Thirty-six (N = 36) young adults volunteered to participate in the study (22 F; M_age = 24.38 ± 4.21 years; M_{years of education} = 16.85 ± 1.80). Participants were recruited via advertisements on university bulletin boards and social media. All participants were right-handed as determined by the Edinburgh Handedness Inventory (Oldfield, 1971) (Edinburgh score = 0.89 ± 0.14 points). All were native Italian speakers who learned English as an L2 and had normal or corrected-to-normal visual acuity. No participant had a history of neurological or psychiatric disease or substance abuse, nor was in treatment with psychiatric medications. Due to task requirements, participants were screened for colour blindness before the MRI scanning session. The study was approved by the Human Research Ethics Committee of the San Raffaele Hospital (Milan, Italy). Informed consent was obtained from all participants.

Participants’ bilingual language background was assessed through the Language History Questionnaire (version 3) (LHQ3) (Li et al., 2020). For each participant, L2 AoA, L2 self-reported proficiency, and language dominance were collected. L2 AoA was operationalized as the lowest age at which participants began to speak, read, write, or listen to in the L2. L2 self-reported proficiency was calculated as the weighted sum of participants’ self-rated proficiency on different components of L2 knowledge (i.e., listening, speaking, reading, and writing). Language dominance was determined, for both L1 and L2, as an aggregate score of self-reported proficiency in a language and the estimated time spent every day using that language in different activities (e.g., listening to podcasts and reading). L2–L1 dominance was computed as the ratio of the dominance score of the L2 against that of the L1. The ratio score ranges from 0 to 1, and indicates to what extent a participant is exposed to both languages (0 = the participant is exposed only to the L1; 1 = the participant is equally exposed to L1 and L2). For a detailed description of the LHQ3 measures and their calculation, see Li et al. (2020). Participants’ objective proficiency in the L2 was assessed through the English Proficiency Test (EPT). The EPT (developed by Transparent Language),¹ includes 40 multiple-choice items. Thirty questions evaluate English grammar and conversational knowledge (e.g., participants had to fill in the blank spaces within a sentence with the correct option), and 10 questions assess text comprehension abilities (i.e., participants had to correctly answer questions regarding short English texts). Based on the scores obtained at the EPT (M = 35.25 points; SD = 3.59; range: 25–40), participants were classified as medium-to-highly proficient. Twenty-one participants (58.33%) learned English exclusively in instructional settings; 4 participants (11.11%) learned English exclusively in immersive contexts; 11 participants (30.55%) learned English in mixed contexts (i.e., classroom + self-learning / classroom + immersion / classroom + self-learning + immersion). The descriptive statistics of bilinguals’ background measures are reported in Table 1.

Italian (L1) words were selected from the Italian adaptation (Montefinese et al., 2014) of the Affective Norms for English Words database (Bradley and Lang, 1999). English (L2) words were selected from the Glasgow Norms database (Scott et al., 2019). Each word was originally rated on a 9-point Likert scale across 6 psycholinguistic dimensions (for details, see Montefinese et al., 2014; Scott et al., 2019). For each language, we selected 2 sets of 50 words, one emotional (i.e., negatively valenced) and one neutral. The final set of stimuli included 200 words. Within each language, emotional and neutral words were comparable in terms of several psycholinguistic variables, but differed in terms of valence and arousal (all ps < 0.001). Across languages, the valence and arousal scores were comparable for both emotional and neutral words (all ps > 0.1) (see Table 2). Crucially, the stimuli were matched across L1 and L2 in terms of their affective dimensions [i.e., L1 (EWs) vs. L2 (EWs) valence: p = 0.161; L1 (NWs) vs. L2 (NWs) valence: p = 0.562; L1 (EWs) vs. L2 (EWs) arousal: p = 0.123; L1 (NWs) vs. L2 (NWs) arousal: p = 0.171].

The study employed a 2 × 2 design with Language (words presented in Italian vs. English) and Word Type (negative and neutral words) as within-subjects factors. An event-related design was used to administer the EST in the MRI scanner. The experimental session consisted of 2 runs, 1 run per language, each comprising 2 blocks of 100 trials: 50 emotional trials and 50 neutral trials (2 runs, 4 blocks and 400 trials in total). Words were printed in capital letters in 1 of 4 colours: red (RGB: 255, 0, 0), blue (RGB: 1, 50, 187), yellow (RGB: 255, 255, 0), and green (RGB: 0, 255, 0). Each word was presented twice in each language (i.e., once per block), and each of the 4 colours was presented 25 times within each block. The two fMRI runs were interleaved with the acquisition of a T1 weighted MR image, and their presentation order was counterbalanced across participants. Within each block, trials were arranged in a fixed pseudo-randomised order, so that no more than 4 consecutive trials belonging to the same condition or having the same word colour were presented to participants.

Each trial started with the presentation of a fixation cross (350 ms) appearing in the centre of the screen, and indicating where the subsequent stimulus would appear. In each trial, a word was presented centrally on a black background for 2,000 ms, irrespective of whether the participant had made a response or not. The inter-trial interval (ITI) was jittered with the Dale’s exponential function (Dale, 1999) (mean = 1,720.86 ms; range = 1,127–3,259 ms). Each run lasted approximately 15 min (∼8 min per block). The total scanning time for each participant (including the acquisition of structural data) was approximately 45 min.

Before completing the EST, participants read standardised instructions and underwent a training session inside the scanner. They were instructed to indicate, as quickly and accurately as possible, the ink colour of the words presented to them, without paying attention to the word content. The training session comprised 56 trials (28 trials – 14 emotional, 14 neutral – for each language condition). The stimuli used in the training session were not employed in the experimental session. Participants’ responses were given by pressing a button on an MRI-compatible four-button response box. The Presentation software² was used to present stimuli and collect responses. Accuracy and response latencies in milliseconds (ms) were recorded.

As a complement of the experimental task, at the end of the fMRI session outside the scanner, affective ratings were collected for the emotional and neutral words used as stimuli in the EST. In particular, participants were asked to evaluate the words on two 7-point scales in terms of Valence (from 1 = fully unpleasant to 7 = fully pleasant) and Arousal (from 1 = completely calm/relaxed to 7 = totally activated). Within each language condition (L1 and L2) words were presented in a randomised order, while the order of the language blocks matched that of the fMRI session. Participants were also asked to indicate whether they knew or not the meaning of each L2 word.

Magnetic resonance images were acquired with a 3 Tesla Philips Ingenia CX MR system (Philips HealthCare, Best, Netherlands) equipped with a 32 channels SENSE head coil at C.E.R.M.A.C. (Centro Eccellenza di Risonanza Magnetica ad Alto Campo) of San Raffaele Hospital in Milan (Italy).

For all participants, a high-resolution T1-weighted MPRAGE (Magnetisation Prepared Rapid Gradient Echo) structural image was acquired with the following parameters: Repetition Time (TR) = 9.9 ms, Echo Time (TE) = 4.9 ms, L2ip Angle = 8°, Field of View (FOV) = 260 mm, matrix size = 256 × 256, number of slices = 243, slice thickness = 1.4 mm, voxel size = 0.7 mm × 0.7 mm × 0.7 mm isotropic; Phase Encoding Direction (PE) = R/L; whole brain coverage.

Functional scans were acquired with a fast speed Echo Planar Imaging (EPI) sequence [TE = 33 ms; TR = 2,000 ms; L2ip Angle (FA) = 85°; number of volumes per run = 218; FOV = 240 × 240 mm; matrix size = 80 × 80; 35 axial slices per volume; slice thickness = 3 mm; interslice gap = 0.75; voxel size = 3 mm × 3 mm × 3 mm; PE = A/P; SENSE factor = 2; whole brain coverage]. Four dummy scans preceded each run to optimise EPI image signal.

Structural and functional data were pre-processed using SPM12,³ running on Matlab 2013b. First, the origin of each T1w image was manually set to match the Anterior Commissure – Posterior Commissure (AC-PC) line. Then, T1w images were bias-corrected for intensity inhomogeneities and segmented using the “unified segmentation and normalisation” function in SPM12 (Ashburner and Friston, 2005). Finally, non-cerebral tissue was removed from bias-corrected structural T1w images by means of the “Image Calculator” SPM function (i.e., skull stripping). Functional volumes were slice-time corrected using the first slice as reference point, then realigned to the first volume and unwarped to correct for motion artefacts and geometric distortions. Realigned functional volumes were coregistered to the bias-corrected skull-stripped structural image and normalised to the standard Montréal Neurological Institute (MNI) template. After normalisation, functional volumes were resampled to 2 mm × 2 mm × 2 mm voxels and smoothed with a 10 mm full width at half-maximum (FWHM) Gaussian kernel. For each participant, functional volumes were checked for excessive head motion (>2 mm).

Descriptive statistics of behavioural variables are reported in Table 3. RTs smaller than 300 ms and greater than 2,000 ms were treated as outliers (e.g., Franken et al., 2009). No trials were removed from subsequent analyses according to this criterion. Response accuracy was consistently high in all conditions and was thus not subjected to further analysis.

In the RTs analysis, no effect reached significance (all ts < 1.4, all ps > 0.1). In the analyses of affective ratings, the analysis of valence showed a significant effect of Word Type (Likelihood Ratio Test [LRT] = 375.47, p < 0.001, b = −1.52, SE = 0.04, z = −31.03) – indicating that emotional (negative) words were perceived as less pleasant than neutral words – and a significant effect of Language (LRT = 5.97, p = 0.01, b = 0.11, SE = 0.04, z = 2.46) – indicating that L1 words were perceived as less pleasant than L2 words. The interaction between Word Type and Language was not significant (LRT = 0.05, p > 0.8). The analysis of arousal showed a significant main effect of Word Type (LRT = 327.06, p < 0.001, b = 1.98, SE = 0.07, z = 27.38), with emotional (negative) words being perceived as more arousing than neutral words. No other effect reached significance (all ps > 0.1). The patterns of Valence and Arousal ratings are presented in Figure 1.

Functional magnetic resonance imaging results are reported in Table 4 and Figure 2. A significant main effect of Word Type was observed, revealing that, regardless of language condition, neutral (vs. emotional) words elicited a stronger left-lateralised activity in the putamen, thalamus, and sensorimotor cortex (see Table 4 and Figure 2a). No significant main effect of Language was observed (p-FWE > 0.05). Significant interactions between Word Type and Language also emerged. In particular, the processing of neutral (vs. emotional) words in the L2 condition elicited a greater left-lateralised activation in the superior frontal cortex, in the sensorimotor cortex, and in the thalamus (see Table 4 and Figure 2b). Moreover, the processing of L1 (vs. L2) emotional words elicited a greater activation in the left thalamus, in the right posterior cingulate cortex, as well as mesial and lateral aspects of the left parietal cortex (see Table 4 and Figure 2c). No further contrast reached significance (p-FWE > 0.05).

We observed a similar pattern of word processing in both our participants’ languages. This finding concurs, for example, with those of Sutton et al. (2007) and Grabovac and Pléh (2014), who presented emotional and neutral words to early proficient bilinguals, and that of Eilola et al. (2007), who presented emotional and neutral words to late proficient bilinguals. Sutton et al. (2007), in particular, presented negative and neutral words to early Spanish-English bilinguals who were dominant in their L2, and found significant effects of word type and language on word processing (i.e., bilinguals were slower in categorising the colour of negative words and faster in categorising the colour of L2 words) but no significant interaction between the two factors (i.e., emotional words produced interference in both languages). Eilola et al. (2007) used positive, negative, taboo, and neutral words in late Finnish dominant bilinguals who reached a high level of proficiency in their L2 (English). Although significant differences in RTs between negative and neutral words, as well as taboo and neutral words were observed, language did not produce a main effect, nor interacted with word type on colour categorisation. A similar pattern of results (a word type effect, with no language differences or interaction effects) was observed in a Hungarian minority group assimilated into a Serbian majority culture who acquired their L2 early and reached high levels of proficiency in that language (Grabovac and Pléh, 2014). At first glance, when interpreted alongside this earlier evidence, the lack of a reduced Emotional Stroop effect in L2 versus L1 in our study may be due to the high level of proficiency attained by our participants in their second language. It is worth noting, however, that although the above mentioned studies did not report a significant interaction between language and word type on colour categorisation, they did report an Emotional Stroop effect of comparable size in L1 and L2. By contrast, we failed to detect any significant difference in responding to negative and neutral words in both our participants’ languages. As suggested by Liao and Ni (2022), who focused on the processing of L2 words in Chinese-English bilinguals and failed to report an Emotional Stroop effect on RTs, null effects of emotionality may be attributed to the shallower level of processing at which the EST operates relative to more explicit tasks. On similar lines, Crossfield and Damian (2021) matched a set of neutral, negative, and positive words on multiple psycholinguistic variables other than valence, and used them both in the EST and in a Lexical Decision Task (LDT) administered to monolingual participants. Results indicated a significant processing advantage for positive words over negative and neutral words in the LDT, whereas valence alone did not produce any significant effects in the EST. It was concluded that significant effects of valence would be constrained to tasks where this variable is relevant for task success, and/or tasks that require a deeper level of processing. The results from the affective rating task we administered to participants after EST completion are compatible with this suggestion. We found, in particular, that negative words were perceived as less pleasant and more arousing than neutral words, and that L1 words were overall perceived as less pleasant than L2 words. Whereas in the EST participants are instructed to ignore the words’ content when categorising their responses, affective ratings are expected to drive, to a larger extent, the allocation of top-down attentional resources to the connotative meaning of the stimuli. The different mechanisms underlying word processing in implicit (i.e., EST) and explicit tasks (i.e., affective rating) may account for the task-specific effects of emotional information reported here and in previous research on emotional interference or attentional bias toward negative content. As a side note, it is worth mentioning that we used a version of the EST that required manual responses (i.e., key presses). The Stroop literature has shown that response modality modulates the magnitude of Stroop interference, which is typically smaller in task versions that require manual (vs. verbal) responses. This is arguably because the interference observed with verbal responses results from the contribution of task, semantic, and response conflicts, whereas task conflict does not significantly contribute to the interference observed with manual responses (see, e.g., Kinoshita et al., 2018; Augustinova et al., 2019; Scaltritti et al., 2022).

Among the numerous studies that used the EST to probe emotional interference processes in both healthy and clinical populations, many reported altered psychophysiological and neural (re)activity to negative stimuli in the absence of any detectable behavioural effect (e.g., Compton et al., 2003; Thomas et al., 2007; Fan et al., 2016). The overall pattern of findings reported in this study is in line with such previous evidence, possibly due to the different sensitivity of behavioural and hemodynamic responses to the emotional content of words in L1 and L2 during EST execution. Building on previous research, we expected to find some general differences in activation as a function of word type in frontal-subcortical networks typically engaged in emotional control, but no (or only marginally) significant interactions between word type and language. Contrary to our predictions, the key fMRI finding reported in this study is the significant interaction between word type and language – with stronger activations for L1 versus L2 emotional words – in sectors of the posteromedial cortex involved in attention modulation, episodic memory, and emotional processing. The posteromedial cortex receives major inputs from the dorsal visual stream and somatosensory areas, as well as from regions with functions related to emotion and social behaviour, including the subgenual ACC and the orbitofrontal cortex (see, Rolls, 2019). Moreover, the ventral aspects of the posterior cingulate cortex (PCC) and the adjacent retrosplenial cortex have reciprocal connections with memory structures in the medial temporal lobe (Vogt et al., 2001; Leech and Sharp, 2014). Consistent with its anatomy and connectivity, the PCC, in particular, has been found to be engaged by a range of tasks that involve the modulation of attentional focus (e.g., Hahn et al., 2007; Pearson et al., 2011), episodic memory (see, Rugg and Vilberg, 2013), the retrieval of emotionally salient contextual information (e.g., Maratos et al., 2001), and the evaluation of emotional versus neutral verbal stimuli (e.g., Maddock et al., 2003). The stronger activations for L1 versus L2 emotional words in relevant sectors of the posteromedial cortex is consistent with the hypothesis of a stronger emotional resonance when processing words in a native language. Noticeably, previous behavioural and psychophysiological evidence suggests that L2 proficiency – more than other variables – modulates the extent to which emotional resonance is dampened in a second language (e.g., Ferré et al., 2010; Caldwell-Harris et al., 2011; Degner et al., 2012; Champoux-Larsson and Nook, 2024). Whereas our behavioural data are compatible, at least in part, with this tenet, our fMRI findings seem to downplay the significance of proficiency in favour of other characteristics of bilinguals’ language background, such as learning context. The “emotional context of learning hypothesis” (Harris et al., 2006) suggests that learning a language in the absence of emotion-based communicative interactions leads to a reduced emotionality of that language due to a weaker connection with emotion regulation systems. Our participants were proficient in their L2, and displayed a relatively balanced use of the two languages at time of testing (i.e., they were equally exposed to both L1 and L2 more than they were exposed to the L1 only, as indexed by the L2–L1 dominance ratio metric). However, they learnt English mainly in instructional or mixed contexts (only ∼10% of participants learned English exclusively via immersion), where the opportunities for affective linguistic conditioning are fewer compared to the contexts of acquisition of L1 (e.g., Pavlenko, 2008; Caldwell-Harris, 2014). Therefore, the increased affective response to negative content experienced in L1 (vs. L2) may be attributed to the stronger emotional resonances associated with that language, acquired in a context that was rich in emotional experiences.

Another aspect worth mentioning is that we failed to observe any significant effect of word type or interaction between word type and language in the amygdala, a key region in the brain circuitry of emotion (LeDoux, 2000). As suggested by Chen et al. (2015), however, the activation of the amygdala has been found more frequently in studies where the evaluation of the emotional content of words was explicitly required by the task at hand, and not when emotional valence was task-irrelevant. In addition, the neural systems associated with emotional reaction have been shown to be more active for emotional pictures than for words. This is arguably because pictures are perceived as more biologically salient and emotionally arousing than written verbal stimuli. In support of this hypothesis, a recent meta-analysis of fMRI data on implicit emotional processing in monolinguals (Feng et al., 2021) suggested that affective pictures and words modulate implicit emotional processing differently, and recruit distinct neural systems. In particular, only negative pictures, and not words, could reliably elicit activation in the amygdala.

Several issues of the present study call for caution when interpreting the results. First, as we did not manipulate dimensions of bilinguals’ language background such as L2 proficiency or learning context, inferences about the contribution of these variables on emotional processing are only tentative. Second, although this is one of the very few studies that paired a behavioural task with fMRI to investigate affective processing in bilinguals, the evidence we provide comes from the investigation of single, decontextualized affective stimuli. The implementation of natural language in more ecological paradigms (e.g., naturalistic viewing paradigms –Sonkusare et al., 2019; Bellini et al., 2024) may help future research drawing a more reliable and lifelike picture of how bilinguals process affective language in everyday communication.

Overall, a number of theoretical and methodological implications can be drawn from our findings. A first theoretical implication is that the role of bilinguals’ languages in their affective repertoires is complex, and arguably conditional upon numerous factors. In our sample – which comprised proficient bilinguals who learned their L2 mostly in instructional or mixed contexts – stronger activations were observed for processing L1 versus L2 emotional words in sectors of the posteromedial cortex involved in attention, memory retrieval, and affective processing. However, the interpretation of fMRI findings could not be guided by our behavioural results, which showed no difference in the automaticity of emotional word processing across L1 and L2. This is possibly due to the different sensitivity of behavioural and hemodynamic responses to the emotional content of words in L1 and L2, but may also suggest that the EST is not suitable to capture potential emotionality effects on word processing across languages. The EST has been widely deployed in clinical studies to investigate individual differences in emotional processing by using emotional words related to a particular individual’s pathology or dysfunctional personality trait (e.g., anxiety, phobia, depression, substance addiction – see, Williams et al., 1996). However, when applied to explore implicit affective word processing in healthy speakers – both bilinguals and monolinguals – this task has produced mixed results, often in contrast with psychophysiological and brain imaging data (see, for reviews, Phaf and Kan, 2007; Jończyk, 2016). The present study provides further evidence that the EST, at least in its colour-word “manual” version, may not be suitable to capture potential emotionality effects that can be seized through fMRI analysis. Future research should possibly make use of other tasks involving emotional interference, or other versions of the EST (e.g., the “word-face” variant) involving more intense emotional conflict (see, Song et al., 2017).