Engagement for emotional prototypicality is shaped by word frequency in reading: evidence from eye movements

Tongwen Hu; Xinying Yuan; Linlin Zhang; Yuru Cheng; Jingxin Wang

doi:10.1017/S0142716425100222

Engagement for emotional prototypicality is shaped by word frequency in reading: evidence from eye movements

Published online by Cambridge University Press: 23 October 2025

Yuru Cheng and

Tongwen Hu: Affiliation:
College of Education, Anqing Normal University, Anqing, China Key Research Base of Humanities and Social Sciences of the Ministry of Education, Academy of Psychology and Behavior, Tianjin Normal University, Tianjin, China
Xinying Yuan: Affiliation:
Key Research Base of Humanities and Social Sciences of the Ministry of Education, Academy of Psychology and Behavior, Tianjin Normal University, Tianjin, China
Linlin Zhang: Affiliation:
Business School, Northeast Normal University, Changchun, China
Yuru Cheng: Affiliation:
Key Research Base of Humanities and Social Sciences of the Ministry of Education, Academy of Psychology and Behavior, Tianjin Normal University, Tianjin, China
Jingxin Wang*: Affiliation:
Key Research Base of Humanities and Social Sciences of the Ministry of Education, Academy of Psychology and Behavior, Tianjin Normal University, Tianjin, China
*: Corresponding author: Jingxin Wang; Email: wjxpsy@126.com

Article contents

Abstract
Introduction
Experiment 1
Experiment 2
General discussion
Limitations
Conclusion
Replication package
Competing interests
Declarations of AI-based tools
Funding statement
References

Rights & Permissions

Abstract

Emotional prototypicality (EmoPro) refers to the degree of emotional representativeness of a word and influences emotional content extraction at the lexical level. However, its effect on more complex semantic structures, such as sentences, remains unclear. This study employed eye-tracking to examine the EmoPro effect during Chinese sentence reading. EmoPro (high vs. low) was manipulated in two experiments, with sentences containing either a positive or negative valence target word. The lexical frequency of these target words was also manipulated to assess its influence on emotional semantics activation during reading. The results show that high EmoPro words consistently evoke greater engagement during both early and late word processing, demonstrating a significant advantage in emotional information retrieval. Word frequency influenced this processing advantage differently for words with a positive or negative valence. For positive valence, high-frequency words facilitated emotional extraction for high EmoPro words; for negative valence, low-frequency words enhanced their salience, leading to faster emotional retrieval. These findings provide the first evidence that EmoPro significantly impacts the processing of words in natural reading. The findings also highlight a complex interplay between affective and linguistic information in emotional semantics embodiment, with word frequency playing a pivotal role in shaping its depth during reading.

Keywords

Emotional prototypicality Emotional words Natural reading Word frequency

Information

Type: Original Article
Information: Applied Psycholinguistics , Volume 46 , 2025 , e36

DOI: https://doi.org/10.1017/S0142716425100222 [Opens in a new window]
Creative Commons: This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution and reproduction, provided the original article is properly cited.
Copyright: © The Author(s), 2025. Published by Cambridge University Press

Introduction

Emotions are a fundamental part of human nature, influencing both physiological and psychological processes (Pekrun et al., Reference Pekrun, Marsh, Elliot, Stockinger, Perry, Vogl, Goetz, van Tilburg, Lüdtke and Vispoel2023) and playing a key role in cognitive functions such as attention (Sharif & Mahmood, Reference Sharif and Mahmood2023). Emotionally salient stimuli, typically marked by high arousal and either positive or negative valence, tend to capture attention and be processed with higher priority, resulting in faster and more accurate recognition (Rouhani et al., Reference Rouhani, Niv, Frank and Schwabe2023). Research on emotionally salient stimuli has focused on visual stimuli, such as pictures (Zsidó et al., Reference Zsidó, Bali, Kocsor and Hout2023), facial expressions (Schindler & Bublatzky, Reference Schindler and Bublatzky2020), and emojis (Kaye et al., Reference Kaye, Rocabado, Rodriguez-Cuadrado, Jones, Malone, Wall and Dunabeitia2023). In contrast, much less attention has been given to how emotions are perceived and interpreted through language, especially in the absence of salient perceptual cues (Hinojosa et al., Reference Hinojosa, Herbert and Kissler2023; Speed & Brysbaert, Reference Speed and Brysbaert2023). Investigating emotional processing during language comprehension is nevertheless important, as written language can convey complex emotional information without visual or sensory cues and plays a vital role in socialization (Snefjella et al., Reference Snefjella, Lana and Kuperman2020). Accordingly, the present study aimed to uncover how emotions conveyed through language are processed.

As words play a central role in language comprehension (Rayner, Reference Rayner2009), research has focused on understanding the influence of emotion on cognitive processes in word recognition (Ferré et al., Reference Ferré, Guasch, Stadthagen-González, Hinojosa, Fraga, Marín and Pérez-Sánchez2024; Hinojosa et al., Reference Hinojosa, Moreno and Ferré2020). The emotional content of words is typically characterized in terms of arousal and valence, where words might be high or low arousal and have either a positive, or negative valence (Scott et al., Reference Scott, O’Donnell, Leuthold and Sereno2009). Research has shown that these emotional word characteristics can influence processes of word recognition across a range of tasks (Hinojosa et al., Reference Hinojosa, Herbert and Kissler2023). In lexical decision tasks, positive and negative valence words are recognized more quickly and accurately than neutral words (Vinson et al., Reference Vinson, Ponari and Vigliocco2014), suggesting that the lexical representation of emotional words is swiftly activated and readily available. Similarly, in masked priming paradigms, emotional word primes typically produce larger semantic priming effects than neutral word primes (Spruyt et al., Reference Spruyt, Hermans, Houwer and Eelen2002), revealing that the emotional content of words can influence their sub-threshold processing (Gaillard et al., Reference Gaillard, Del Cul, Naccache, Vinckier, Cohen and Dehaene2006). Other studies show an influence of emotional content in implicit semantic processing tasks (which do not require participants to make overt semantic judgments). Studies using the dot probe (Sutton & Altarriba, Reference Sutton and Altarriba2011) or the Stroop tasks (Siakaluk et al., Reference Siakaluk, Knol and Pexman2014) typically manipulate the valence of a cue word, requiring participants to indicate either the location (dot probe) or the color (Stroop) of the stimulus. Results from these tasks show that participants are better at judging the location of emotional words compared to neutral words (Sutton & Altarriba, Reference Sutton and Altarriba2011), but poorer at identifying the color of emotional words compared to neutral words (Siakaluk et al., Reference Siakaluk, Knol and Pexman2014). These findings are explained in terms of emotional words having a greater influence on attention allocation, by more strongly capturing a reader’s attention. Electrophysiological studies additionally reveal a neural basis for this attentional bias by revealing larger effects for emotional compared to neutral words in event-related potentials (ERPs), with differential effects observed in the Early posterior negativity (EPN) (250–300 ms, Schindler & Kissler, Reference Schindler and Kissler2016) and Late positive complex (LPC) (500–700 ms, Kaltwasser et al., Reference Kaltwasser, Ries, Sommer, Knight and Willems2013). Overall, these effects suggest that cognitive engagement with emotional words, even in the absence of explicit emotional processing demands, activates their emotional content, captures attention, and plays a central role in lexical access.

There is nevertheless disagreement over the nature of the effects. Some studies suggest a positive valence bias, such that participants exhibit a larger word recognition advantage for positive compared to neutral words than for negative compared to neutral words (e.g., Hinojosa et al., Reference Hinojosa, Mercado, Albert, Barjola, Peláez, Villalba-García and Carretié2015; Kuperman et al., Reference Kuperman, Estes, Brysbaert and Warriner2014; Scott et al., Reference Scott, O’Donnell, Leuthold and Sereno2009). Others show a negative valence bias, with a larger word recognition advantage for negative compared to neutral words than for positive compared to neutral words (e.g., Espuny et al., Reference Espuny, Jiménez-Ortega, Casado, Fondevila, Muñoz, Hernández-Gutiérrez and Martín-Loeches2018; Sutton & Altarriba, Reference Sutton and Altarriba2011). Yet other studies suggest that emotional words, whether positive or negative, have general processing advantages over neutral words (e.g., Ferré et al., Reference Ferré, Fraga, Comesaña and Sánchez-Casas2015; Kousta et al., Reference Kousta, Vinson and Vigliocco2009). Notably, not all emotional words rapidly activate emotional content during lexical processing (Haro et al., Reference Haro, Calvillo, Poch, Hinojosa and Ferré2023). Research has shown substantial variability among emotional words in their capacity to trigger emotional activation, with some being significantly more effective than others (Zhang et al., Reference Zhang, Wu, Meng and Yuan2017). These differences have often been overlooked in previous studies, contributing to inconsistent findings regarding valence effects (Betancourt et al., Reference Betancourt, Guasch and Ferré2023). As such, it is essential to clarify how variability in emotional content activation for words influences lexical processing (Wu & Zhang, Reference Wu and Zhang2020).

This has led some researchers to classify emotional words in terms of how strongly their emotional content influences processes of word recognition. A key distinction concerns the extent to which words explicitly represent emotional meanings versus evoke emotional experiences (Pavlenko, Reference Pavlenko2008). Emotion-label words, such as “happy” and “sad,” directly denote emotional states. In contrast, emotion-laden words, such as “failure” and “death,” evoke emotional responses through associative meanings, despite not explicitly referring to emotions. Although both types convey emotional content, it is argued that they differ in their lexical processing due to the distinct mechanisms through which emotional meaning is activated during lexical access (Wu & Zhang, Reference Wu and Zhang2020). For instance, in semantic priming tasks (Kazanas & Altarriba, Reference Kazanas and Altarriba2015, Reference Kazanas and Altarriba2016) and emotion classification tasks (Gu & Chen, Reference Gu and Chen2024), emotion-label words have been shown to activate emotional semantics more rapidly than emotion-laden words. This results in stronger priming effects, faster emotional valence judgments, and more intense emotional responses. These findings suggest that emotion-label words have a distinct advantage in semantic memory accessibility, enabling them to evoke emotional experiences more efficiently. This distinction has been consistently observed across multiple languages, including English (Knickerbocker et al., Reference Knickerbocker, Johnson and Altarriba2015), Chinese (Zhang et al., Reference Zhang, Wu, Yuan and Meng2019), and Spanish (Betancourt et al., Reference Betancourt, Guasch and Ferré2024), and across tasks, such as the Simon task (Altarriba & Basnight-Brown, Reference Altarriba and Basnight-Brown2011) and the Flanker task (Zhang et al., Reference Zhang, Wu, Yuan and Meng2019). Despite these findings, there is no clear consensus on the differentiation between the two word types. Some researchers argue that emotion-label words trigger stronger emotional activation during lexical processing (Zhang et al., Reference Zhang, Wu, Meng and Yuan2017), while others find the opposite (Liu et al., Reference Liu, Fan, Tian, Li and Feng2023).

A potential reason for this inconsistency is the lack of an objective and reliable standard for classifying emotional word types, as most researchers rely on subjective judgments (Hinojosa et al., Reference Hinojosa, Moreno and Ferré2020). To address this limitation, researchers have begun selecting emotional words based on their degree of typicality in representing specific emotions, defined as emotional prototypicality (EmoPro) (Pérez-Sánchez et al., Reference Pérez-Sánchez, Stadthagen-Gonzalez, Guasch, Hinojosa, Fraga, Marín and Ferré2021). This approach, inspired by Rosch’s (Reference Rosch, Rosch and Lloyd1978) work and grounded in prototype theory (Russell, Reference Russell1991), proposes that emotion concepts, similar to other “fuzzy” categories like “bird,” are organized in a prototypical structure and vary in how strongly they represent a specific emotion. This implies that “some words are more representative of the semantic category of emotions than others, and therefore convey a particular emotional meaning more strongly” (Ferré et al., Reference Ferré, Guasch, Stadthagen-González, Hinojosa, Fraga, Marín and Pérez-Sánchez2024, p. 745). For instance, “sadness” is considered a more prototypical emotion word than “failure.”

The prototypicality approach, widely employed in emotion research over the past few decades (e.g., Galati et al., Reference Galati, Sini, Tinti and Testa2008; Shaver et al., Reference Shaver, Schwartz, Kirson and O’Connor1987), involves asking participants to rate the extent to which candidate words refer to an emotion. This method has been applied across various languages, including French (Niedenthal et al., Reference Niedenthal, Auxiette, Nugier, Dalle, Bonin and Fayol2004), Spanish (Pérez-Sánchez et al., Reference Pérez-Sánchez, Stadthagen-Gonzalez, Guasch, Hinojosa, Fraga, Marín and Ferré2021), and more recently in Chinese (Wu, Reference Wu2023). This approach has encouraged researchers to examine how EmoPro influences language processing. Lexical decision studies in Spanish (Haro et al., Reference Haro, Calvillo, Poch, Hinojosa and Ferré2023) and Chinese (Wu et al., Reference Wu, Wu and Gao2024) show that words with high EmoPro facilitate more efficient visual word recognition by enhancing access to emotional content. Such findings suggest that highly prototypical emotion words may have stronger semantic ties to emotional information, making them more effective in activating the emotional meanings they convey.

According to the Embodied Theory of Semantic Representation (Vigliocco et al., Reference Vigliocco, Meteyard, Andrews and Kousta2009), the advantage of high EmoPro words in retrieving emotional information likely arises from their semantic representations, which develop in tandem with affective experiences during socialization and emotional interactions. This process integrates affective experiential information (e.g., feelings, interoception; Ferré et al., Reference Ferré, Guasch, Stadthagen-González, Hinojosa, Fraga, Marín and Pérez-Sánchez2024) into word meanings (Tang, Fu, et al., Reference Tang, Fu, Wang, Liu, Zang and Kärkkäinen2023). As a result, activating the meaning of a high EmoPro word simultaneously triggers its associated emotional experiences. In contrast, low EmoPro words have semantic representations more akin to general vocabulary, with an initial emphasis on conceptual understanding during learning. Their affective experiential associations emerge more gradually as semantic knowledge becomes more refined (Tang, Fu, et al., Reference Tang, Fu, Wang, Liu, Zang and Kärkkäinen2023), meaning that emotional activation typically follows full semantic processing. Current evidence for the processing advantage of high EmoPro words primarily comes from lexical decision tasks (e.g., Hao et al., 2023; Wu et al., Reference Wu, Wu and Gao2024). In these tasks, words are presented individually for recognition, and so are not encountered within a normal language context or read normally as part of a sentence. As a result, such tasks may provide an incomplete picture of how emotional language is processed in real-world contexts and when reading text naturally (Alexander & Buzzell, Reference Alexander and Buzzell2024). Extrapolating findings from isolated-word studies to natural language processing poses significant challenges (Alexander & Buzzell, Reference Alexander and Buzzell2024). In everyday contexts, emotional information is typically conveyed through sentences that embed emotional words within broader linguistic structures (Scott et al., Reference Scott, O’Donnell and Sereno2012). Accordingly, it is also important to investigate EmoPro under more naturalistic reading conditions, to move fully understand its influence on everyday language processing. To our knowledge, no research has yet explored EmoPro effects in natural reading contexts. Eye-tracking offers a powerful tool for investigating word recognition processes in natural reading, as it allows for the continuous recording of eye movements under relatively normal reading conditions that preserve natural reading behaviors. This approach has been shown to yield valuable insights into the cognitive mechanisms underlying sentence comprehension and provide an ecologically valid method for investigating how emotion influences language processing (Scott et al., Reference Scott, O’Donnell and Sereno2012). Accordingly, the present research used eye-tracking methodology to investigate the effects of EmoPro on the processing of words during natural reading, aiming to generate more realistic and nuanced evidence of a processing advantage for high EmoPro words.

In addition to this concern, the embodied theory (Vigliocco et al., Reference Vigliocco, Meteyard, Andrews and Kousta2009) suggests that the semantic activation of emotional words depends not only on affective experiences but also on the linguistic experiences associated with words (Villani et al., Reference Villani, Lugli, Liuzza, Nicoletti and Borghi2021). Thus, linguistic information, such as word frequency, age of acquisition, and semantic diversity, will also influence the processing of emotional semantics (Sheikh & Titone, Reference Sheikh and Titone2013; Diveica et al., Reference Diveica, Muraki, Binney and Pexman2024). These linguistic factors serve as a bridge between lexical form and semantic representation in the reader’s mental lexicon, supporting both word recognition and semantic retrieval (Chapman & Martin, Reference Chapman and Martin2022). Among them, word frequency, defined as the rate at which a word appears in a given corpus, plays a critical role in lexical access during natural reading (Mei et al., Reference Mei, Chen, Xia, Yang and Liu2025). Moreover, studies show that word frequency can limit the influence of other experiential information associated with words. For example, the effect of action-related experiential information, where word meanings activate the sensorimotor system (e.g., “grasp” activates motor-related brain regions), is observed only with words having a low frequency rather than words with high frequency (Sheikh & Titone, Reference Sheikh and Titone2013). Word frequency also appears to moderate access to emotional semantics, permitting the extraction of emotional content only for lower frequency (Kuperman et al., Reference Kuperman, Estes, Brysbaert and Warriner2014). This suggests that emotional meaning is shaped not just by experience but also by linguistic factors like word frequency. To fully understand how EmoPro influences language processing, it will be important to consider both types of information, the reader’s emotional experience and their linguistic knowledge. Accordingly, building on previous research examining the role of EmoPro in natural reading, this study investigates how word frequency affects the ability of high EmoPro words to facilitate emotional meaning retrieval during sentence processing.

In summary, the present uses eye-tracking methods to overcome limitations of previous research, which predominantly relied on isolated word presentations. By embedding emotional words with varying degrees of typicality in representing specific emotions within sentence contexts, we examined how such differences affected natural reading processes. In addition, the study manipulated word frequency to explore its role in moderating emotional information retrieval at different levels of EmoPro during sentence reading. Given that valence—a critical emotional attributes—significantly influences lexical processing, and to avoid potential confounding effects from mixed-valence stimuli (Kazanas & Altarriba, Reference Kazanas and Altarriba2015), our study adopted an approach consistent with prior research (Knickerbocker et al., Reference Knickerbocker, Johnson and Altarriba2015; Tang & Ding, Reference Tang and Ding2024). Specifically, it examined these questions separately under positive and negative valence conditions to provide a clearer perspective on how semantic representation information varies across valence. Our study, therefore, consisted of two experiments: one focusing on positive valence (Experiment 1) and the other on negative valence (Experiment 2). Both adopted a 2 (EmoPro: high vs. low) × 2 (Word Frequency: high vs. low) within-subjects design and were conducted in Chinese. Importantly, the selection of materials and control of lexical attributes were conducted separately for each experiment. This ensured that the attributes of target words were independently tailored to positive and negative valences, rather than being combined into a single experiment with uniformly controlled words. The specific hypotheses for each experiment are detailed below.

Experiment 1

Experiment 1 examined how EmoPro and word frequency influenced the processing of specific target words in sentences during reading. Each participant’s eye movements were recorded as they read a series of trials comprising an isolated sentence presentation, which was followed by a simple comprehension question (unrelated to the emotion of the target word) on a proportion of trials.

Drawing from prior studies, we hypothesized that EmoPro would exert a significant influence in natural reading contexts, with high EmoPro words maintaining their advantage in emotional content retrieval. According to the model of motivated attention and affective states, emotional stimuli are highly relevant to survival and self-defense, making them central to attention allocation and facilitating more comprehensive cognitive processing (Lang et al., Reference Lang, Bradley and Cuthbert1990). High-prototypical emotional words, characterized by the simultaneous activation of emotional information embedded in their semantics, are hypothesized to generate faster attentional biases (Altarriba & Basnight-Brown, Reference Altarriba and Basnight-Brown2011) due to this concurrent activation (Haro et al., Reference Haro, Calvillo, Poch, Hinojosa and Ferré2023; Wu et al., Reference Wu, Wu and Gao2024). Once this emotional information captures attention, it is likely to maintain attention for an extended period before disengagement occurs (Siakaluk et al., Reference Siakaluk, Knol and Pexman2014). Consequently, we hypothesize that high EmoPro words will result in prolonged fixation times while reading.

Predictions regarding the influence of word frequency were based on the Semantic Network Model (Collins & Loftus, Reference Collins and Loftus1975). According to this model, word semantic representations are stored in a hierarchically organized network, with prototype concepts located at higher levels. The closer a word’s semantic distance is to a prototype concept, the faster its associated semantic memory is activated (Kumar et al., Reference Kumar, Steyvers and Balota2022). This activation is also influenced by the strength of connections, determined by word familiarity (Collins & Loftus, Reference Collins and Loftus1975). Word frequency reflects the familiarity of a word to readers, with higher frequency indicating greater familiarity (Rayner, Reference Rayner2009). Consequently, frequent word occurrences can strengthen connections to prototype concepts, making their activation more automatic (Collins & Loftus, Reference Collins and Loftus1975). Positive words are stored with higher density and shorter semantic distances in the network (Unkelbach et al., Reference Unkelbach, Fiedler, Bayer, Stegmüller and Danner2008), and their frequent occurrences further enhance the automaticity of semantic activation. High EmoPro words are inherently closer semantically to emotional prototype concepts than low EmoPro words (Ferré et al., Reference Ferré, Guasch, Stadthagen-González, Hinojosa, Fraga, Marín and Pérez-Sánchez2024). Frequent occurrences of positive EmoPro words offer a stronger advantage in activating emotional semantics. This facilitates the efficient extraction of emotional information from positive high EmoPro words, capturing and sustaining readers’ attention. Therefore, we hypothesized that the activation advantage of positive high EmoPro words would be more pronounced at high frequencies, leading to greater attention and longer fixation durations.

Participants

The sample comprised 40 undergraduates, 31 of whom were female, with ages ranging from 18 to 25 years (M _age = 20.77 years, SD = 2.01), based on sample sizes from previous EmoPro studies (Wu et al., Reference Wu, Wu and Gao2024). Participants were right-handed, native Chinese speakers with normal or corrected vision. They had no history of reading disorders. Statistical power was evaluated with the Mixedpower package (Kumle et al., Reference Kumle, Võ and Draschkow2021) in R (R Development Core Team, 2016), running 1000 simulations based on the gaze duration effects of word frequency as described by Scott et al. (Reference Scott, O’Donnell and Sereno2012) in natural reading. The analysis confirmed that the current sample size was sufficient to detect similar effects, achieving statistical power greater than 0.9 at α = 0.05.

Materials and apparatus

Following the EmoPro rating approach used by Haro et al. (Reference Haro, Calvillo, Poch, Hinojosa and Ferré2023) and Wu et al. (Reference Wu, Wu and Gao2024), this study selected 513 potential two-character emotional words from the Cai and Brysbaert (Reference Cai and Brysbaert2010) corpus. The selected words were divided into five sets, and participants rated each word on a 5-point scale to indicate the extent to which it referred to an emotion (1 = “This word does not refer to an emotion,” 5 = “This word clearly refers to an emotion”). For the rating task, 338 participants were recruited and randomly assigned to one of the five sets. Each set was rated by at least 50 participants, ensuring comprehensive coverage across all word sets. To assess the validity of our ratings, we conducted correlation analyses between our ratings and those in the EmoPro Chinese word databases developed by Wu (Reference Wu2023) and Zheng et al. (Reference Zheng, Zhang, Guo, Guasch and Ferré2023). The results showed a strong positive correlation between the ratings (Wu’s database: r (208) = 0.91, p < 0.001; Zheng et al.’s database: r (155) = 0.81, p < 0.001). We also calculated intraclass correlation coefficients (ICCs) to assess the inter-rater reliability of the ratings for all word sets, following the procedure outlined by Pérez-Sánchez et al. (Reference Pérez-Sánchez, Stadthagen-Gonzalez, Guasch, Hinojosa, Fraga, Marín and Ferré2021). The average ICC (2, k) across all versions was 0.97 (SD = 0.01, coefficient of variation = 0.6%, range = 0.97–0.98), demonstrating high inter-rater reliability. These results show that the ratings are reliable and can be used as the target word database for our study. This database is available at https://osf.io/jr7zm/.

Based on the number of materials used by Wu et al. (Reference Wu, Wu and Gao2024), 80 two-character positive target words were selected, controlling for various lexical properties: pleasantness (F(3, 76) = 1.65, p = 0.19), arousal (F(3, 76) = 1.52, p = 0.22), abstractness (F(3, 76) = 0.74, p = 0.53), number of strokes (F(3, 76) = 0.30, p = 0.83), word family size (F(3, 76) = 0.87, p = 0.46), and age of acquisition (F(3, 76) = 0.68, p = 0.57). The final set included 40 high-prototypical emotion words (Mean _{emotional prototypicality} > 3) and 40 low-prototypical emotion words (Mean _{emotional prototypicality} < 3). These 80 target words were further categorized by frequency: 40 high-frequency words (Mean _{per million frequency} > 15, Mean _{log frequency} > 2) and 40 low-frequency words (Mean _{per million frequency} < 10, Mean _{log frequency} < 2). Pleasantness, arousal, and abstractness were rated by 27 non-reading experiment participants on a 9-point scale (1 = unpleasant, calm, concrete; 9 = pleasant, excited, abstract). Stroke count and word frequency data were taken from the Cai and Brysbaert (Reference Cai and Brysbaert2010) corpus, while word family size (the number of words containing a given character) was obtained from the Chinese Lexical Database (CLD; Sun et al., Reference Sun, Hendrix, Ma and Baayen2018). Age of acquisition information was sourced from the corpus of Xu et al. (Reference Xu, Li and Guo2021). The mean values of these variables for the target words are presented in Table 1. High- and low-EmoPro words showed significant differences in emotional prototypicality (Mean _{high EmoPro} = 3.56 (SD = 0.55), Mean _{low EmoPro} = 2.38 (SD = 0.40), t(78) = 10.78, p < 0.001), but no significant differences in word frequency (per million frequency for Mean _{high EmoPro} = 12.52 (SD = 18.42), per million frequency for Mean _{low EmoPro} = 14.10 (SD = 19.04), t (78) = 0.38, p = 0.71; log frequency for Mean _{high EmoPro} = 2.10 (SD = 0.78), log frequency for Mean _{low EmoPro} = 2.26 (SD = 0.72), t(78) = 1.00, p = 0.32). In contrast, high- and low-frequency words exhibited significant differences in frequency (per million frequency: Mean _{high frequency} = 18.71 (SD = 21.34), Mean _{low frequency} = 7.91 (SD = 13.70), t(78) = 2.69, p = 0.009; log frequency: Mean _{high frequency} = 2.51 (SD = 0.55), Mean _{low frequency} = 1.85 (SD = 0.79), t(78) = 4.31, p < 0.001), but no significant differences in emotional prototypicality (EmoPro for Mean _{high frequency} = 2.97 (SD = 0.74), EmoPro for Mean _{low frequency} = 2.96 (SD = 0.78), t(78) = 0.09, p = 0.93).

Table 1. Means (Standard deviations) for positive target word properties

Note. HF = high frequency. LF = low frequency. EmoPro = emotional prototypicality. HEP = high emotional prototypicality. LEP = low emotional prototypicality. These abbreviations remain consistent in following tables.

Experimental sentences were created by embedding positive high- and low-EmoPro words with high and low frequency, resulting in 80 distinct sentence structures. We also used the Mixedpower package (Kumle et al., Reference Kumle, Võ and Draschkow2021) to run 1000 simulations to assess whether the current number of materials has sufficient statistical power to detect the gaze duration moderation effects of word frequency, as described by Scott et al. (Reference Scott, O’Donnell and Sereno2012) in natural reading. The simulation results indicated that the statistical power at α = 0.05 exceeded 0.95 for the current size of experimental sentences.

Target words appeared centrally within the sentences, avoiding the first or last three positions (see Table 2 for examples). Sentence length ranged from 21 to 33 characters, with an average of 25 characters (SD = 3.49). Sentence contexts were designed to be non-predictive to minimize top-down processing, validated by a cloze test with 15 participants. The test showed low predictability for target words (Mean _{predictability} = 0.01%, SD = 4%), with no significant differences across conditions (F < 2.50, p > 0.10). In addition, 40 participants, not involved in the cloze test or reading experiment, rated the sentences on a 1–5 scale for understandability and plausibility. The sentences were rated as highly understandable (Mean _{understandability} = 4.12, SD = 0.33) by 20 participants and highly plausible (Mean _plausibility = 4.08, SD = 0.26) by another 20 participants, with no significant differences across conditions (Fs < 2.75, ps > 0.10). The mean values of understandability, plausibility, and predictability for each condition are presented in Table 1.

Table 2. Examples of experimental sentences for positive target words

Note. Target words presented in bold and italics above, but were not bold and non-italicized during experimental sessions.

The experimental sentences were organized into four sets using a Latin square design. Each set contained 80 experimental sentences, 40 fillers, and 10 practice sentences, totaling 130 sentences. Each participant read one list, with each target word and sentence appearing only once. The sentences were presented in a random sequence during the experiment, with practice sentences appearing before the experimental and filler sentences.

Eye movements from the right eye were captured using the SR Eyelink 1000 Plus tracker, sampling data at 1000 Hz during binocular viewing. The sentences were displayed in Song 32-point font (43 × 43 pixels) on a 24-inch DELL display (1920 × 1080 pixels; 120 Hz refresh rate). Participants sat 70 cm away from the screen, with their heads supported by a chin rest to minimize movement. The visual angle of each character was approximately 0.97°, typical for reading.

Procedure

The experiment was approved by the Ethics Committee of Tianjin Normal University (EC2024032508) and conducted in accordance with the principles of the Declaration of Helsinki. Participants completed the experiment individually, reading the sentences silently at their own pace to ensure comprehension. A three-point horizontal calibration was performed at the beginning, aligning with the sentence line to achieve spatial accuracy of 0.35° or better for all participants. Each trial began with the presentation of a fixation cross (“+”) at the leftmost position of the sentence. The sentence appeared when participants fixated on this spot. After reading each sentence, participants pressed a response key. Approximately 30% of sentences were followed by yes/no questions, balanced between “yes” and “no” responses. Participants’ average accuracy was about 94% (SD _accuracy = 9%). The entire experiment lasted about 20 minutes for each participant, with a 3-minute break midway through. Participants received ¥20 as compensation upon completion of the experiment.

Result

Consistent with previous research (Scott et al., Reference Scott, O’Donnell and Sereno2012), we calculated eye movement metrics for the areas of interest (AOI) containing the target word. The key metrics included first fixation duration (FFD), single fixation duration (SFD), gaze duration (GD), and total fixation time (TT). FFD refers to the duration of the first fixation in an AOI, SFD refers to the duration of a single fixation in the AOI, GD refers to the total time spent within the AOI before the gaze moves elsewhere, and TT refers to the cumulative fixation duration within an AOI (Rayner, Reference Rayner2009). FFD, SFD, and GD reflect early processing differences during lexical access, with FFD and SFD serving as the earliest indicators of lexical processing (Rayner, Reference Rayner2009). In contrast, TT reflects the overall process of semantic integration, encompassing both lexical and post-lexical processing (Rayner, Reference Rayner2009).

Standard procedures (Knickerbocker et al., Reference Knickerbocker, Johnson and Altarriba2015) were followed, excluding fixations shorter than 80 ms and longer than 1000 ms. Trials were also excluded based on the following criteria: (1) three or fewer fixations for the sentence (0.2%); (2) loss of tracking due to coughing, head movements, or blinks within the AOI (0.1%); and (3) fixations exceeding 2.5 standard deviations from the mean within the AOI (2%). The remaining data were analyzed using a linear mixed-effects model, using the lme4 package in R (Kuznetsova et al., Reference Kuznetsova, Brockhoff and Christensen2017). EmoPro, word frequency, and their interaction were treated as fixed effects, with participants and sentences as crossed random effects. The analysis began with a maximal random model, including random slopes for all variables for both participants and sentences. If the model did not converge, the model structure was then simplified, starting with removing correlations for sentences, then excluding random slopes. The final random effects structure, established after successful model convergence, is detailed in the statistical effects table (see below). Log-transformed fixation-time effects are reported (Sheikh & Titone, Reference Sheikh and Titone2013). Mean fixation times for target words across all conditions are presented in Table 3, and statistical effects are summarized in Table 4.

Table 3. Means (Standard Deviations) for eye-movement measures for positive target words

Note. FFD = first fixation duration; SFD = single fixation duration; GD = gaze duration; TT = total fixation time. These abbreviations remain consistent in following tables.

Table 4. Results of linear mixed effects models for positive target words

Note. *p<0.05. ** p<0.01. *** p<0.001. Freq = word frequency. This abbreviation remains consistent in following tables.

The results revealed significant main effects of EmoPro on FFD (b = −0.04, SE = 0.01, t = −2.69, p = 0.007), SFD (b = −0.04, SE = 0.01, t = −2.92, p = 0.004), GD (b = −0.04, SE = 0.02, t = −2.51, p = 0.01), and TT (b = −0.07, SE = 0.02, t = −3.65, p<0.001). High EmoPro elicited longer fixations, indicating they captured more attention compared to low EmoPro. A significant interaction between word frequency and EmoPro was found for FFD (b = 0.06, SE = 0.03, t = 2.01, p = 0.04) and SFD (b = 0.06, SE = 0.03, t = 2.09, p = 0.04). This suggests that word frequency moderated the advantage of high EmoPro in extracting emotional information during the initial stages of processing. To further assess the interaction between word frequency and EmoPro, we conducted post hoc comparisons employing a Bonferroni correction using the emmeans function (Lenth et al., Reference Lenth, Buerkner, Herve, Love, Miguez, Riebl and Singmann2022). Post hoc comparisons revealed the differences between high and low EmoPro appeared only under high-frequency (FFD: Mean _{(high EmoPro)} = 225 (SE = 5.31), Mean _{(low EmoPro)} = 211 (SE = 5.32), contrast estimate = 0.06, SE = 0.02, t = 3.31, p<0.001; SFD: Mean _{(high EmoPro)} = 225 (SE = 5.35), Mean _{(low EmoPro)} = 210 (SE = 5.36), contrast estimate = 0.07, SE = 0.02, t = 3.52, p<0.001), as shown in Figure 1. No significant differences were found in low-frequency (FFD: Mean _{(high EmoPro)} = 228 (SE = 5.30), Mean _{(low EmoPro)} = 225 (SE = 5.31), contrast estimate = 0.01, SE = 0.02, t = 0.48, p = 0.63; SFD: Mean _{(high EmoPro)} = 227 (SE = 5.34), Mean _{(low EmoPro)} = 224 (SE = 5.33), contrast estimate = 0.01, SE = 0.02, t = 0.59, p = 0.55).

Figure 1. The interaction between EmoPro and word frequency in FFD (with standard errors bars) for positive target words.

Note. The interaction patterns for FFD and SFD were consistent, so only the interaction plot for FFD is presented.

Moreover, the main effect of word frequency was significant for both early and late measures (FFD: b = 0.04, SE = 0.01, t = 2.72, p = 0.007; SFD: b = 0.03, SE = 0.01, t = 2.37, p = 0.02; GD: b = 0.04, SE = 0.02, t = 2.33, p = 0.02; TT: b = 0.04, SE = 0.02, t = 2.14, p = 0.03). High-frequency words elicited shorter fixations, reflecting faster lexical access compared to low-frequency words.

Discussion

Experiment 1 investigated whether EmoPro affects the retrieval of emotional information during Chinese natural reading, with words embedded in sentences under positive valence. Furthermore, it examined how word frequency, as a key form of linguistic information, moderates the retrieval advantage of emotional information associated with high-prototypical emotion words in natural language contexts.

Findings from Experiment 1 indicate that under positive valence, the degree of emotional semantic representation significantly affected reading processes. Consistent with previous research (Haro et al., Reference Haro, Calvillo, Poch, Hinojosa and Ferré2023; Wu et al., Reference Wu, Wu and Gao2024), high EmoPro words facilitated more efficient retrieval of emotional information from the lexicon. This advantage was evident in both early (e.g., FFD, SFD, GD) and later (e.g., TT) stages of lexical processing: high EmoPro words enabled faster access to emotional semantics, drawing and maintaining readers’ attention. Sustained attentional engagement resulted in longer fixation durations for high-prototypical emotion words during reading. In addition, the interaction between EmoPro and word frequency in the earliest measures (FFD, SFD) suggested that under positive valence, word frequency affected the retrieval of affective experiential information only during the initial stages of lexical access. In the initial stages of lexical access, the emotional semantics advantage of high EmoPro words appeared exclusively at high frequency, as frequent word occurrence enhances the automatic activation of emotional semantics. However, once lexical meaning was fully accessed (GD) and integrated during the late processing stage (TT), word frequency no longer had an impact. This suggests that the influence of word frequency on the emotional semantics of positive words is relatively weak during lexical access (Scott et al., Reference Scott, O’Donnell and Sereno2012), affecting only the initial stages of word processing.

Experiment 2

Experiment 2 further examined the effect of EmoPro under negative valence during reading and how word frequency moderates this effect. The experimental design followed the same procedure as in Experiment 1.

Regarding the impact of EmoPro in natural reading under negative valence, we anticipated results similar to Experiment 1, showing an emotional semantics advantage for high EmoPro words. We expected negative high EmoPro words to access emotional semantics more quickly, capture readers’ attention, and lead to longer fixation duration. However, we expected a different moderating effect of word frequency under negative valence compared to Experiment 1. Negative words are characterized by lower storage density and greater semantic distance within semantic memory (Unkelbach et al., Reference Unkelbach, Fiedler, Bayer, Stegmüller and Danner2008). Although frequent occurrences of negative words can strengthen their connection to semantics, this effect is constrained by their semantic distance. High word frequency primarily strengthens semantic activation when semantic distance is short (Collins & Loftus, Reference Collins and Loftus1975). Conversely, negative stimuli possess a distinct advantage due to heightened alertness, which is vital for survival and significantly influences attention (Sakaki et al., Reference Sakaki, Gorlick and Mather2011). Their heightened survival value makes negative information more attention-grabbing and engaging. However, the attentional advantage of negative stimuli varies with their frequency of occurrence. When negative words are frequent, their salience diminishes, reducing their ability to capture attention and sustain engagement. In contrast, infrequent negative words, due to their lower familiarity, are more likely to elicit alertness and engagement (Scott et al., Reference Scott, O’Donnell and Sereno2012). Thus, the impact of salience on attention to negative words is more pronounced at low frequencies (Scott et al., Reference Scott, O’Donnell and Sereno2012). The emotional semantic advantage of negative high EmoPro words arises from their faster extraction of emotional information during lexical access. When these words are frequent, the salience of their emotional information may decrease, leading to a reduced semantic advantage. In contrast, when these words are infrequent, their emotional information remains salient, making them more engaging and capturing readers’ attention more effectively. Therefore, we hypothesized that the emotional semantics advantage of negative high EmoPro words would be more pronounced at low frequencies. Specifically, we expected low-frequency negative high EmoPro words to exhibit heightened alertness in emotional meanings, capturing readers’ attention and leading to longer fixation duration.