Differential Modulation of Retroactive Emotional Learning on Concrete and Abstract Word Processing

ABSTRACT

The present study employed a sensory preconditioning paradigm combined with behavioral measures and event‐related potential (ERP) techniques to investigate whether emotional information could retroactively modulate the processing of previously encoded neutral concrete and abstract words. Behavioral results indicated that word memory performance was retroactively modulated by subsequent emotional conditioning: when the objects paired with abstract words were associated with aversive (vs. neutral) sounds subsequently, recognition accuracy was significantly higher and reaction times faster; for concrete words, however, no significant differences in accuracy or reaction time emerged between the two sound conditions. Additionally, liking ratings for both word types decreased overall. ERP results elicited by the words revealed that, in early attentional stages, abstract words whose paired objects were associated with aversive sounds evoked larger P200 amplitudes than those in the neutral condition. During memory retrieval, abstract words indirectly linked to aversive sounds (compared to the neutral condition) elicited smaller N400 and larger late positive component (LPC) amplitudes, reflecting enhanced early attentional capture, increased familiarity, and optimized recollection. In contrast, concrete words whose paired objects were associated with aversive sounds (compared to the neutral condition) elicited a larger right‐lateralized early posterior negativity (EPN), suggesting rapid activation of early emotional processing, but showed no significant changes in mid‐to‐late memory‐related components. These findings indicate that retroactive emotional learning selectively enhances memory for abstract words, supporting the representational substitution hypothesis and behavioral tagging theory, whereas its influence on concrete words is primarily manifested in early emotional modulation, consistent with partial expectations of the multimodal induction hypothesis.

Impact Statement

How does emotion retroactively reshape past memories? This study reveals the key lies in the semantic grounding of the memory content. For abstract concepts, retroactive emotional learning acts as a compensatory sensory anchor, deeply strengthening their memory encoding. For concrete words, it merely triggers rapid early emotional arousal. These findings provide novel neural and behavioral evidence supporting the representational substitution hypothesis and behavioral tagging theory, advancing our understanding of how emotion dynamically interacts with language and memory systems.

1. Introduction

Human memory is a dynamic and adaptive system that can be shaped by emotional events. The emotionality of an event not only significantly influences memory for the event itself (Yonelinas and Ritchey 2015) but can also retroactively strengthen memories of previously encoded neutral information (Ballarini et al. 2009; Dunsmoor et al. 2022; Ramírez Butavand et al. 2020). This phenomenon, known as retroactive memory enhancement (RME), is crucial for individuals to extract meaning from seemingly ordinary experiences and guide future behavior. However, existing research predominantly employs simple perceptual stimuli (e.g., shapes, faces), which limits explanations of how emotionally organized knowledge and experiences in real life are retroactively modulated by emotion. In particular, the mechanisms through which emotion differentially affects memory across semantic categories—such as concrete versus abstract words—remain unclear. Therefore, this study aims to investigate whether and how emotional stimuli differentially enhance retroactive memory for neutral concrete and abstract words, thereby uncovering the underlying mechanisms of the interplay among emotion, language, and memory.

Neurobiological models provide an important framework for understanding the mechanisms underlying RME effects. The “synaptic tag‐and‐capture (STC)” theory posits, at the cellular and molecular levels, that weak stimulation sets a transient, glutamate‐mediated “tag” at the synapse, and that a strong stimulus (such as emotional arousal) occurring within several hours thereafter releases a large amount of plasticity‐related proteins (PRPs). The transient “tag” “captures” the released proteins, thereby transforming the weak memory into a stable long‐term memory (Frey and Morris 1997). The “behavioral tagging” theory is the behavioral counterpart and extension of STC, referring to the analogous phenomenon observed in animal or human experiments: presenting a strong, emotionally arousing or novel experience within approximately several hours after a weak learning event can significantly enhance memory for that weak event (Dunsmoor et al. 2022). Critically, this theoretical framework defines “retroactive influence” as the strengthening or modification of newly encoded, not‐yet‐stabilized memory traces by a subsequently occurring event, rather than requiring the initial memory to have first completed consolidation into a stable long‐term form. Memory consolidation is a dynamic, time‐dependent process, and within minutes to hours after initial encoding, memory traces remain in a labile state, highly susceptible to subsequent neuromodulatory events. Indeed, the “synaptic tags” set by a weak learning event are transient but typically last several hours, such that a strong event occurring within this temporal window can retroactively “capture” and strengthen the previously set weak tags. However, RME can also lead to excessive memory generalization, manifesting as intrusive and generalized memories in conditions such as post‐traumatic stress disorder (PTSD), affecting an individual's emotional and cognitive functioning. Thus, RME reflects both the flexibility and adaptability of memory systems and may also become a potential source of memory distortion in psychopathological states.

Further research has elucidated the modulatory effects of emotion on previously neutral memories and their underlying mechanisms. Zhu et al. (2022) employed a “sensory preconditioning” experimental paradigm to systematically investigate how emotion strengthens previously formed neutral memories. In that study, participants first learned associations between two neutral stimuli, such as pairings of object pictures (Preconditioned Stimulus, PS) and face pictures (Conditioned Stimulus, CS). Subsequently, CS+ was paired with a screaming aversive sound (Unconditioned Stimulus, US), while CS− was paired with a neutral sound. The results showed that memory for the PS‐CS associations formed during the initial learning phase was significantly enhanced after the CS was paired with the US. Furthermore, Wong et al. (2019) observed in rat experiments that animals could integrate the previously acquired PS‐CS association with the subsequent CS‐US emotional association into a unified memory representation. Consequently, during the test phase, even though the PS was never directly paired with the US, the rats exhibited fear responses to the PS. In summary, emotional information not only retroactively enhances the strength of neutral memories but also generalizes to associated neutral memories, thereby imbuing them with emotional significance.

Despite advancements in the research on retroactive memory enhancement (RME) and emotional generalization, a notable limitation persists: the majority of evidence has been derived from studies employing simple, nonlinguistic perceptual stimuli. For instance, Dunsmoor et al. (2015) used object pictures (animals and tools) to demonstrate that fear conditioning selectively and retroactively strengthens memory for previously encoded neutral images belonging to the same semantic category as the conditioned stimuli, an effect contingent upon a post‐encoding consolidation window. Extending this line of inquiry, Zhu et al. (2022) utilized a sensory preconditioning paradigm with neutral faces and object pictures, revealing that emotional learning retroactively promotes memory integration for prior neutral associations through rapid reactivation of overlapping neural traces in the hippocampus and stimulus‐relevant neocortex. Similarly, Cooper et al. (2024) employed animal and tool pictures as category exemplars to show that threat conditioning indirectly modifies neural representations of an entire semantic category pre‐associated with the conditioned cue, with representational changes persisting beyond 24 h in the hippocampus and perirhinal cortex. Most recently, Lambert et al. (2025) examined the boundary conditions of behavioral tagging, finding that retroactive memory enhancements for category pictures encoded prior to fear conditioning can endure for up to 1 month, provided the initial encoding was sufficiently weak and occurred without threat expectancy. Collectively, these findings from nonlinguistic picture‐based paradigms converge to indicate that emotional events can selectively bolster memory for previously neutral perceptual stimuli through conceptual or categorical overlap, supported by dynamic interactions among the hippocampus, amygdala, prefrontal cortex, and category‐selective cortical regions (for a review, see Dunsmoor et al. 2022). However, while such materials are invaluable for isolating sensory modalities and elucidating core mechanisms, their conceptual structure remains relatively simple and their semantic depth shallow, thereby limiting the extent to which they can illuminate how experiences that are organized and transmitted through language in daily life become retroactively modulated by emotion.

Notably, the “Affective Neurolinguistics” framework advanced by Hinojosa et al. (2020, 2023) underscores that the interplay between emotion and language permeates multiple levels—including early semantic access (Liu et al. 2023), the integration of syntactic structures (Poch et al. 2022), and comprehension within complex pragmatic contexts (Hernández‐Gutiérrez et al. 2022)—constituting a highly dynamic cognitive system. In everyday experience, ostensibly mundane verbal information, such as a neutral description or isolated words, may only acquire personal significance when subsequently linked to an emotional event. Therefore, extending RME research to encompass linguistic symbols—the primary vehicles of human knowledge, particularly concrete and abstract words that diverge fundamentally in their embodied grounding—is essential for unraveling the interactive mechanisms among emotion, language, and memory.

It is well known that the human knowledge system is largely organized and transmitted through language. Words, as the basic units of language, can be divided into concrete and abstract words, reflecting different semantic conceptual categories. According to the Dual Coding Theory (Paivio 1991) and recent neuroscience evidence (Bi 2021), concrete words can be encoded and retrieved via dual channels—the language system and the sensorimotor system—activating both language‐related brain regions (e.g., left temporal lobe) and areas associated with sensory experience and motor execution (e.g., sensorimotor cortex). In contrast, abstract words, lacking direct sensory referents, rely more on symbolic representation within the language system itself (e.g., dorsal anterior temporal lobe) for their semantics (Vigliocco et al. 2014; Bi 2021). These theories are supported by empirical research, further confirming the dual coding advantage of concrete words and the reliance of abstract words on a single language system. For instance, concrete words are processed faster and more accurately than abstract words (Kounios and Holcomb 1994; Holcomb et al. 1999; West and Holcomb 2000). This fundamental representational difference raises a key question: Do emotional stimuli differentially affect the retroactive enhancement of memory for concrete versus abstract words?

Currently, there are two competing theoretical hypotheses: the Multimodal Induction Hypothesis and the Representational Substitution Hypothesis. The Multimodal Induction Hypothesis posits that concrete words can directly link to rich sensorimotor experiences, which serve as “channels” for emotion induction, facilitating the rapid activation of emotional content (Niedenthal 2007; Hess and Blairy 2001). Consequently, the impact of emotion on concrete words is more pronounced. Conversely, the Representational Substitution Hypothesis suggests that, due to their lack of direct sensory grounding, the representation of abstract words relies more heavily on the linguistic system and its associated emotional and social information (Kousta et al. 2011; Vigliocco et al. 2014). According to this view, emotion to some extent substitutes for the missing sensory foundation, compensating for the semantic ambiguity and perceptual distance of abstract words, potentially rendering emotional stimuli more influential on abstract words.

Although the aforementioned two hypotheses have been partially empirically tested in the interaction between emotion and word processing (Yao et al. 2018, 2024), it remains unclear whether they apply to the “retroactive enhancement” of word memory by emotional stimuli and the emotional learning process. To investigate this theoretical divergence, the present study builds upon the sensory preconditioning paradigm used by Zhu et al. (2022), introducing concrete and abstract words as memory materials to systematically examine the retroactive enhancement effect of emotional stimuli (aversive screams vs. neutral sounds) on the memory of neutral concrete and abstract words. Based on the competing hypotheses, we made the following predictions for behavioral outcomes: If the Representational Substitution Hypothesis holds, emotional learning should selectively enhance memory performance (i.e., higher accuracy and faster reaction times) for abstract words due to emotion compensating for their lack of sensorimotor grounding. Conversely, if the Multimodal Induction Hypothesis prevails, concrete words, by virtue of their direct sensory channels, would show a greater retroactive memory benefit from emotional learning. Furthermore, the study will employ a word liking rating task to assess whether emotional pairing indirectly alters the affective valence of the words, thereby providing behavioral evidence for the emotion‐ascription mechanism. Experiment 2 further incorporates ERP techniques, focusing on neural activities across different time windows, to reveal the dynamic process of emotion modulating the memory of concrete and abstract words.

In ERP research, the early stages of word processing show particular sensitivity to emotional information in the P200 and early posterior negativity (EPN) components. An increased P200 amplitude typically reflects the rapid capture of attention by emotional stimuli (Kissler et al. 2009; Hinojosa et al. 2020), while the EPN is closely associated with the automatic processing of emotional content and emotional arousal (Citron 2012; Schacht and Sommer 2009). During the mid‐to‐late stages of memory retrieval, the N400 and the late positive component (LPC) are linked to the processes of familiarity and recollection, respectively (Curran 2000; Rugg and Curran 2007). Research indicates that emotional information can modulate these components: for instance, emotional information can enhance the N400 effect, suggesting that emotion may promote item memory by boosting familiarity (Xie et al. 2024), whereas the LPC is generally enhanced during recollection processes associated with emotional information (Ventura‐Bort et al. 2017; Xie et al. 2024).

Accordingly, the present study will test two competing hypotheses regarding the differential effects of retroactive emotional enhancement on concrete versus abstract words. If the Representational Substitution Hypothesis holds, emotion should selectively enhance memory for abstract words by compensating for their lack of sensorimotor grounding; this would manifest as a larger N400 and LPC effect for abstract words relative to concrete words following emotional learning, possibly accompanied by early emotion‐related modulations of the P200 or EPN that reflect rapid semantic compensation. Conversely, if the Multimodal Induction Hypothesis prevails, concrete words—by virtue of their direct sensorimotor channels—should benefit more from retroactive emotional enhancement, which would be reflected in more pronounced memory effects for concrete words during both familiarity‐based retrieval (N400) and recollection (LPC), along with early P200 or EPN enhancements indicating rapid integration of emotional information via sensorimotor pathways. By integrating behavioral responses with ERP techniques and examining the dynamic time courses of these components (P200, EPN, N400, and LPC), this study aims to elucidate the temporal dynamics and neural mechanisms underlying the retroactive memory enhancement of emotional stimuli on neutral concrete and abstract words.

2. Experiment 1

2.1. Participants

The sample size was calculated using G*Power 3.1 software. With a medium effect size (f = 0.25), a Type I error probability (α = 0.05), and a statistical power of 0.80, the analysis indicated a minimum required sample size of 27 participants. The experiment actually recruited 32 university students (9 male), with a mean age of 19.78 years (SD = 1.52). All participants were native Chinese speakers, had normal or corrected‐to‐normal vision and normal hearing, were right‐handed, and reported no history of neurological or psychiatric disorders. This study was approved by the Research Ethics Committee of the School of Psychology at the authors' institution (Approval No. 2025‐03‐06). All participants provided written informed consent prior to participation. Each participant received ¥40 (approximately $5.5 USD) as compensation for their time.

2.2. Materials

Word materials were selected from the pool of neutral concrete and abstract nouns rated by Yan et al. (2018), comprising 112 words in total. For the formal experiment, 96 words were used (48 concrete and 48 abstract), with the remaining 16 words allocated to the practice phase. Word frequency information was obtained from the Chinese Lexical Database (CLD; Sun et al. 2018), a large‐scale lexical resource for simplified Mandarin Chinese providing frequency, stroke count, and other lexical properties.

Eighteen participants (4 male, M = 25.39, SD = 3.87) who did not participate in the formal experiment rated the concreteness and liking of the word materials using a 9‐point Likert scale (1 = “very disliked, uncomfortable” or “very abstract,” 9 = “very liked, very comfortable” or “very concrete”). The rating results showed a significant main effect of word type on concreteness, F(1, 94) = 1504.62, p < 0.001. In contrast, no significant differences were found in liking ratings, word frequency, or stroke count (see Table 1). Furthermore, liking ratings did not differ significantly between the two experimental conditions (aversive vs. neutral), F(1, 94) = 0.65, p = 0.42.

TABLE 1.

Descriptive statistics of the lexical dimensions for concrete and abstract words (M ± SD).

Dimension	Concrete words (n = 48)	Abstract words (n = 48)	F value	p
Concreteness	7.96 ± 0.54	2.92 ± 0.72	F(1,94) = 1504.62	< 0.001
Liking rating	5.26 ± 0.63	5.4 ± 0.7	F(1,94) = 1.07	0.3
Word frequency	34.73 ± 115.2	82.87 ± 164.6	F(1,94) = 2.76	0.1
Stroke count	15.67 ± 4.48	15.02 ± 3.56	F(1,94) = 0.61	0.44
Imageability	7.88 ± 0.64	3.63 ± 0.77	F(1,94) = 870	< 0.001
Valence	5.24 ± 0.67	5.27 ± 0.77	F(1,94) = 0.05	0.82
Arousal	3.98 ± 0.58	4.13 ± 0.62	F(1,94) = 1.64	0.2
Semantic diversity	0.2 ± 0.03	0.19 ± 0.02	F(1,94) = 6.98	0.01

Furthermore, we conducted additional analyses on valence, arousal, imageability, and semantic diversity. Nineteen university students (6 males; M = 25.11, SD = 3.11) who did not participate in the main experiment rated valence (1 = very unpleasant, 9 = very pleasant), arousal (1 = not at all arousing, 9 = very arousing), and imageability (1 = no concrete image, 9 = very vivid image) for all 96 words using a 9‐point Likert scale.

Moreover, semantic diversity, which indexes the variability of a word's contextual usage across semantic space (Hoffman et al. 2013), was computed for each item. Using pretrained Chinese word vectors (sgns_merge_bigram‐char) implemented in the R package PsychWordVec (Bao 2023), each word was represented by a 300‐dimensional embedding. The semantic diversity index was derived by calculating the standard deviation of cosine similarity values between the target word and all other words in the corpus. Higher values reflect greater contextual variability and less rigid semantic boundaries.

ANOVAs revealed no significant differences between concrete and abstract words in terms of valence (F(1, 94) = 0.05, p = 0.82) or arousal (F(1, 94) = 1.64, p = 0.2). As expected, imageability ratings were significantly higher for concrete words than for abstract words (F(1, 94) = 870, p < 0.001). Critically, abstract words exhibited significantly greater semantic diversity than concrete words (F(1, 94) = 6.98, p = 0.01).

Picture materials consisted of 32 novel images selected from The Novel Object and Unusual Name Database (Horst and Hout 2016). During the formal experimental phase, 24 of these images were used as memory materials, while the remaining 8 images were allocated to the practice phase. To ensure that the images paired with aversive versus neutral sounds did not differ systematically in their baseline emotional properties, an independent sample of 18 university students (6 males; M = 25.28, SD = 3.1) who did not participate in the main experiment evaluated the two sets of images (those subsequently paired with aversive screams vs. those paired with neutral sounds) on valence, arousal, and familiarity using a 9‐point Likert scale. The results revealed no significant differences between the two image sets on any of the three dimensions: valence, F(1, 22) = 2.46, p = 0.13; arousal, F(1, 22) = 0.05, p = 0.83; and familiarity, F(1, 22) = 1.7, p = 0.21.

Four female aversive screams and four female neutral sounds were selected from an online audio resource library (https://www.audiodown.com). Twelve participants (4 male, M = 20.67, SD = 2.23) who did not participate in the formal experiment rated the valence and arousal of the audio materials using a Likert 9‐point scale (1 = “very unpleasant” or “not at all exciting”, 9 = “very pleasant” or “very exciting”). The rating results revealed a significant main effect of valence, F(1,11) = 192.46, p < 0.001, with aversive screams (M = 1.17, SD = 0.08) rated as significantly more unpleasant than neutral sounds (M = 4.54, SD = 0.25). A significant main effect of arousal was also observed, F(1,11) = 261.3, p < 0.001, indicating that aversive screams (M = 8.56, SD = 0.15) elicited higher arousal than neutral sounds (M = 4, SD = 0.36).

Furthermore, following the approach of Zhu et al. (2022), we analyzed the acoustic properties of the sound clips using Praat software. As shown in Table 2, while the frequency characteristics differed between the aversive and neutral sounds (given the inherent high pitch of screams), the loudness (decibel levels) was relatively comparable across all sounds, mitigating potential confounding effects from physical intensity differences. Given that the effects of emotional learning are primarily driven by arousal and valence (Kensinger 2004), explicit familiarity ratings for the auditory stimuli were not collected.

TABLE 2.

Acoustic characteristics of the eight voice clips.

	Duration (s)	Frequency (Hz)				Power (db)
		Median	Mean ± SD	Min	Max	Median	Mean ± SD	Min	Max
Scream1	2	499.22	422.61 ± 149.91	100.02	548.45	82.05	81.93 ± 2.1	71.84	85.35
Scream2	2	473.45	498.48 ± 51.74	168.13	599.05	84.8	82.34 ± 8.06	47.3	88.23
Scream3	2	515.07	498.89 ± 51.92	323.79	541.64	74.94	74.65 ± 2.59	54.08	78.73
Scream4	2	509.2	482.65 ± 74.71	249.46	579.73	77.73	72.55 ± 10.52	37.97	82.83
“Ah”	2	207.79	207.86 ± 4.15	185.59	220.17	76.65	74.92 ± 6.36	56.84	81.38
“Eh”	2	211.3	211.36 ± 2.95	205.66	226.7	77.82	75.41 ± 6.6	56.81	82.48
“Eih”	2	215.09	216.02 ± 3.63	209.82	229.25	75.43	74.56 ± 4.01	60.21	79.15
“Oh”	2	209.28	208.8 ± 5.21	190.67	224.2	78.6	77.04 ± 4.61	58.74	80.64

2.3. Design and Procedure

The experiment employed a 2 (Word Type: concrete, abstract) × 2 (Stimulus Condition: aversive, neutral) within‐subjects design. The dependent variables were the accuracy and reaction time (RT) for identifying old words in the word recognition task, as well as the word liking ratings. Stimulus presentation and RT recording were implemented using PsychoPy 2023.2 software. The experiment consisted of a learning phase and a test phase.

The sensory preconditioning paradigm was used, comprising two subphases: preconditioning learning and emotional learning.

In the preconditioning learning phase, participants were required to learn the associative relationships between neutral words and novel object pictures. To control the duration of the experiment and prevent participant fatigue, a many‐to‐one pairing design was adopted: four words of the same type (all concrete or all abstract) were paired with the same novel object picture. A total of 96 trials (48 concrete and 48 abstract words) were presented in two blocks, with each trial lasting approximately 9 s. Participants could take a short break between blocks if needed. The total duration of this phase was approximately 14 min. To encourage active encoding and associative formation between the stimuli, participants were instructed to vividly imagine an interactive scenario between each word and its corresponding object in a certain context (e.g., the object being used by, accompanying, or placed in the same scene as the concept denoted by the word). All trials were presented in a pseudo‐random order, ensuring that the same picture did not appear consecutively more than twice.

In the emotional learning phase, the previously used novel object pictures were represented as overlapping cues, each randomly paired with either an aversive (terror) or a neutral sound. There were 24 trials in total (one for each object picture), with each trial lasting 12 s. The total duration of this phase was approximately 5 min. Participants were instructed to imagine an interaction between the object and the sound (e.g., relating the sound to the scenario in which the object was situated). Each object picture appeared only once during this phase to reduce potential habituation due to repeated exposure to the emotional stimuli and to maintain their effectiveness (Figure 1A). After the learning phase, participants rested for approximately 10 min before proceeding to the test phase.

FIGURE 1.

Example of a trial during the learning and test phases.

During the test phase, participants first completed a word recognition task, followed by a word liking rating task to avoid potential order effects on memory and affective judgments.

First, in the word recognition task, participants judged whether presented words were “old” (from the learning phase) or “new.” There were 96 old words and an additional 48 new words serving as lures, resulting in 144 trials. Each trial lasted up to 5 s, and the total duration of this task was within 12 min. Participants pressed the “M” key if the word was from the preconditioning learning phase and the “Z” key if it was new, with the key assignment counterbalanced across participants. The presentation order was pseudo‐random, ensuring that no more than three consecutive trials required the same response. The word disappeared immediately after a response or automatically advanced to the next stimulus if no response was made within 3 s.

After completing the word recognition task, participants then performed a “liking” rating task for the words (i.e., those memorized during the preconditioning phase), which were presented randomly on the screen. A total of 96 trials were presented, with each trial lasting approximately 6 s on average (until a response was given or a 10‐s timeout). The total duration of this task was approximately 9 min. They were asked to rate how much they liked each word on a scale from 1 (“very disliked, uncomfortable”), through 5 (“neutral, no particular feeling”), to 9 (“very liked, very comfortable”). The word disappeared immediately after a rating was given or automatically advanced if no response was made within 10 s (see Figure 1B). The entire experiment lasted approximately 50 min on average.

2.4. Statistical Analysis

Behavioral data analysis was performed using JASP 0.19.1 (JASP Team 2022). Prior to statistical analysis, data cleaning was conducted as follows: trials with incorrect responses in the word recognition task and trials with reaction times (RTs) exceeding 3 s were excluded, accounting for 17.81% of the total trials. Subsequently, for the remaining trials, RTs exceeding ±3 standard deviations from the mean RT of each condition per participant were removed (1.89%).

We conducted traditional repeated‐measures analyses of variance (ANOVAs) for accuracy and RT data, with Word Type (concrete, abstract) and Stimulus Condition (aversive, neutral) as within‐subject factors. Where appropriate, simple effects analyses were performed to decompose significant interactions, using Bonferroni correction for multiple comparisons.

In addition to frequentist ANOVA, we performed Bayesian repeated‐measures ANOVAs to quantify the strength of evidence for the alternative hypothesis (H₁) relative to the null hypothesis (H₀). An advantage of the Bayesian approach is that it provides probabilistic evidence for the presence or absence of effects, rather than relying solely on binary null hypothesis testing. We report Bayes factors as BF₁₀ when supporting H₁, and BF₀₁ when supporting H₀. According to conventional interpretation guidelines (Wagenmakers et al. 2018), BF₁₀ values greater than 1 indicate evidence for H₁, with values > 3 considered moderate evidence, > 10 strong evidence, and > 30 very strong evidence. Conversely, BF₀₁ values greater than 1 indicate evidence in favor of H₀.

For the word liking ratings, a similar 2 × 2 repeated‐measures ANOVA was conducted, followed by Bayesian ANOVA to assess the robustness of the observed effects.

2.5. Results

A repeated‐measures analysis of variance (ANOVA) was performed on accuracy (see Figure 2A). The results revealed a significant main effect of Word Type, F(1, 31) = 7.6, p = 0.01, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.2, BFincl = 3.68, indicating that accuracy was higher for concrete words (M = 0.84, SD = 0.02) than for abstract words (M = 0.79, SD = 0.02). There was no main effect of Stimulus Condition, F(1, 31) = 2.54, p = 0.12, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.08, BFincl = 0.41. The interaction between Word Type and Stimulus Condition was marginally significant, F(1, 31) = 3.88, p = 0.06, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.11, BFincl = 4.82. Simple effects analyses showed that for concrete words, accuracy did not differ significantly between the aversive and neutral sound conditions, F(1, 31) = 0.35, p = 0.56, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.01. The Bayesian analysis provided moderate evidence for the null hypothesis, BF ₀₁ = 4.5, indicating that the absence of an emotional modulation effect on concrete word accuracy was substantiated by the data. For abstract words, however, accuracy was significantly higher in the aversive sound condition than in the neutral condition, F(1, 31) = 6.62, p = 0.02, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.18, BF ₁₀ = 3.12. Further comparisons indicated that in the neutral sound condition, accuracy was significantly higher for concrete words than for abstract words, F(1, 31) = 10.7, p < 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.25, BF ₁₀ = 11.76. In contrast, under the aversive sound condition, there was no significant accuracy difference between the word types, F(1, 31) = 0.46, p = 0.5, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.02, and the Bayesian analysis again provided moderate evidence for the null hypothesis (BF ₀₁ = 4.28), confirming that the concreteness advantage was eliminated following aversive learning. Thus, the null findings concerning the lack of emotional modulation for concrete words were supported by moderate Bayesian evidence, whereas other comparisons (e.g., the main effect of Stimulus Condition) remained inconclusive due to only weak or ambiguous evidence.

FIGURE 2.

Comparisons of accuracy and reaction time across conditions in the word recognition task. (A) Mean accuracy (%) for concrete and abstract words under aversive and neutral sound conditions. (B) Mean reaction times (s) for concrete and abstract words under aversive and neutral sound conditions. Error bars represent the standard error of the mean (SEM). *p < 0.05, **p < 0.01, ***p < 0.001.

Regarding reaction time (RT; see Figure 2B), there was a significant main effect of Word Type, F(1, 31) = 13.94, p = 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.31, BFincl = 14.84, with RTs being faster for concrete words (M = 0.86 s, SD = 0.03) than for abstract words (M = 0.90 s, SD = 0.03). There was also a significant main effect of Stimulus Condition, F(1, 31) = 7.59, p = 0.01, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.2, BFincl = 2.5, where RTs were faster under the aversive sound condition (M = 0.86 s, SD = 0.03) than under the neutral condition (M = 0.90 s, SD = 0.03). The interaction between Word Type and Stimulus Condition was significant, F(1, 31) = 6.91, p = 0.01, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.18, BFincl = 12.7. Simple effects tests revealed that for concrete words, RTs did not differ significantly between the aversive and neutral sound conditions, F(1, 31) = 0.00, p = 0.97, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.00. Critically, the Bayesian analysis provided moderate evidence for the null hypothesis, BF ₀₁ = 5.29, indicating that the absence of retroactive emotional modulation on concrete word RTs was substantiated by the data. In contrast, for abstract words, RTs were significantly faster in the aversive sound condition than in the neutral condition, F(1, 31) = 16.25, p < 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.34, BF ₁₀ = 86.3. Furthermore, under the neutral sound condition, RTs were significantly faster for concrete words than for abstract words, F(1, 31) = 16.86, p < 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.35, BF ₁₀ = 104.02. Under the aversive sound condition, however, there was no significant RT difference between the two‐word types, F(1, 31) = 0.49, p = 0.49, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.02, and the Bayesian analysis again provided moderate evidence for the null hypothesis (BF ₀₁ = 4.22), confirming that the concreteness advantage in RTs was eliminated following aversive learning.

Analysis of word liking ratings using repeated‐measures ANOVA (see Figure 3) revealed a marginally significant main effect of Word Type, F(1, 31) = 3.84, p = 0.06, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.11, BFincl = 1.53. There was a significant main effect of Stimulus Condition, F(1, 31) = 12.48, p < 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.29, BFincl = 13.44. Liking ratings were significantly lower for words in the aversive sound condition (M = 5.67, SD = 0.1) compared to those in the neutral sound condition (M = 5.82, SD = 0.09). There was no significant interaction between Word Type and Stimulus Condition, F(1, 31) = 0.89, p = 0.35, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.03, and the Bayesian analysis yielded anecdotal evidence in favor of the null hypothesis for the interaction (BFincl = 0.4, corresponding to BF ₀₁ = 2.5). This pattern indicates that the retroactive transfer of negative valence occurred to a similar extent for both concrete and abstract words, with no differential modulation by word type.

FIGURE 3.

Comparison of liking ratings across conditions. Mean liking ratings for concrete and abstract words under aversive and neutral sound conditions. Error bars represent the standard error of the mean (SEM). *p < 0.05, **p < 0.01.

Results from Experiment 1 demonstrated that retroactive emotional learning significantly facilitated memory for abstract words, but did not show a significant effect on memory for concrete words. Specifically, when the objects paired with abstract words were associated with aversive sounds, recognition accuracy was significantly higher and reaction times (RTs) were faster compared to the condition where they were associated with neutral sounds. For concrete words, however, there were no significant differences in either accuracy or RTs between the two sound conditions. Moreover, when the paired objects were linked to neutral sounds, concrete words were recognized with higher accuracy and faster RTs than abstract words; yet this advantage disappeared when the objects were associated with aversive sounds.

Results from the word liking ratings indicated that, for both concrete and abstract words, ratings became significantly more negative when their paired objects were associated with aversive sounds. This suggests that retroactive emotional learning not only altered the emotional valence of the words—reflecting a generalization process of emotional information—but more importantly, this learning experience also significantly modulated the processing efficacy of the words. To further elucidate the underlying neural mechanisms, Experiment 2 combined ERP techniques to conduct an in‐depth analysis of electrophysiological components during word processing under different conditions.

3. Experiment 2

3.1. Participants

The sample size was calculated using G*Power 3.1 software. Based on a medium effect size (f = 0.25), a Type I error rate (α = 0.05), and a statistical power of 0.8, the minimum required sample size was estimated to be 27 participants. Thirty‐two university students were initially recruited. Behavioral data were analyzed from all 32 participants (11 males, M_age = 19.38 years, SD = 1.94). For the EEG analysis, two participants were excluded due to excessive artifacts, leaving a final sample of 30 participants (11 males, M_age = 19.34 years, SD = 2.03). All participants were native Chinese speakers, had normal or corrected‐to‐normal vision and normal hearing, were right‐handed, and reported no history of neurological or psychiatric disorders. This study was approved by the Research Ethics Committee of the School of Psychology at the authors' institution (Approval No. 2025‐03‐06). All participants provided written informed consent prior to participation. Each participant received ¥70 (approximately $9.5 USD) as compensation for their time.

3.2. Materials

The materials were identical to those used in Experiment 1.

3.3. Design and Procedure

The learning phase was identical to that of Experiment 1. After the learning phase, participants rested for approximately 10 min.

During the test phase, EEG signals were recorded while participants performed the word recognition task. Each trial began with a 2‐s fixation cross (“+”), followed by the presentation of a word for 1 s. After the word disappeared, a cue (“*”) appeared, and participants then made their keypress response. This delayed response paradigm was employed to minimize artifacts from motor preparation and execution. To ensure a sufficient number of trials per condition, each old word was presented twice, resulting in a total of 288 trials (192 old words + 96 new lures). These trials were divided into two blocks, with a 5–10‐min break in between to prevent fatigue. Each trial lasted up to 5 s, and the task was completed within approximately 24 min. The trial sequence was pseudo‐randomized, ensuring that no more than three consecutive trials required the same response. Following the recognition task, participants completed the word liking rating task.

3.4. EEG Recording and Analysis

Electroencephalogram (EEG) was recorded using a 64‐channel Ag/AgCl electrode cap and the ERP recording system manufactured by Brain Products (Germany). The sampling rate was set to 1000 Hz with a bandpass filter from 0.05 Hz to 100 Hz. EEG data were preprocessed using EEGlab. Processing steps included: (1) applying a 0.1–30 Hz band‐pass filter, (2) segmenting the data into epochs from 200 ms before to 1000 ms after the onset of each target word, with a baseline correction from −200 ms to 0 ms relative to word onset, and (3) re‐referencing the data offline to the average of the bilateral mastoids. Independent Component Analysis (ICA) was further applied to remove artifacts arising from eye blinks, cardiac activity, skin potentials, and muscle movements. Epochs containing amplitudes exceeding ±100 μV were rejected; on average, 5.54% of trials were excluded.

Finally, for each participant, all valid EEG epochs were averaged separately for each experimental condition to obtain the ERP waveforms. Following the methodological approach of Liu et al. (2023) in studying the neural mechanisms of emotional word processing, and based on the typical scalp topographies of each ERP component reported in prior literature, representative electrodes were selected and regionally averaged to enhance the signal‐to‐noise ratio and focus on component‐specific neural activity. Specifically, analyses focused on four components: P200, EPN, N400, and LPC (for details on electrode sites and time windows, see Table 3). The mean amplitude of each component was subjected to a 2 (Word Type: concrete, abstract) × 2 (Stimulus Condition: aversive, neutral) repeated‐measures analysis of variance (ANOVA). All statistical analyses were performed using the JASP 0.19.1 software package.

TABLE 3.

The time window and electrode selection for the ERP components.

	Time window (ms)	Electrode
P200	180–220	C1, C2, Cz, FC1, FC2, FCz
EPN	250–350	LH: P5, P7, PO7; RH: P6, P8, PO8
N400	300–500	F1, F2, FZ, FC1, FC2, FCZ
LPC	500–700	CP1, CP2, CP3, CP4, CPZ, P1, P2, P3, P4, PZ

Abbreviations: LH, Left Hemisphere; RH, right Hemisphere.

3.5. ERP Results

For the P200 component (see Figure 4), there was no main effect of Word Type, F(1, 29) = 0.71, p = 0.41, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.02. The main effect of Stimulus Condition was significant, F(1, 29) = 5.33, p = 0.03, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.16, with the aversive sound condition (M = 5.01, SD = 0.74) eliciting a larger P200 amplitude than the neutral condition (M = 4.64, SD = 0.7). The interaction between Word Type and Stimulus Condition was significant, F(1, 29) = 4.85, p = 0.04, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.14. Simple effects tests revealed that for concrete words, the difference between the aversive and neutral sound conditions was not significant, F(1, 29) = 0.36, p = 0.55, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.01. For abstract words, however, the aversive sound condition elicited a significantly larger P200 amplitude than the neutral condition, F(1, 29) = 7.80, p = 0.01, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.21. This indicates that the aversive emotional stimulus enhanced early attentional capture for abstract words, suggesting that abstract words received greater early processing resources under emotional learning conditions.

FIGURE 4.

Comparison of P200 component (180 ms–220 ms) mean amplitudes over central regions (based on the average of electrodes C1, C2, Cz, FC1, FC2, and FCz).

For the EPN component (see Figure 5), a significant main effect of Word Type was observed over the right hemisphere, F(1, 29) = 12.9, p = 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.31. Concrete words (M = 0.76, SD = 0.45) elicited a larger (i.e., more negative) EPN amplitude compared to abstract words (M = 1.26, SD = 0.4). There was no main effect of Stimulus Condition, F(1, 29) = 0.74, p = 0.4, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.03. However, the interaction between Word Type and Stimulus Condition was significant, F(1, 29) = 5.66, p = 0.02, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.16.

FIGURE 5.

Comparison of EPN component (250 ms–350 ms) mean amplitudes over the right posterior region (based on the average of electrodes P6, P8, and PO8).

Simple effects tests revealed that for concrete words, the aversive sound condition elicited a significantly larger EPN amplitude than the neutral condition, F(1, 29) = 4.29, p = 0.047, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.13. In contrast, for abstract words, the difference between the two sound conditions was not significant, F(1, 29) = 1.57, p = 0.22, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.05. These results indicate that concrete words, when associated with emotional learning, triggered stronger early automatic emotional processing, and this effect showed a right‐hemispheric predominance.

For the N400 component (see Figure 6), the main effect of Word Type was significant, F(1, 29) = 12.39, p = 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.3, with concrete words (M = −0.68, SD = 0.4) eliciting a larger N400 amplitude (i.e., more negative) than abstract words (M = 0.29, SD = 0.39). There was no main effect of Stimulus Condition, F(1, 29) = 1.37, p = 0.25, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.05. However, the interaction between Word Type and Stimulus Condition was significant, F(1, 29) = 4.3, p = 0.047, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.13.

FIGURE 6.

Comparison of N400 component (300 ms–500 ms) mean amplitudes over the frontal region (based on the average of electrodes F1, F2, FZ, FC1, FC2, and FCZ).

Simple effects tests revealed that for concrete words, there was no significant difference between the aversive and neutral sound conditions, F(1, 29) = 0.61, p = 0.44, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.02. For abstract words, the difference approached significance, with the neutral condition eliciting a larger N400 amplitude than the aversive condition, F(1, 29) = 3.87, p = 0.06, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.12. This pattern suggests that emotional learning enhanced the memory familiarity for abstract words, whereas the processing of concrete words was not significantly modulated by the aversive sound conditions.

For the LPC component (see Figure 7), there was no main effect of Word Type, F(1, 29) = 0.76, p = 0.39, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.03. The main effect of Stimulus Condition was also not significant, F(1, 29) = 2.16, p = 0.15, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.07. However, a significant interaction between Word Type and Stimulus Condition was observed, F(1, 29) = 5.29, p = 0.03, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.15.

FIGURE 7.

Comparison of LPC component (500 ms–700 ms) mean amplitudes over the posterior parietal region (based on the average of electrodes CP1, CP2, CP3, CP4, CPz, P1, P2, P3, P4, and Pz).

Simple effects tests indicated that for concrete words, the LPC amplitude did not differ significantly between the aversive and neutral sound conditions, F(1, 29) = 0.29, p = 0.59, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.01. In contrast, for abstract words, the aversive sound condition elicited a significantly larger LPC amplitude than the neutral condition, F(1, 29) = 6.34, p = 0.02, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.18. This suggests that emotional learning facilitated deeper recollection‐based processing for abstract words, potentially related to the consolidation and elaboration of their memory representations, whereas concrete words showed no evidence of emotional modulation during this late memory stage.

3.6. Liking Ratings

A repeated‐measures ANOVA (2 Word Type × 2 Stimulus Condition) was conducted on the liking ratings from Experiment 2 (see Figure 8). The results showed a significant main effect of Stimulus Condition, F(1, 31) = 15.46, p < 0.001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.33, BFincl = 25.88. Liking ratings were significantly lower for words in the aversive sound condition (M = 5.89, SD = 0.16) compared to those in the neutral sound condition (M = 6.08, SD = 0.15). The main effect of Word Type was not significant, F(1, 29) = 2.51, p = 0.13, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.08, BFincl = 1.2, and the interaction was also not significant, F(1, 31) = 0.14, p = 0.71, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = 0.004. This pattern fully replicates the findings from Experiment 1, confirming that retroactive emotional learning reliably reduced liking ratings for both word types without differential modulation by concreteness.

FIGURE 8.

Comparison of liking ratings across conditions. Mean liking ratings for concrete and abstract words under aversive and neutral sound conditions. Error bars represent the standard error of the mean (SEM). *p < 0.05, ***p < 0.001.

4. Discussion

The present study systematically investigated the differential impact of retroactive emotional learning on the memory processing of neutral concrete and abstract words using a sensory preconditioning paradigm. Findings from both the behavioral (Experiment 1) and ERP (Experiment 2) measures consistently indicate that emotional learning significantly modulated the encoding and retrieval processes of word memory. Notably, a significant memory facilitation effect was observed specifically for abstract words, which exhibited a distinct neural processing dynamic compared to concrete words. The cognitive and neural mechanisms underlying these effects for abstract and concrete words are discussed separately in the following sections.

4.1. Retroactive Enhancement of Abstract Word Memory by Emotion and Its Neural Mechanisms

Behavioral results revealed that when the objects paired with abstract words were associated with aversive sounds, recognition accuracy was significantly higher and reaction times were faster compared to the neutral sound condition. In contrast, for concrete words, no significant differences in accuracy or reaction time emerged between the two sound conditions. Furthermore, when paired objects were linked to neutral sounds, concrete words were recognized with higher accuracy and faster reaction times than abstract words; however, this advantage disappeared when the objects were associated with aversive sounds.

ERP findings further elucidated the neural dynamics of abstract words under emotional learning conditions. During the early attentional stage, abstract words whose paired objects were linked to aversive sounds elicited a larger P200 amplitude than those in the neutral condition, indicating that emotional information rapidly captured attentional resources and facilitated early processing. During the memory retrieval stage, abstract words indirectly associated with aversive sounds elicited a smaller N400 and a larger late positive component (LPC) compared to the neutral condition, reflecting enhanced memory familiarity and augmented recollection‐based processing, respectively (Xie et al. 2024). These findings collectively indicate that emotional learning not only optimizes the early allocation of attention to abstract words but also significantly enhances the efficiency of their memory encoding and retrieval.

These findings can be interpreted from the following two theoretical perspectives. On the one hand, the Representational Substitution Hypothesis (Kousta et al. 2011; Vigliocco et al. 2014) posits that emotional information plays a more central role in the representation of abstract words, partially “substituting” for their lacking sensorimotor grounding. Since abstract words lack direct perceptual referents, their semantic access relies primarily on the prefrontal‐temporal language network (Bi 2021), leading to weaker memory performance in the absence of emotional reinforcement. The “sensory preconditioning” paradigm employed in the current study enabled the retroactive generalization of emotional information to previously learned neutral words, thereby integrating originally neutral lexical items into a unified memory representation endowed with emotional significance. This process effectively compensates for the sensory deficit of abstract words, consequently facilitating their memory performance.

On the other hand, Behavioral Tagging Theory proposes that emotional arousal can trigger neuromodulatory systems (e.g., dopamine, norepinephrine), which “capture” and strengthen weakly encoded memory traces within a specific temporal window (Dunsmoor et al. 2022). According to the Dual Coding Theory (Paivio 1991), abstract words are primarily processed via a single linguistic‐symbolic system, whereas concrete words activate a dual system encompassing both sensorimotor and linguistic‐symbolic codes. Consequently, abstract words exhibit lower encoding strength under baseline conditions, better fitting the criteria of “weak memories.” The retroactive memory enhancement effect observed here likely operates by reinforcing these initially weak memory traces, thereby significantly boosting the consolidation and retrieval efficiency for abstract words.

4.2. Modulation of Concrete Word Memory by Retroactive Emotional Learning and Its Neural Mechanisms

Behavioral results from Experiment 1 revealed that liking ratings for both concrete and abstract words significantly decreased following emotional learning, indicating successful generalization of emotional information to originally neutral words, thereby imbuing them with emotional valence. However, unlike abstract words, concrete words did not exhibit a significant enhancement in memory performance (accuracy or reaction time) under the aversive sound condition compared to the neutral condition. This dissociation between affective evaluation and memory performance suggests that, for concrete words, emotional learning effectively transferred valence to the words but did not translate into measurable benefits in recognition efficiency.

The ERP results provided further insight into this selective modulation. A significant Word Type × Stimulus Condition interaction was observed for the early posterior negativity (EPN) component over the right hemisphere: the aversive (vs. neutral) sound condition elicited a larger EPN amplitude only for concrete words, whereas no such emotional modulation was found for abstract words. This interaction indicates that the effect is not merely a domain‐general threat response but is specifically tied to the representational properties of concrete words.

Within the framework of the Multimodal Induction Hypothesis (Niedenthal 2007), this selective EPN enhancement can be interpreted as follows. Concrete words, by virtue of their rich sensorimotor representations, can directly activate perceptual and motor schemas. When a concrete word is retroactively associated with an aversive sound, its existing sensorimotor representations facilitate a rapid “perceptual reexperiencing” of the emotional content, which is reflected in an early right‐lateralized posterior negativity. This interpretation is consistent with previous evidence showing that the right hemisphere plays a privileged role in processing evolutionarily significant threat stimuli (Langeslag and van Strien 2018) and that emotionally learned pseudowords recruit right posterior regions during the EPN time window (Gu et al. 2022). In contrast, abstract words, lacking such direct sensory bridges, do not show this early EPN modulation, consistent with their reliance on a more linguistically mediated route.

Furthermore, concrete words under the emotional learning condition did not exhibit the same enhancement in memory‐related components (e.g., N400 and LPC) as observed for abstract words. The N400, which reflects familiarity‐based retrieval, and the LPC, which indexes recollection‐based processing, showed no significant emotional modulation for concrete words. This absence of late‐stage memory effects, together with the presence of an early EPN enhancement, suggests that for concrete words, the retroactive attribution of emotional meaning primarily manifests as increased early emotional arousal rather than as structural alterations in the memory system. Because concrete words already possess stable perceptual and semantic grounding, the additional emotional information may be rapidly processed at a sensory‐affective level without substantially changing the strength or efficiency of their memory representations.

Taken together, these findings partially support the Multimodal Induction Hypothesis (Niedenthal 2007), which posits that concrete words can rapidly activate emotional content through their sensorimotor channels. However, our results also indicate that the memory system for concrete words is less sensitive to emotional modulation compared to that for abstract words. This asymmetry suggests that while concrete words enjoy a natural advantage in early emotional integration, this advantage does not necessarily translate into enhanced memory retrieval, possibly because their baseline memory strength is already high due to dual coding (Paivio 1991). In contrast, abstract words, which lack such direct sensory anchors, rely more on emotional compensation to boost their memory performance, as reflected in the later N400 and LPC effects.

The present study systematically reveals differential regulatory pathways of retroactive emotional learning on concrete versus abstract words. The results support the “Representational Substitution Hypothesis,” demonstrating that emotional information can compensate for the weaker perceptual grounding and memory encoding of abstract words, providing direct cognitive and neural evidence for their memory facilitation. This deepens our understanding of the interactive mechanisms among emotion, language, and memory. Furthermore, it offers a valuable experimental paradigm and neural indices for investigating the neural basis of emotional memory generalization, such as the intrusive memories related to linguistic concepts observed in conditions like PTSD.

However, the aversive emotional stimulus employed in this study was limited to screaming sounds produced by female speakers. This choice may constrain the generalizability of our findings, as previous research has shown that emotional voice processing can be modulated by speaker gender and the valence of the emotional expression (Belin et al. 2008; Bestelmeyer et al. 2014). Future research could incorporate emotional stimuli of different valences (e.g., pleasant sounds) and different genders (e.g., male vocalizations) to examine whether the current findings generalize across emotion types and sensory channels. Additionally, while the ERP results revealed neural activity differences across time windows, their spatial resolution is limited, making it difficult to precisely localize the key brain regions (e.g., amygdala, prefrontal, or temporal areas) responsible for integrating emotional information with lexical semantics. Future studies could combine techniques such as functional magnetic resonance imaging (fMRI) to provide more precise functional neuroanatomical evidence for theoretical hypotheses like “representational substitution” and “behavioral tagging.”

5. Conclusion

The present study demonstrates that retroactive emotional learning dynamically modulates the memory processing mechanisms of neutral words, with particularly pronounced facilitative effects on the encoding and retrieval efficiency of abstract words. For concrete words, the influence is primarily evident in the early emotional processing stage. By compensating for the sensory representational deficits of abstract words, emotional learning enhances their attentional allocation and memory consolidation, thereby supporting both the Representational Substitution Hypothesis and Behavioral Tagging Theory. Conversely, concrete words, endowed with rich perceptual‐semantic grounding, exhibit a more rapid activation of emotional information primarily during early stages, without significantly altering their memory structure. These findings not only deepen the understanding of the interactive mechanisms among emotion, language, and memory but also provide a valuable reference for exploring the neural underpinnings of memory generalization in emotional disorders.

Author Contributions

Jingjing Guo: conceptualization, writing – review and editing, funding acquisition, project administration, resources, methodology, supervision. Ailifeire Aili: conceptualization, methodology, investigation, writing – original draft, formal analysis, data curation.

Funding

This work was supported by the National Social Science Found of China (22BYY202) and the National “111” center (B25068).

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.