The Depth Beyond the Lines: Piloting of the Psycholinguistic Test Battery for Polish Poetry Study

Abstract

We present a psycholinguistic test battery designed to examine the cognitive and affective processes involved in reading Polish poetry. This toolkit combines reader profiling (vocabulary, memory and reading proficiency) with tasks that assess the influence of lexical, textual, affective and poetic features on recognition, context restoration and association generation. Pilot data confirmed the reliability of the measures and their sensitivity to recognised psycholinguistic effects. Vocabulary size and delayed memory rehearsal strongly predicted performance in content restoration, while recognition and association latencies were closely related, indicating shared retrieval mechanisms. Structural and affective properties also influenced responses: line-final words improved recognition but impeded association, with these effects being moderated by word length and frequency. Words that were negatively valenced, abstract and hardly imaginable were restored more accurately than positive or concrete ones. These findings demonstrate the potential of the battery for profiling readers and provide new insights into how Polish poetic language engages memory and associative processes.

1. Introduction

Reading is a multidimensional phenomenon involving cognitive and perceptual, socio-emotional, and background knowledge domains (e.g., Cain and Oakhill 1999; Kim 2020; Smith et al. 2021). The reading process is further influenced by modalities, i.e., paper vs. screen (Hakemulder and Mangen 2024; Mangen and Kuiken 2014), and text genres (Hanauer 1998; McCarthy et al. 2024; Zwaan 1994) yet the genre effect is challenged by recent studies (Hartung et al. 2017; Triantafyllopoulos et al. 2021).

Empirical studies of literature (ELS) is a rapidly growing interdisciplinary field (Alber and Strasen 2020; Kuiken and Jacobs 2021; Lauer 2015) that explores how literary reading influences our cognition and emotions. Within ELS, empirical research on poetry has emerged as a distinct subfield. Since Roman Jakobson introduced the poetic function of language (Jakobson 1975) and advocated for the application of rigorous scientific methods to study poetry (Jakobson 1976), several influential frameworks have been developed, including Cognitive Poetics (Tsur 2008), Neurocognitive Poetics (Jacobs 2014, 2015), and Psychopoetics (van Peer and Chesnokova 2022). They combine literary analysis with modern empirical techniques such as eye tracking, fMRI, or EEG. However, the methodological framework of cognitive poetics is still being developed (e.g., Brône and Vandaele 2009).

Poetry is an example of “unconventional” texts (De Beaugrande and Dressler 1982), being charged with poetic devices, specific prosody, syntax violations, and visuo-spatial organization (e.g., Kulawik 1994, pp. 52–130; Lukin 1999). These traits challenge the reader’s ability to create a coherent text representation in mind. The quality of these representations often depends on linguistic markers, narrative types, readers’ background knowledge, etc. (Moss and Schunn 2015; Saux et al. 2021; van Dijk and Kitsch 1983). Once poetry steps outside conventional narrative frameworks, it initiates comprehension tactics that differ from those used with other text types.

In recent decades, researchers have uncovered remarkable patterns in how poetry shapes readers’ experience, stirring feelings, awakening emotions, and guiding the reading process. Poetry is read more slowly, provoking more regressive eye movements with a higher number of fixations compared to prose (Blohm et al. 2022; Fechino et al. 2020; Peskin 2007), and metrical anomalies exhibit disruptive eye movements and increased rereadings (Beck and Konieczny 2021). Geyer et al. (2020) reported that juxtaposition of two images and the usage of cut markers complicate meaning resolution and reading fluency. They likely act as foregrounding elements and thus de-automatizing the reading process.

In Gao and Guo’s (2018) study, Chinese poetry was evaluated as more beautiful than prose and induced intense activation of “aesthetic centers” in the human brain. Obermeier et al. (2015) observed the facilitating effect of rhyme and meter in different levels of processing and increased liking of rhymed and metered pieces. Kraxenberger and Menninghaus (2017) have shown that affinity with poetry affects aesthetic evaluation. Cross and Fujioka (2019) observed the right hemisphere dominance and left hemisphere reduction while rhyme processing by lyricists compared to non-lyricists. Experts are likely more attuned to phonological incongruity, employing more holistic problem-solving and decision-making strategies (Cross 2023). Lea et al. (2021) suggested experts’ advanced skills in phonological anticipation, and Kao and Jurafsky (2012) reported exceptional skills of poetry experts in lexical and poetic forms reception. Fokin et al. (2022) observed that poets tend to keep a steady pace, progressing through the line, suggesting that they perceive all words as equally important.

Menninghaus and Wallot (2021) observed rhymed and metered poetry is more beautiful and moving and provokes the deceleration of the reading pace, likely caused by “savoring effect”. Xue et al. (2019, 2023) cemented the list of predictors important for poetry reading, including lexical, textual, affective-semantic features such as orthographic neighborhood and dissimilarity, sonority score, valence, etc. During the rereading session, affective features became even more important. This is in line with findings on the key role of affective characteristics of words in poetry (Aryani et al. 2016; Lüdtke et al. 2014). Hugentobler and Lüdtke (2021) reported that higher semantic cohesion induces fewer and less varied associations, contributes to mental image creation, and influences basic (word identification) and higher-order (integration) processes.

A recent large-scale study by Porter and Machery (2024) introduces a new angle to this debate by comparing AI-generated and human-authored poetry. Readers were unable to distinguish between two types of poetry and evaluated aesthetically higher AI-generated pieces while human-written pieces more often “made no sense”. Researchers concluded that more straightforward artificial poetry unambiguously communicates image, mood, and emotions, and does not require in-depth comprehension and understanding. Similar patterns were previously observed in Köbis and Mossink’s (2021) study, yet readers found AI-generated poems more semantically poor.

Together, these findings provide extensive evidence on how poetry shapes reading and cognitive processes, revealing the unique influence of poetry on comprehension and aesthetic appreciation. However, the rich tradition of Polish poetry remains underrepresented in empirical research. The underrepresentation leads to the perpetuation of a methodological lacuna and conceptual blindness to particular language families and poetic traditions, creating a systematic bias. Hence, it is crucial to conduct empirical research that is as linguistically and formally diverse as possible, expanding the research scope and thus increasing the empirical material on the basis of which universal patterns might be observed. In this study, we enlarge the extant body of work by including Polish poetic material.

2. The Present Study

2.1. State-of-the-Art

Only a few papers investigate Polish poetry using empirical and quantitative methods. Comparing English and Polish texts, Grabska-Gradzińska et al. (2012) found that literary texts, i.e., poetry and prose, tend to follow a scale-free structure, meaning a high dispersion of the lexical network with only a few significant nodes (concepts) having a high number of connections; scientific texts, in turn, consist of a pool of highly frequent and well-connected words (for review see Nodus Labs 2012; Sole and Valverde 2004). Osowiecka and Kolanczyk (2018) observed increased indices of creative thinking after reading Szymborska’s “Utopia” compared to the cooking book description, aligning genre-specific hypotheses (Hanauer 1998). Yet, these isolated studies do not apply mixed-method approaches or propose methodological tools suitable for empirical research on Polish poetry. This highlights the need for methodologically robust instruments to investigate poetry comprehension in controlled settings.

For the experiment, we selected poems by the most renowned modern Polish poets: Wisława Szymborska, Adam Zagajewski, and Zbigniew Herbert. Their practices arouse permanent interest among literary scholars (e.g., Czapliński and Śliwiński 2000; Hutnikiewicz and Lam 2000), Szymborska was honoured with the Nobel Prize in Literature in 1996 while Zagajewski was its finalist and Herbert—a candidate in 1968. Also, all three won numerous international awards. The empirical study of these authors offers fresh insights on their poetry and its influence on readers. In the present study, we introduce a psycholinguistic test battery for the empirical study of poetry and poetry readers.

2.2. Aims and Hypotheses

The main purposes of the study are to develop an experimental toolkit for the empirical study of poetry and readers’ cognitive and behavioral responses and to pilot the battery under controlled circumstances. The following research question guided our exploration:

RQ1. 

What means are best-fitted for comprehensive empirical study of modern poetry?

RQ2. 

How does reading modern poetry impact cognitive processes involved in verbatim and content remembrance, associative processes, and topic identification?

RQ3. 

What are the roles of textual, lexical, poetic, and affective features in framing these effects?

Based on the proposed questions, we assume:

H1. 

The in-depth analysis of modern poetry comprehension can be conducted, employing the two-step experimental toolkit consisting of preliminary reading-related skills assessment and control study of reading-related cognitive processes;

H2. 

Selected means are capable of (a) measuring verbatim memory, content remembrance, and associative processes; and (b) the reader’s ability to identify the poetic topic.

H3. 

Affective features are the key drivers in poetry reception, shaping verbatim memory, content remembrance, and associative responses. Textual word properties, particularly position in the line, are salient predictors of remembrance but are refined by poetic devices, e.g., rhyme or enjambment. Lexical features contribute to reception, but their influence is contingent on affective word properties, which can either enhance or diminish their impact.

3. Materials and Methods

Stimuli

The selected poems exemplify two of the most prevalent forms of contemporary Polish poetry (Kulawik 1994; Pszczołowska 2001). Herbert’s poem adheres to regular syllabic-accentual verse, while Szymborska’s and Zagajewski’s poems are written in free verse (the so-called sentence poem). All three adhere to the syntactic norms of the Polish language, use conventional lexicon, and utilize analogous poetic devices, such as enumeration, syntactic parallelism, and verse anaphora, while being devoid of rhyme regularities.

The poems are comparable in readability index (SMOG) (McLaughlin 1969; Scott 2025), number of lines, and average word length, diverging in visual layout, prosodic contour, and poetic tropes employed (e.g., metaphor, enjambment) (

Table 1. Stimulus poems descriptives.

Author	Poem	Readability (SMOG)	Nlines	Nwords	Mword_length	Meter	Rhyme
Wisława Szymborska	“Muzeum” (Museum)	5/7 (10.46)	20	97	5.7	no	irregular
Zbigniew Herbert	“Mój ojciec” (My Father)	5/7 (9.78)	24	121	5.6	yes	no
Adam Zagajewski	“Z życia przedmiotów” (From lives of things)	5/7 (10.2)	20	98	5.6	no	internal

). Thus, poems represent the broad but nevertheless comparable spectrum of poetic practice.

4. Research Variables

The research variables reflect distinct yet interacting layers of poetic architectonics (Jacobs 2015)—lexical, textual, affective, and poetic—which shape how readers cognitively engage with poetic texts.

Lexical properties of words (e.g., frequency and length) typically follow Zipf’s law (Lestrade 2017; Linders and Louwerse 2023; Zipf 1949), where high-frequency words tend to be shorter. These properties directly affect eye movement behavior (Rayner 1998), appear to be universal across alphabetic languages (Kuperman et al. 2024), and also influence memory in line with Kolmogorov complexity (Ehret 2018), which links length to the restoration of information content. Word length represents the number of letters per word. Word frequencies were derived from Polish Web Corpus 2019 (Jakubíček et al. 2013; Lexical Computing CZ s.r.o. n.d.) and log-transformed to normalize the distribution and mitigate skewness typical for raw word frequencies. Textual features included position within the line (initial, medial, or final) and the word’s serial position in the poem. Word placement is a crucial feature for poetic study since it represents syntactic peculiarities (including such devices as enjambement) and spatial organization (Blohm et al. 2022; Fechino et al. 2020). It was observed that overall line-final words are better remembered and induce the closure effect (Kuperman et al. 2010; Menninghaus and Wallot 2021).

Poetic variables included rhyme, enjambment, metaphor, and enumeration, and were manually annotated (0 = absent; 1 = present). These poetic tropes were chosen due to their well-documented effect on reading behavior and experienced emotions (Jacobs and Kinder 2018; Rasse et al. 2020; Stamenković et al. 2023). Metaphors were identified based on conceptual mappings and figurative language markers, e.g., wój wyścig z suknią nadal trwa (my race with the dress is still on). Enjambments are well-known as disturbing elements that impede reading and comprehension fluency (Delente 2024; van ‘t Jagt et al. 2013) and create a tension between syntactic and versification lines, adding fragmentarity but also flow to the poetic piece (Mastalski 2023; Tsur 2015):

Omszały woźny drzemie słodko zwiesiwszy wąsy nad gablotką

W. Szymborska

The moss-grown guard in golden slumber props his moustache on the Exhibit Number.

(transl. S. Barańczak and C. Cavanagh)

Enumerations function as prosodic scaffolding, contributing rhythmicity and cohesion to the poetic structure. It adds rhythmicity and references to the clause, creating a sense of stasis and motion, manipulating it with different degrees of connections between elements involved (e.g., Brickey 2022; Vedder 2022). In the study, enumerations are understood as phonological parallelisms and list-like elements, sequentially located within one or several lines and linked to one semantic construct, e.g.,

dotykamy i oka i powietrza, ciemności i światła, Indii i Europy

(A. Zagajewski)

We touch the eye and air, the darkness and the light, India and Europe.

(literal transl. D.F.)

Rhyme, as a basic parallelistic, rhythm-related element, creates the beauty, poeticity, and melodiousness of poetry (Menninghaus and Wallot 2021), purposefully delivers a wide spectrum of emotions (Kraxenberger and Menninghaus 2017; Johnson-Laird and Oatley 2022), and evokes intense feelings and aesthetic appreciation (Obermeier et al. 2015). In the paper, rhyme is understood as rhymed consonances in any line position:

w niebieskich chmurach aromatu

smakował uśmiech w wargach wąskich

Z. Herbert

In the blue clouds of aroma

He savored a smile in his narrow lips

(literal transl. D.F.)

The selected poems lack regular metrical and rhyming patterns. So, these dispersed rhythmic structures require readers to rely more heavily on localized poetic cues.

Finally, affective features refer to the emotional meaning of a word and readers’/listeners’ feelings evoked by a word (Kuperman et al. 2014). They were taken from Imbir’s (2016) norms gathered using the Self-Assessment Manikin (SAM) 9-point scale (Lang 1980) and included: Valence (unpleasant vs. pleasant); Dominance (submissive vs. control); Origin (emotional vs. rational); Arousal (calmness vs. excitement); Significance (unimportant vs. important); Concreteness1 (concrete vs. abstract); Imageability (low vs. high degree of mental imagery)2. A recent study of Aka et al. (2021) revealed the positive influence of some affective features (valence, arousal, concreteness) on word remembrance. Other studies revealed the impact of affective word properties on poetry reading, liking, aesthetic, and emotional appreciation (e.g., Aryani et al. 2016; Hugentobler and Lüdtke 2021; Johnson-Laird and Oatley 2022; Ullrich et al. 2017). Thus, lexical affectiveness may play a key role in poetry reception, creating a tone and atmosphere of a poem but also modulating readers’ perception of it.

Table 2. Descriptives of features used in the study.

Type (N)	Variable	M_all (SD)	Zagajewski Mean (SD)	Szymborska Mean (SD)	Herbert Mean (SD)
Lexical (316)	log_frequency	2.206 (1.377)	2.084 (1.335)	2.021 (1.354)	2.452 (1.405)
	length	5.269 (2.810)	5.551 (2.781)	5.165 (2.637)	5.124 (2.971)
Textual * (316)	position in text	53.633 (31.357)	49.5 (28.434)	49.000 (28.145)	61.000 (35.074)
Poetic ** (316)	enumeration	0.158 (0.366)	0.235 (0.426)	0.175 (0.382)	0.083 (0.276)
	metaphor	0.326 (0.479)	0.214 (0.412)	0.423 (0.497)	0.339 (0.475)
	rhyme	0.500 (0.501)	0.490 (0.502)	0.423 (0.497)	0.570 (0.497)
	enjambement	0.111 (0.314)	0.82 (0.275)	0.041 (0.200)	0.200 (0.394)
Affective *** (121)	valence	5.380 (1.275)	5.090 (1.537)	5.380 (1.008)	5.560 (1.076)
	arousal	3.740 (0.876)	3.850 (1.096)	3.820 (0.810)	3.580 (0.535)
	origin	5.520 (0.898)	5.380 (1.014)	5.900 (0.715)	5.550 (0.857)
	dominance	5.360 (1.058)	5.160 (1.169)	5.480 (1.061)	5.420 (0.774)
	concreteness	4.180 (1.608)	4.396 (1.562)	3.936 (1.684)	3.967 (1.582)
	imageability	6.740 (1.340)	6.650 (1.195)	6.540 (1.569)	6.880 (1.261)
	significance	4.240 (0.854)	4.450 (0.749)	4.060 (0.954)	4.040 (0.870)

Note *. Position in line is limited to three—initial, middle, and final—and converted into numbers, thus the mean is equal to 2 for each poem. Note **: Poetic devices were coded as binary variables. The mean values in the table show the proportion of words in each text that contain the respective poetic device; Note ***. Affective features for poems are represented in medians due to the large number of missing data.

summarizes the variables included in the analysis.

The interactions of all the above-mentioned features are summarized in Figure 1 (see for details Fokin (forthcoming). In short, the matrix reveals strong relations between lexical and affective word properties, the key role of word frequencies, whereas textual features barely correlate with linguistic and affective ones.

5. Methodology

5.1. General Overview

The test battery (six experiments lasting around 60 min) consists of two main blocks:

• Preliminary experiments (three tests lasting 15 min) include the Polish adaptation of the SLS-Berlin test (Lüdtke et al. 2019) to assess one’s reading proficiency, hereafter referred to as the Polish Reading Proficiency Test (PRPT); the Polish Vocabulary Size Test (Fokin et al. 2025), developed to evaluate participants’ receptive vocabulary size; and the Polish MoCA, a verbal short-term memory subtest (Gierus et al. 2015), which explores the accuracy of word remembrance. Together, they are aimed at creating a reader’s linguistic profile.
• Control experiments (3 tests lasting 45 min) assessing high-level cognitive processes such as memory-based processes, task-switching, inferencing, and contextual associations: Context-Primed Associations Task, designed to test key predictors impacted by verbatim memory, switching between intermediate- and long-term memory buffers, and associative processes. The Cloze-Test (CT) is employed to study the quality of text representation and poetic content remembrance, and the Topic Inferential Task aims to reveal inferential skills.

5.2. Procedure

Participants first signed an informed consent form and an agreement for the processing of personal data. They were then instructed about the procedure and proceeded with the study. First, they completed the PRPT on the experimental computer, followed by the MoCA. Next, participants read a practice verse—the first quatrain of Ewa Lipska’s “Kraj podobny do innych”—displayed on the screen, and then completed the CPAT, CT, and TIT tasks to familiarize themselves with the design. After reading the practice verse, they were asked to repeat aloud the words from a delayed memory task.

After the practice session, the control experiment began. Participants read the three poems presented in random order, one by one. After reading each poem, they completed the CPAT (via the LimeSurvey platform, version 6.4; Lime Survey GmbH 2023) and CT tasks. This sequence was repeated three times—once for each poem.

Participants’ eye movements were recorded using an SMI 500 eye tracker throughout the reading phase. The CPAT was completed on participants’ personal devices, while the Cloze Test (CT) was administered on the experimental computer.

Finally, after all three poems and the respective tasks were completed, participants performed the TIT task on the experimental computer.

5.3. Polish Reading Proficiency Test (PRPT)

PRPT was adapted to Polish from the SLS-Berlin Test (Lüdtke et al. 2019), aiming to bridge the critical gap in assessing reading proficiency. Here, participants read 77 sentences and responded aloud whether the sentence aligned with general world knowledge (correct) or violated it (incorrect). Then, they evaluate sentence complexity on a 5-point scale. Readers’ complexity evaluations were averaged for each sentence and compared with the German original. Sentences were translated from German, considering Polish cultural and language nuances. The Polish adaptation is comparable with the original in the number of syllables but exceeds it in the number of letters and long words, and yields to it the number of words in a sentence (see Appendix A,

Table A1. Descriptive statistics of 77 Polish sentences and their German original.

					95% CI
	Polish Mean (SD)	German Mean (SD)	t	p	Lower Bound	Upper Bound
number of letters	64.03 (19.19)	74.36 (17.80)	−4.7244	1.036 × 10−5	59.669	68.382
number of syllables	26.68 (8)	25.38 (5.72)	1.421	0.159	24.86	28.491
number of words	9.81 (2.75)	10.96 (2.91)	−3.687	<0.000	9.181	10.429
number of long words (>6 letters)	4.96 (1.96)	4.27 (1.46)	3.089	0.003	4.515	5.407

).

5.4. Polish Vocabulary Size Test (PVST)

PVST (Fokin et al. 2025) was newly developed using a procedure introduced by Golovin (2015). It measures the receptive vocabulary width of native Polish speakers with high precision. During the test, participants decide whether they know at least one of the word’s meanings (yes/no) and the word’s synonym in several multiple-choice questions. Test-takers responded to around 40 items within around 2.5 min.

Once the test is completed, participants provide their age and native language (Polish vs. non-Polish), and respond whether they answered honestly (when the test taker did not tick this box, their responses were not counted). Participants completed the test on their personal devices.

5.5. Polish Verbal Short-Term Memory Test

Polish MoCA verbal subtest (Gierus et al. 2015) evaluates one’s verbal memory capacity. The participant first listens to the five words, pronounced by the examiner, then s/he recalls them aloud twice: in around 2.5 min (right after PVST) and in around 10 min after the practice trial. While the MoCA protocol traditionally uses a ~5 min interval between encoding and delayed recall, our study implemented a longer delay (~10 min) with intervening tasks involving linguistic processing. This variation aims to assess the durability of intermediate-term memory (Kamiński 2017; Rosenzweig et al. 1993) under cognitive load and interference, critical circumstances of the present research design.

5.6. Context-Primed Association Test (CPAT)

CPAT combines elements of the Recognition Memory Task (Rich 2017) and the Free Association Test (see Goroshko 2001; Aitchison 1987). The test was designed with three main goals: 1. to assess the speed and accuracy of retrieving items presented outside their poetic context, i.e., verbatim memory; 2. to measure task-switching between recognition and associative tasks, assuming that both rely on verbal–semantic processing; and 3. to collect associative data, given the lack of associative dictionaries in Polish.

It is widely acknowledged that the mental lexicon is organized as a network based on associative principles, where activation of one node spreads gradually across the entire network (Collins and Loftus 1975). Word meaning is defined by both context (Schmid 2005) and background knowledge (Duch et al. 2008). The part of speech also influences the types of associations produced (Anstatt 2008; Gawarkiewicz 2024; Kurcz 1967). Thus, this test is well-suited for studying associations activated by poems and the influence of poetic content on the variation and richness of associative responses.

Participants are presented with 24 words (16 nouns and 8 adjectives) in total—five stimuli and three foils per poem, across three poems. The control items differed in their placement in a line (7 line-final vs. 7 line-middle). In addition, some of them represent poetic devices (three enjambments, four rhymes, and one—a component of a metaphor). First, participants answered a recognition question: “Was this word present in a poem?” (yes/no). If they answered “yes”, they were then asked to recall as much context as they remembered (open question with “don’t remember” accepted as a valid response). Regardless of their recognition answer, test-takers were subsequently asked to provide the first five words that came to mind upon seeing the stimulus. This sequence was designed to anchor participants in the poem’s content, so that their associations would be influenced by contextual recall.

We measured response times for the recognition and association tasks, and properties of the associations themselves.

5.7. Cloze-Test (CT)

CT evaluates participants’ verbatim memory and the ability to use contextual clues to restore gaps (Mcgee 1981, pp. 145–56; Sadeghi 2021, pp. 65–83). Up to 10 words were removed from the poem, and participants filled in the gaps from memory by pronouncing the missing word. Contrary to the common practice of removing every fifth word, we wanted to examine how the word position in line and poetic tropes influence verbatim recall. Thus, the removed items were dispersed over the text but did not follow each other. The proportion of the removed items is as follows: five initial, fifteen middle, ten final words; out of them, there were fourteen rhymes and ten enjambments. We followed the common practice of exact scoring (Sadeghi 2021, pp. 73–77) (correct vs. incorrect).

5.8. Topic Inferential Task (TIT)

TIT examines participants’ abilities to draw topical inferences within poems. The test was adapted from Cain et al. (2004) and Long et al. (1997). Participants read six two-line excerpts (twelve lines total) drawn from three poems, for example: “w niebieskich chmurach aromatu/smakował uśmiech w wargach wąskich.” Each excerpt was followed by three word pairs (prime–target), such as chmura–zadowolenie (“cloud–satisfaction”), chmura–arogancja (“cloud–arrogance”), and niepogoda–chandra (“bad weather–depression”). These pairs served as semantically appropriate, inappropriate, or filler trials, respectively (see Supplementary Materials). The participants’ task was to choose a word pair that better fits the topic of a sentence. We measured the accuracy of the responses.

6. Participants

Thirty-eight Polish native speakers (Mage = 22.5; female = 75%) participated in this study either voluntarily or via SONA panel (https://uw.sona-systems.com/ (accessed 25 November 2025)) in exchange for ECTS credits. The majority were BA students (84%); only five (13%) participants were enrolled in Philology programs, and only two studied Polish philology.

Six participants (15%) reported reading poetry at least a few times per week, preferring either regular metered or free verse poetry. Only one individual reported writing poetry. Thus, the sample was largely inexperienced in both literary training and poetry reading/writing, making it well-suited for the pilot study, establishing a baseline for future comparisons. All participants had normal or corrected-to-normal vision and were naïve to the purpose of the research.

7. Data Analysis

Raw data were checked manually for extreme and missing values and removed before applying an automatic cleaning procedure. In CPAT, 19 trials of one participant were discarded due to technical problems, 22 binary response time values of 26 participants, and 15 response time values in the associations’ test were preliminarily removed, being extremely high. Then, we removed outliers exceeding the participants’ specific cut-offs (Mean ± 3 SD). Among affective features, only 667 values were available based on Imbir’s (2016) data. RTs were additionally log-transformed to normalize distribution (available Removed for peer-review). Statistical analyses were conducted via JASP software (JASP Team 2023; version 0.18.3.), employing binomial Generalized Linear-Mixed Models, Pearson’s correlation, paired- and one-sample t-Tests, and Linear Mixed Regression Models.

8. Results

8.1. Polish Reading Proficiency Test (PRPT)

We compared the means of sentence complexity between the Polish and German versions of the SLS test to evaluate their comparability.

Table 3. Pearson’s correlations between sentence characteristics in PRPT.

Variable	1	2	3	4	5	6
1. PRPT Mean	—
2. Difference in Means SLS-Berlin vs. PRPT	−0.074	—
3. Nsyllables	0.698 ***	−0.376 ***	—
4. Nletters	0.735 ***	−0.346 **	0.969 ***	—
5. Nwords	0.676 ***	−0.406 ***	0.878 ***	0.882 ***	—
6. Nwords_long	0.580 ***	−0.228 *	0.774 ***	0.829 ***	0.616 ***	—

Note. * p < 0.05; ** p < 0.01, *** p < 0.001.

reveals strong positive correlations between sentence-level psycholinguistic features (e.g., syllables, letters, word length) and average complexity ratings, but a strong negative link with the difference scores between versions. A paired-samples t-test confirmed this effect (t(76) = 7.55, p < 0.001, d = 0.86), supported by reduced range and standard deviation in the Polish version (range_polish = 1.580; range_german = 2.680; SD_{polish_difficulty} = 0.269; SD_{german_difficulty} = 0.504). This indicates that although PRPT reflects systematic relations at the sentence level, its sentences were overall easier than those in the German original, yielding, on average, a one-point lower complexity rating and a narrower distribution of item difficulty. This suggests the necessity to consider syntactic specificities of the Polish language, which were not addressed within the current adaptation. Nevertheless, both versions are moderately to strongly correlated (r(75) = 0.58, z = 0.66) (see Supplementary Materials), supporting the construct validity of the Polish adaptation (Messick 1988).

While the reduced difficulty in the Polish version limits its discrimination among highly proficient readers, the adaptation remains psycholinguistically grounded and suitable for inclusion in the present test battery.

8.2. Polish MoCA Verbal Subscale (P-MoCA)

To demonstrate that immediate and delayed responses in the P-MoCA test measure the same phenomenon, we examined their relationship. We also expected links between P-MoCA results and other memory-related measures, i.e., response times in the recognition and free association tasks in CPAT. In addition, we included vocabulary size (PVST) scores, as this test was completed before the delayed recall and might modulate performance.

We found that participants who performed better in immediate recall also performed better in delayed recall (r = 0.832; p < 0.001), confirming a strong connection between the two measures. Better memory scores in P-MoCA (i.e., lower values) were associated with larger vocabulary size (r_{_immediate} = −0.124; p < 0.001; r_{_delayed} = −0.162; p < 0.001). This is not surprising: people who remember better also tend to possess a broader vocabulary. Finally, response times in both recognition and free association tasks tended to correlate with MoCA delayed, but not immediate, scores (r_{_recognition} = 0.110; p = 0.002; r_{_associations} = 0.117; p < 0.001) (see Appendix A,

Table A2. Pearson’s correlations between MoCA scores and memory-related variables.

Variable		1	2	3	4	5
1. MoCA_immediate	Pearson’s r	—
2. Vocabulary Size	Pearson’s r	−0.124 ***	—
3. MoCA_delayed	Pearson’s r	0.832 ***	−0.162 ***	—
4. Response Time_recognition	Pearson’s r	0.049	−0.002	0.110 **	—
5. Response Time_associations	Pearson’s r	0.070	−0.016	0.117 ***	0.198 ***	—

Note. ** p < 0.01, *** p < 0.001. Statistically significant correlations are bolded.

).

8.3. Polish Vocabulary Size Test (PVST)

The present study utilized one of the early iterations of PVST, although the exact version cannot be retrospectively confirmed. However, the lack of validated tests does not allow for measuring vocabulary size differently. The current findings should be interpreted with caution, especially in regard to psychometric properties.

Nevertheless, Pearson’s correlation analysis reveals strong associations between PVST’s key variables: larger vocabulary scores are linked to higher accuracy (r = 0.279) and longer test duration (r = 0.274), while test accuracy is negatively related to test duration (r = −0.258) (see Supplementary Materials). This suggests that participants with higher vocabulary scores were more engaged with the task; longer total duration likely reflects the decision-making process, which in turn supports higher accuracy. At the same time, the negative accuracy–duration link indicates that more confident responses required less time per decision, whereas hesitation was associated with lower accuracy.

PVST yielded a normally distributed range of scores (M = 46.526; SD = 23.861; skewness = 0.981), suggesting meaningful individual variation in vocabulary size among participants. However, participants varied significantly in their receptive vocabulary knowledge (range 18,000–108,000; M = 46.270; SD = 23.821; Supplementary Materials).

8.4. Context-Primed Associations Test (CPAT)

The main target of CPAT—the associations provided by participants—was not analyzed in the present paper, as an in-depth qualitative analysis is required. These data will be published separately.

Table 4. Context-Primed Associations Test descriptives.

	Response Time, log(RT)	Response_Timecontext log(RT_cont.)	Response_Timeassociations log(RT_assoc.)	Accuracy
Nresponses	817	424	863	862
Mean (in seconds) (SD)	6.156 (3.393), 1.693 (0.484)	13.180 (6.693), 2.246 (0.477)	17.451 (5.508), 2.794 (0.361)	0.665 (0.472)
p-value	<0.001, <0.001	<0.001, 0.009	<0.001, <0.001	<0.001
Range	2.000–19.42 s, 0.690–2.97	3.85–37.50 s, 1.35–3.62	7.20–38.81 s, 1.97–3.66

presents descriptives of key variables. Overall, readers performed above chance in the recognition memory task, having ~67% of correct responses. Separated by the poem, results are slightly more accurate for Zagajewski’s poem compared to the other two (M_zagajewski = 0.719; M_szymborska = 0.626; M_herbert = 0.649).

We then ran a statistical model (binomial GLMM) to see what modulates readers’ certainty when deciding whether a word had appeared in the poem. Here, two factors are crucial: the speed of decision-making and the placement of the word. First, the quicker the response, the more accurate it is (B = −0.919; t = −3.789, p < 0.001)—delays and mistakes are both due to uncertainty.

Second, line-final words are remembered better than line-middle ones (B = 1.356; t = 3.291; p = <0.001; Figure 2). However, the advantage of line-final words fades if the reader hesitates: as response time increases, the memory boost for line-final words disappears (see Appendix A,

Table A3. Anova summary of General Linear Mixed-Model on response accuracy in CPAT.

Effect	df	ChiSq	p
Intercept	1	29.188	<0.001
position_in_line	1	7.040	0.008
log(response time)	1	15.076	<0.001
position in line × log(RT)	1	5.276	0.022

Note. ID (36) was used as a random factor with 409 observations. A binomial general mixed model was employed, and accuracy was used as a dependent factor tested with the likelihood ratio test method. Model fit is singular, which might inflate the reported p-value.

). In other words, the effect of ending words is strongest when recognition is immediate; otherwise, the word placement effect weakens.

To explore the relationship between lexical retrieval in the recognition task, context restoration, and performance in the free association task, we conducted a Pearson’s correlation analysis (see Supplementary Materials). The results revealed that word recalling is linked to the generation of associations, but not to the production of the context (RT_recognitionxRT_associations r = 0.198; p < 0.001; RT_recognitionxRT_context r = 0.038; p = 0.449). So, the quick decision-making in word recognition expedites associative processes, evidencing the overlap between these cognitive performances and a possible prime effect.

Lastly, we examined which factors influence the generation of associations and found an interesting pattern: unlike in the recognition task, where line-final words improved accuracy, in the free association task, they instead led to longer hesitations before producing associations (B = 0.238; t = 4.493; p < 0.001). However, two interactions moderated this effect. First, the interaction between word length and line position (B = −0.023; t = −3.472; p < 0.001) indicated that longer words at the end of the line slowed responses less than shorter ones. Second, the interaction between word length and word frequency (B = −0.027; t = −2.844; p = 0.007) showed that short, high-frequency words contribute to the associative process. Results suggest that while line-final words generally decelerate associative responses, short and frequent words at the end of a line facilitate the generation of associations.

8.5. Cloze-Test (CT)

In CT, overall readers’ performance was below average, with only 33% of correct responses (Mean = 0.334; SD = 0.472). Yet there were differences in accuracy across poems: while reconstruction of Herbert’s and Szymborska’s poems is comparable (M_{_accuracy_herbert} = 0.383; M_{_accuracy_szymborska} = 0.354), Zagajewski’s poem significantly challenged participants (M_{_Accuracy_zagajewski} = 0.265).

We first obtained that word length and valence and their interaction crucially affected participants’ performance (

Table 5. Fixed Effects Estimates: valence and word length impact on CT accuracy.

Term	Estimate (SE)	t	p
Intercept	8.022 (2.441)	3.286	0.001
length	−1.082 (0.343)	−3.160	0.002
valence	−1.081 (0.410)	−2.635	0.008
valence * length	0.127 (0.058)	2.185	0.029

Note. ID (36) was used as a random factor with 660 observations (450 observations were removed due to the missing data); A binomial general mixed-model was employed, and accuracy was used as a dependent factor tested with the likelihood ratio test method; Model fit is singular, which might inflate the reported p-value.

). Results demonstrate that short (up to five letters) and pleasant (valence ~ 4) words were remembered and restored much better compared to long and unpleasant ones (see here Appendix A,

Table A5. Estimated Marginal Means of length × valence on response accuracy (CT).

			CI 95%
Length	Valence	Estimate_Accuracy (SE)	Lower	Upper
4.604	4.432	0.697 (0.055)	0.581	0.793
6.502	4.432	0.464 (0.044)	0.380	0.551
8.399	4.432	0.246 (0.036)	0.181	0.324
4.604	5.549	0.570 (0.43)	0.484	0.652
6.502	5.549	0.393 (0.35)	0.326	0.463
8.399	5.549	0.240 (0.37)	0.175	0.320
4.604	6.666	0.432 (0.54)	0.330	0.540
6.502	6.666	0.327 (0.43)	0.248	0.417
8.399	6.666	0.237 (0.49)	0.154	0.346

Note. responses are not on the response scale. Significant changes in rates are bolded.

). Word valence is a stronger predictor of accuracy than word length; thus, increasing valence penalizes the advantage of length, and the accuracy decreases.

Most intriguing is the interaction of affective word properties while reconstructing a poetic content. Calm words (low in arousal), such as palmy (palms) or trwać (to last) restored significantly better (B = −7.186; t = −4.947 p < 0.001) than exciting ones (high in arousal). Decrease in imageability score, i.e., hard-to-imagine words, (B = −2.535; t = −4.119; p < 0.001) and increase in concreteness, i.e., abstract words3 (B = 1.922; t = 3.114; p = 0.002), additionally contribute to accuracy (see Appendix A,

Table A6. Estimated Marginal Means of imageability × concreteness on response accuracy (CT).

			CI 95%
Imageability	Concreteness	Estimate (SE)	Lower	Upper
5.553	2.082	0.319 (0.123)	0.134	0.587
5.553	2.082	0.407 (0.072)	0.276	0.552
5.553	2.082	0.501 (0.055)	0.394	0.607
6.810	3.628	0.353 (0.066)	0.236	0.491
6.810	3.628	0.296 (0.042)	0.220	0.385
6.810	3.628	0.345 (0.054)	0.155	0.365
8.066	5.175	0.388 (0.066)	0.268	0.523
8.066	5.175	0.205 (0.058)	0.114	0.343
8.066	5.175	0.095 (0.044)	0.037	0.223

Note. responses are not on the response scale. Significant changes in rates are bolded.

). Thus, words such as żyć (to live) or słodko (sweet) were better reconstructed than concrete and easy-to-imagine dywan (carpet) or glina (clay).

Further, the concreteness of a word modulates the effects of lexical imageability. Concrete and imaginative words, e.g., ściana (wall), are restored better, but an increase in word abstractness penalizes the effect of lexical imagery (Appendix A, Table A6). So, abstract but easy-to-imagine words, e.g., smakować (to taste), were remembered worse.

8.6. Topic Inferential Task (TIT)

TIT revealed a main effect of a poem in favor of Herbert’s text compared to the other two (Figure 3). The response accuracy to Herbert’s verses was substantially higher, approaching the ceiling effect (B = 0.985; CI95% = [0.577, 1.000]), whereas responses to Szymborska’s (B = 0.575; CI95% = [0.434, 0.705]) and Zagajewski’s (B = 0.530; CI95% = [0.410, 0.646]) poems were more evenly divided. No reliable effects of priming pairs or priming pair x poem were observed.

9. Discussion and Conclusions

The present pilot study aimed to develop and test a psycholinguistic battery tailored for the empirical investigation of Polish poetry processing. Our goals were 1. to create a tool for linguistic profiling of readers and 2. to explore how lexical, textual, affective, and poetic word properties influence cognitive and behavioral responses.

RQ1. 

What means are best-fitted for comprehensive empirical study of modern poetry?

As preliminary measures, we selected the SLS-Berlin Reading Proficiency Test (Lüdtke et al. 2019), Vocabulary Size test (Fokin et al. 2025), and verbal MoCA subscale (Gierus et al. 2015), all previously shown to measure crucial text comprehension skills (Chateau and Jared 2000; Just and Carpenter 1980; Matchin et al. 2024). The control experiments included the Context-Primed Associations Task (CPAT), the Cloze-Test (CT), and the Topic Inferential Task (TIT), intended to measure verbatim and content memory, associative processes, participants’ ability to restore the context, and topic inference. The correlations among the scores of all tests included in the study are summarized in Figure 4, revealing potentially similar cognitive processes underlying word remembrance and associative mechanisms related to delayed memory recall. Contextual word restoration, in turn, requires activation of the mental lexicon and verbal memory, resulting in a strong association between the Cloze Test scores and the MoCA and PVST results.

The Polish Reading Proficiency Test (PRPT) showed a moderate-to-strong correlation with the German original and consistent relations between sentence-level features, supporting its validity for further application. Yet, PRPT proved easier: in contrast to the German original, which strongly predicted score differences (r = −0.850, p < 0.001), our adaptation did not (r = −0.074; p = 0.524), evidencing its reduced sensitivity.

We observe moderate-to-strong associations between 1. immediate and delayed lexical rehearsal (r = 0.83); 2. weak but statistically significant links between P-MoCA delayed scores and the vocabulary size (r = −0.18), recognition memory (r = −0.11), and association generation times (r = −0.12); 3. moderately strong association between times in recognition and association generation tasks (r = 0.20) (CPAT); 4. strong connection between context restoration scores and the vocabulary size (r = 0.45).

The Topic Inferential Task (TIT) showed an imbalance in design, leading to uneven responses across poems and no expected priming effects. However, TIT’s results may illustrate differences in topic distinguishing: the semantic mainline in Herbert’s poem—the disappearance of the narrator’s father—is depicted in a more straightforward way, while the meanings of Szymborska’s and Zagajewski’s poems are more fuzzy and dispersed, with no pronounced main topic. Thus, the results may demonstrate the higher semantic and narrative coherence in the first poem compared to the last two, and participants identify it more easily. However, since TIT did not demonstrate the expected priming effects, we decided to remove it from the battery until calibrated. This will also streamline the test battery and reduce participants’ cognitive load.

To sum up, we confirmed our hypothesis by revealing the capability of the proposed test battery to capture meaningful links across the measures, supporting the feasibility of this two-step design for further studies. PRPT requires further calibrations and stimulus set refinements, and TIT, in its current form, cannot yet capture topic inference. Overall, the combined battery demonstrates strong potential for linguistic profiling of readers and examining memory, vocabulary size, and associative processes.

RQ2. 

How does reading modern poetry impact cognitive processes involved in verbatim and content remembrance, associative processes, and topic identification?

We hypothesized that the control experiments in their current state may capture differences in verbatim memory, content remembrance, and the generation of associations.

The main finding is a moderately strong interaction between word recognition and association–generation times (Figure 3), with no relation to context recall times. Two perspectives may account for this. First, textual information is stored in an episodic buffer (Baddeley 2000, 2003), from which it can later be extracted, and the storage process relies on the similar cognitive functioning that supports lexical retrieval from long-term memory. Second, quick and correct recognition responses may prime the associative network through preactivation, enabling faster retrieval of associations. By contrast, context recall was intentionally designed as a filler to reduce guessing, but it also demands recalling large pieces of textual representations, which are likely stored in the intermediate-term memory (Kamiński 2017), where participants often fail (see below).

The Context-Primed Associations Test’s results suggest that readers likely construct a situation model (Zwaan and Radvansky 1998)—a mental representation of a poem’s narrative, imagery, and structure. However, the latest model proposed by Zwaan (Zwaan 2025) cannot fully explain the observed relations. Our results imply the presence of an intermediate memory store, where larger chunks of textual information might be maintained. This store appears to be a buffer between short-term memory, accessed via a recognition memory task, and long-term memory, engaged during the association generation. It seems task-dependent: knowing the upcoming task, readers initiate attention-guiding mechanisms (e.g., Moss and Schunn 2015). It allows them to retrieve poem-specific information beyond standard working-memory limits (Baddeley 2003). Thus, comprehenders request a situation model through this buffer, performing both working- and long-term memory-related tasks. The quality of this model shapes both recognition accuracy and association generation, which emphasizes mental simulation of time, space, and causality. It is further evidenced from low but significant correlations between recognition and association generation times and delayed memory rehearsal, with no relation to immediate recall.

Context recall is strongly related to vocabulary size and delayed memory rehearsal. It suggests that CT requires additional cognitive consumption, accessing textual representations, appropriate word selection, and then grammatical manipulations with selected items (Appendix A,

Table A8. Stimuli difficulties assessment based on Rasch modelling.

Item	English	Mean	xsi_ Difficulty	Item	English	Mean	xsi_ Difficulty
zagubione	lost (adj.pl.)	0.027	4.044	żyję	to live (1 pers. sing.)	0.297	1.016
przenika	to penetrate (3pers.sing.)	0.054	3.278	głowę	head (sing.acc.)	0.324	0.876
smakował	to taste	0.056	3.255	zdejmował	to take off (past. 3 pers.sing.)	0.389	0.534
starości	old age (sing. gen.)	0.081	2.806	zdjęto	removed (past. impers.)	0.389	0.534
Wstyd	shame	0.108	2.454	odwrócił	to turn around (past pers.sing.)	0.417	0.395
odległych	distant (pl.loc.)	0.111	2.426	oczy	eyes (pl.acc.)	0.444	0.260
kochałiście	to love (2.pers. pl.)	0.135	2.168	światło	light (sing.acc.)	0.459	0.187
glina	clay (nom.sing)	0.162	1.925	ścianie	wall (sing.loc.)	0.514	−0.071
przedmiotów	objects (pl.gen.)	0.189	1.710	apetytu	appetite (sing.gen.)	0.514	−0.071
powietrza	air (sing.gen.)	0.189	1.710	ojciec	father (1 pers. sing.)	0.583	−0.409
słodko	sweet (adv.)	0.189	1.710	dywanie	carpet (sing.loc.)	0.583	−0.409
wieczności	eternity (sing.acc.)	0.216	1.517	palmy	palms (pl.nom.)	0.583	−0.409
szarej	grey (sing.fem.loc.)	0.243	1.339	trzystu	three hundred (sing.gen.)	0.676	−0.878
krople	drops (pl.acc.)	0.278	1.132	rzeczy	things (pl.acc.)	0.892	−2.444
				trwa	to last (3 pers.sing.)	0.919	−2.791

Note. The greater the difficulty, the less often participants responded correctly. Thus, “trwa” was accurately filled by most of the readers, whereas “zagubione” was filled by only a few. One item (cichutko) was excluded from the analysis due to the absence of the correct responses.

). In line with Rugg et al. (1998), who showed that recognition is more accurate than recall and both involve distinct neural activities, our findings confirm: CT taps into long-term memory processes but not word recognition alone and draws the situation model for retrieval and usage of a context when filling the gaps.

Yet, the participants’ mean performance in CT was low (M = 0.334), approaching a floor effect. To evaluate whether it is a task-design issue, we applied a Rasch model (R Core Team 2025; Robitzsch et al. 2020). It enables observing how well stimuli fit the measured dimension, i.e., the ability to restore poetic content. The model indicated trustworthy item functioning (EAP Reliability = 0.792; Correlations = 0.975; Appendix A,

Table A7. The dichotomous Rasch model summary on CT responses.

Metrics	Statistics
Deviance	1005.08
Log Likelihood	−502.54
AIC	1065
BIC	1113
EAP Reliability	0.792
Variance	0.951
Correlations	0.975
Regression Coefficients	0

Note. EAP Reliability reflects the precision of the model; Covariances is the variance of the latent dimension; Correlations are connections between latent dimensions.

; Supplementary Materials). High item and person fit suggest that CT reliably measures participants’ ability to restore poetic content, despite overall low scores.

These low scores may instead reflect properties of modern poetry itself: lack of coherence, resistance to prediction, and high cognitive demand. Participants inexperienced in modern poetry reading may additionally struggle to use contextual cues and text coherence to restore the content. Unlike rhymed or metered verse, which offers prosodic regularities that aid memory (e.g., Lea et al. 2021), the selected poems were rich in syntactic interruptions, enjambments, and semantic complexity. Moreover, preliminary Exploratory Factor Analysis on the lexical content (Fokin, forthcoming) revealed the prevalence of abstract, hard-to-imagine, and calm words with an increase in perceived significance, engaged in the abstract–significance continuum. Such forms of lexical content contain a foreground trajectory of perception (Kuiken and Shawn 2017), challenging participants and requiring reflective and close reading. Non-expert readers, in turn, may not be intended for in-depth analysis or puzzle-solving, preferring unambiguous communication (Porter and Machery 2024). Additionally, they lack clear “reference points” to build a qualitative textual representation. As a result, readers fail in successful content restoration due to the lack of expertise, motivation, and inference skills.

It is worth noting the poem’s influence on participants’ memory: words from Zagajewski’s poem were better recognized, while its content was less accurately restored. This may illustrate the phenomenon of foregrounding: Zagajewski’s text is non-uniform, with numerous punctuation marks and a disproportion between long and short lines. The absence of a clear narrative streamline disrupts attention, making it dispersed. As a result, some particular words are remembered, whereas the content is not. Such structural complexity makes it difficult to form a coherent mental image of the piece, yet at the same time, it captures attention, which may facilitate short-term retention of individual words.

In sum, consistent with our hypothesis, the proposed tests show promise in capturing memory-based processes in poetry reading. Although certain tasks require refinement and the full battery may need adjustment to reduce cognitive load, the study yielded valuable insights. First, cognitive processes involved in reading Polish poetry appear to align with those observed for linguistic materials in other languages, indicating potential universality in language processing mechanisms. Our participants were inexperienced poetry readers, and thus their responses may reflect general, background-independent mechanisms of poetry reception—processing poetry much like other texts. Linguistic properties of words modulate memory encoding and retrieval, either facilitating or complicating them. Importantly, we observed that poetic devices and affective word properties—unique to each poem and to Polish poetic language—modulate contextual restoration. This finding demonstrates that readers are guided by the poem’s affective contour, which directs attention and triggers specific behavioral responses. In this sense, the stimuli (poems) and readers’ responses (words) are both unique and language-dependent. Our findings thus provide novel cognitive and behavioral evidence of how readers engage with Polish poetic texts.

RQ3. 

What is the role of textual, lexical, poetic, and affective features in framing these effects?

We hypothesized that affective word features are key drivers in poetry reception; word position in line is a salient predictor of remembrance, being refined by poetic tropes, and lexical word properties are subordinated by affective ones, which can either enhance or diminish the impact of lexical properties.

The main discovery here is the central role of word line position in recognition and association–generation processes. More accurate responses in the recognition task (CPAT) are linked with faster responses to line-final words compared to middle words (Appendix A,

Table A4. Anova summary of Linear Mixed-Model on response time in the Free Association Task in CPAT.

Term	df	F	p
length	1, 76.78	0.117	0.773
position in line	1, 105.55	20.472	<0.001
word frequency	1, 65.54	3.335	0.072
Length × position in line	1, 279.23	12.175	<0.001
Length × word frequency	1, 46.78	8.121	0.006

). This result aligns with earlier findings in neurocognitive poetics (Fechino et al. 2020; Jacobs 2015; Lüdtke et al. 2014) and empirical stylistics (Menninghaus and Wallot 2021), showing that final words are often semantically loaded and foregrounded (e.g., through rhyme or enjambment). Such words are likely more memorable due to wrap-up and savoring effects (Jacobs 2014; Menninghaus and Wallot 2021). Readers anticipate closure and therefore allocate more attention to the line’s end, while middle words remain less salient and attract less focus.

However, association–generation shows a different pattern. Here, word position is modulated by lexical properties such as word length and frequency. Contrary to recognition, final words slow down associative generation (B = 0.237, SE = 0.052; t = 4.525, Supplementary Materials). This may stem, again, from their poetic foregrounding: while recognition can be facilitated by familiarity, association becomes more difficult if these words carry semantic weight, involve syntactic disruption (e.g., via enjambment), or create incoherent textual representations that reduce priming effects. The negative effect of line-final position is reduced by length, with short final words leading to quicker associations. Likewise, length × frequency interactions show that short, high-frequency words accelerate association production. Taken together, these results demonstrate that word position and fundamental lexical properties are key predictors of memory-based and associative processes. They may reflect principles described by Kolmogorov and Zipf (Ehret 2018; Linders and Louwerse 2023) as well as known mechanisms of memory retrieval and storage (Baddeley 2003; Lea et al. 2021; Rosenzweig et al. 1993; Rugg et al. 1998): short and high-frequency items are easier to identify, remember, and manipulate. Readers focus more on foregrounded elements (see Kuiken and Jacobs 2021; Hakemulder et al. 2025; Hakemulder 2004), making them easier to recognize because of their salience. Yet, when accessing long-term lexical storage, these same elements can hinder retrieval, while basic lexical properties such as length and frequency reliably reduce the negative effects of line position.

Finally, short and negatively valenced words were restored more accurately in CT compared to positive or long-word items. Moreover, an increase in valence (words becoming more positively-colored) decreased the negative effect of long words, leading to higher accuracy. Higher remembrance of negatively-colored words was previously observed by Inaba et al. (2005), and our results are aligned with the prior ones.

Analysis of affective features revealed the crucial role of arousal, imaginability, and concreteness, as well as their interactions. Hardly imagined and abstract words were better restored than concrete and easy-to-imagine ones, contrasting well-established patterns that concrete and imaginative words are better understood, imagined, and remembered (Aka et al. 2021; Altarriba et al. 1999; Imbir 2016). The effect of arousal was moderated by imaginability, reflecting that imaginative content plays a greater role than emotional arousal in poetry processing. A possible explanation here is that abstract and hard-to-imagine words may induce associative and reflective processes—a kind of mind-drift—and readers may react to them as semantically salient or foregrounded figures (see Fokin, forthcoming). Previously, we assumed that CT involves multi-step cognitive processes. Thus, only here do we observe the impact of affective items, suggesting that when restoring content, participants may access long-term lexical storage, and abstract words may trigger larger networks than concrete and familiar ones. So, despite failing to confirm a prominent role of affective features in overall memory processes, we revealed their importance for high-order lexical manipulations and access. We also established the key role of word position in line on readers’ memory, suggesting the importance of visuo-spatial organization for memory-based functions.