. The patients/participants provided their written informed consent to participate in this study.

KH and JL contributed equally to the conception and design and report of the study. All authors have read and agreed to the final version of the manuscript.

We acknowledge support by the Open Access Publication Fund of the Freie Universität Berlin.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.