Are brain activity changes underlying rare word production after learning speciﬁc or do they extend to semantically related rare words?

Learning words in the mother tongue is a fundamental lifelong skill that involves complex cognitive and neural changes. In adults, newly learned words affect the organization of the lexical e semantic network and, compared to words that have been in the lexicon for longer, they activate the same cortical areas, but more extensively and/or intensively. It is however still unclear (1) which brain and cognitive processes underlying word production change when infrequent/unknown words are compared before and after learning and (2) whether integrating newly learned words impacts word speciﬁc processes or has a broader impact on unlearned words. The present study aims to investigate the electrophysiological changes underlying the production of rare words induced by learning and the effect of learning on an unlearned list of rare words belonging to the same semantic categories. To this end, 24 neurotypical adults learned one of two matched lists of 40 concrete rare words from 4 semantic categories. EEG (electroencephalographic) recordings were acquired during a referential word production task (picture naming) of the learned and unlearned words before and after the learning phase. The results show that the production of rare word is associated with event-related (ERP) differences between before and after learning in the period from 300 to 800 msec following the presentation of the imaged concept (picture). These differences consisted in a larger involvement of left temporal and parietal regions after learning between 300 and 400 msec i.e., the time window likely corresponding to lexical and phonological encoding processes. Crucially, the ERP changes are not restricted to the production of the learned rare words, but are also observed

Native language word learning Electrophysiology Microstates Word production Lexicalesemantic integration a b s t r a c t Learning words in the mother tongue is a fundamental lifelong skill that involves complex cognitive and neural changes.In adults, newly learned words affect the organization of the lexicalesemantic network and, compared to words that have been in the lexicon for longer, they activate the same cortical areas, but more extensively and/or intensively.It is however still unclear (1) which brain and cognitive processes underlying word production change when infrequent/unknown words are compared before and after learning and (2) whether integrating newly learned words impacts word specific processes or has a broader impact on unlearned words.The present study aims to investigate the electrophysiological changes underlying the production of rare words induced by learning and the effect of learning on an unlearned list of rare words belonging to the same semantic categories.
To this end, 24 neurotypical adults learned one of two matched lists of 40 concrete rare words from 4 semantic categories.EEG (electroencephalographic) recordings were acquired during a referential word production task (picture naming) of the learned and unlearned words before and after the learning phase.
The results show that the production of rare word is associated with event-related (ERP) differences between before and after learning in the period from 300 to 800 msec following the presentation of the imaged concept (picture).These differences consisted in a larger involvement of left temporal and parietal regions after learning between 300 and 400 msec i.e., the time window likely corresponding to lexical and phonological encoding processes.
Crucially, the ERP changes are not restricted to the production of the learned rare words, but are also observed when participants try to retrieve words of a list of semantically and lexically matched rare words that they have not learned.
The ERP changes on unlearned rare words are weaker and suggest that learning new words induces boarder effects also on unlearned words.© 2024 The Author(s).Published by Elsevier Ltd.This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Introduction
Word learning is a skill that is used daily, even in adulthood (Hartshorne & Germine, 2015), whether to learn new words in the mother tongue (L1), in a second language (L2), or when recovering/(re)learning L1 following brain damage.Word learning in native language occurs in a variety of contexts: often, when encountering words in the media or in conversations, and more actively/systematically, when adults learn a new discipline or profession.Learning words is a fundamental cognitive skill, but is also complex since new words need to be integrated into the mental lexicon and interconnected with other words.Integrating a new word involves several processing steps, from short-term learning to full consolidation, which means that word learning takes time (Ven et al., 2015).This also leads to changes at the neural level (Fargier & Laganaro, 2020;Cornelissen et al., 2004;Takashima et al., 2017, Takashima, Bakker, Van Hell, Janzen, & McQueen, 2014;Paulesu et al., 2014) that has been particularly studied by comparing activations for newly learned words to those of words that were already known.Nonetheless, changes that occur before and after learning, as well as the changes induced by learning in the lexical system, have been poorly studied in the context of word production.In particular, we do not know (1) which specific word encoding processes allow to produce a rare word that has been recently learned and (2) whether the lexicalesemantic integration of newly learned words leads to a spillover effect to related words.In the present study, by comparing word production before and after learning a list of rare word and another one matched but not learned, we will investigate the electrophysiological correlates underlying the production of rare words following learning in neurotypical adults, as well as the effect of learning on semantically related rare words to investigate the changes induced by learning in the interconnected lexical-semantic system.
In the following, we will first summarize the cognitive and brain processes involved in word learning before focusing on the electro/magnetoencephalographic (EEG/MEG) eventrelated (ERP) correlates underlying the recognition/comprehension and production.

Cognitive and neuroimaging processes of new word learning
Learning a new word begins with the first exposure to an unfamiliar word (Gonz alez-fern andez & Schmitt, 2020), and it is only after several occurrences of exposure that it is generally possible to recognize/understand it.It has been suggested that after a period of about 24 h following explicit word learning, new words can be qualified as part of the lexical network, based on the observation that newly learned words influence the recognition and/or production of known words (Clay et al., 2007;Davis et al., 2009;Gaskell & Dumay, 2003;Kapnoula et al., 2015;Tamminen et al., 2010).This phenomenon refers to lexicalesemantic integration, which marks the emergence of connections between the newly learned word and words already present in the mental lexicon.The new word becomes a part of the interconnected and distributed lexicon through its semantic features and/or its associated words (Vivas et al., 2019).Learning changes the weight of the connections in the semantic system through experience (Oppenheim et al., 2010).Most studies have investigated word learning processes using word recognition paradigms, leaving aside the production.Yet, one of the ultimate goals of learning words is to be able to produce the word, a skill that is generally achieved later in the learning process (Keenan & MacWhinney, 1987).The focus on word recognition/comprehension tasks in word learning dominates behavioral studies as well as neuroimaging studies as stressed below.In terms of brain activations, Davis and Gaskell (2009) have modeled the word learning process as involving first the hippocampal memory systems (short-term learning) and the brain's neocortical networks as a result of lexicalesemantic integration (Complementary Learning System (CLS) model).Some neuroanatomical studies have attempted to validate the CLS model by contrasting neural activations after short-term learning and after a one-week consolidation period in recognition (Takashima et al., 2017).As predicted by the CLS model, short-term learning initially involved activation of the bilateral hippocampus, likely reflecting episodic memory retrieval, whereas after one week of consolidation, several cortical areas generally involved in semantic and phonological processing were observed (Functional Magnetic Resonance Imaging (fMRI): Takashima et al., 2017;Takashima, Bakker, Van Hell, Janzen, & McQueen, 2014; Positron Emission Topography (PET): Paulesu et al., 2014).This suggests that learning involves a change in the structures required for production and/or recognition.Neurofunctional studies have also focused on the differences in neural activation between known and newly learned words after lexicalesemantic integration.They highlight the involvement of the same cortical structures, but with increased responses and more extensive brain activation for novel words, both in recognition/understanding and in production (Cornelissen et al., 2004;Davis et al., 2009;Gr€ onholm et al., 2005;Hult en et al., 2009).It has been put forward that this is due to enhanced lexicalesemantic and/or lexicalephonological retrieval processes, and thus continuous consolidation following integration.
Anatomical neuroimaging studies have subsequently demonstrated the involvement of different structures during word learning, starting with hippocampal regions and then involving brain structures common to already known words, but differences remain in the recognition/production of frequent words.Although the observed changes have been attributed to lexicalesemantic and/or lexicalephonological retrieval processes, the lack of temporal resolution of fMRI studies does not allow to address the cognitive processes underlying change during learning.Production and recognition involve the recruitment of a large number of cognitive processes which through the time courses have been largely investigated with EEG/MEG approaches (see meta-analyses for production: Indefrey, 2011;recognition: Holcomb & Grainger, 2006).Some studies have more specifically investigated the cognitive processes involved in new word learning, as summarized in the next section.

Electrophysiological correlates of new word learning
Studies of event-related potentials (ERPs) have mainly focused on the comparison between newly learned and already known words, mostly using word recognition/comprehension tasks such as semantic judgment (Kaczer et al., 2018), semantic decision (Perfetti et al., 2005), meaning generation (Frishkoff et al., 2010), semantic priming in a semantic association task (Bakker et al., 2015a,b).These studies have identified two stages in the word learning process as predicted by the CLS model.First, amplitude differences have been observed in several late components between newly learned and frequent words immediately after learning (Bakker et al., 2015a;Frishkoff et al., 2010;Perfetti et al., 2005).This has been interpreted as a lack of lexicalesemantic integration or an episodic memory trace for newly learned words.This interpretation is also related to the disappearance of these differences after a consolidation period of at least 24 h (Frishkoff et al., 2010), when ERP components of newly learned words do no longer differ from frequent words.Nevertheless, it has also been highlighted that the word learning process is not an all or nothing process, as some persistent ERP differences can last over time, revealing a kind of gradualness in the learning process (Bakker et al., 2015b).To our knowledge, only two studies have investigated word production of newly learned words (a MEG study by Cornelissen et al., 2004; an EEG study by Fargier & Laganaro, 2020).In the first MEG study, the cortical dynamics of the production of newly learned infrequent words from one semantic category (tools, n ¼ 20) in native language was investigated among five participants.The results show that learning the names and/or meanings of unfamiliar objects activates the same cortical network as familiar words (Cornelissen et al., 2004).Similar results were obtained in the study of Fargier and Laganaro (2020) by comparing frequent and newly learned words of two semantic categories (tools and musical instruments, n ¼ 50) using spatio-temporal (microstate) ERP analyses.They showed that the correct production of newly learned words compared to frequent words involves similar brain processes, but some of these processes are lasted longer for new words.Based on the temporal dynamics of word production (Indefrey, 2011), they concluded that both lexical and phonological encoding processes last longer in the production of newly learned words.
In sum, most studies on the neural changes underlying the integration of new words have only investigated words that are correctly recognized after learning; the few studies on correctly produced new words suggest that the encoding processes only differ in the time-dynamics from those of words that have been used for longer.However, which cognitive processes change before and after learning to enable the production of newly learned words are not really addressed in any of these studies.Furthermore, the mental lexicon and the semantic system being highly distributed and interconnected (Oppenheim et al., 2010), the integration of a newly learned word may impact the connected semantic concepts and words.However, while previous studies limited word learning to a few semantic categories, the possible impacts of learning on semantically related words that have not undergone learning has not been investigated so far and may be a window on the understanding of the interconnected language system.

Present study
Studies on word learning have mainly focused on neural structures and ERP components comparing the neural processing of newly learned words with familiar words after consolidation and showed persistent differences with already known words.However, they have not investigated the processing changes underlying word learning by comparing the production of newly learned words before and after learning, i.e., when they cannot be retrieved and/or produced and when they are produced correctly.Moreover, a possible general effect of learning and integrating new words into the lexical system has not been considered in previous studies.In other words, the question of the present study is which word production processes allow a newly learned word to be encoded and retrieved after learning?Which of these processes are specific to correctly produced learned words and which are also observable in non-learned semantically related rare words?The aim of the present study is to investigate the neural processing of newly learned words by comparing the ERPs related to word production before and after learning for a list of learned rare words and a matched list of unlearned rare words.The ERP changes will be investigated by relating them to the time windows of the processes involved in word production (Indefrey, 2011) in order to gain insight into the word encoding processes that change before and after word learning.
To this end, participants will successively learn a list of unfamiliar words in their native language associated with real objects/concepts from four distinct semantic categories (tools, musical instruments, flowers, and fruits/vegetables).They will be tested behaviorally and ERP recordings will be performed during a picture naming task before and after the learning phase.The design thus allows, on one hand, to identify the electrophysiological correlates of newly learned word production immediately after learning, but also, on the other hand, to identify the effects of learning new words on the production of other rare words from the same semantic categories, but without concrete learning.The second aim of the present study is to better understand whether word learning induces changes only in the newly learned words themselves, or whether the effects extend to words of the same semantic categories.

Method
We report how we determined our sample size, all data exclusions, all inclusion/exclusion criteria, whether inclusion/ exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study.

Participants
33 native French speakers 1 (19 females) participated in the study.They were aged between 20 and 35 years old (mean 25 years old) and were all right-handed according to the Edinburgh Handedness Scales (Oldfield, 1971; mean 86.9).They reported normal or corrected-to-normal vision and had no neurological, learning or developmental disabilities.These inclusion criteria were applied before the study.All participants gave their written informed consent and were paid for their participation.The study protocol was approved by the local research ethics committee of the University of Geneva (N PSE.PSE20201003.06).

Overall study design
The study took place over three days, including test and learning sessions.Participants were presented with two matched lists of 40 rare concrete words in French (see Section 2.2.2).The "learned list" was learned during two learning sessions (Day 1 and Day 2), while the "non-learned list" was only used during the test sessions.On the first day ('Day 1'), participants were tested on the production of all the 80 words with a colour picture naming task during an EEG recording.
Immediately after the Day 1 test, they learned one of the two lists of 40 rare words (for about 30 min with a digital learning approach).On Day 2, learning continued on this list for 1 h, resulting in a total of one and a half hours of learning for the 40 words (seven exposures to each word).On Day 3, a second test session with a picture naming task was performed and EEG was acquired.The learned/non-learned lists were counterbalanced between subjects (see Fig. 1).

Picture naming and rare word learning tasks
80 rare words (Appendix A) with very low lexical frequency (<.38 per million words) were selected from the French database Lexique (New et al., 2001) on the basis of a pre-test on 88 rare words.In such pilot, only words with a maximum of 30% naming among 20 participants were selected.The 80 stimuli belonged to four semantic categories (flowers, fruits/ vegetables, tools and musical instruments) with 20 rare words from each category.Stimuli were divided in two lists of 40 items, matched on relevant psycholinguistic variables (i.e., number of syllables, number of phonemes, lexical frequency and sonority values of the first phoneme).For each item, four different colored pictures were selected from free internet databases and cut out to be the same size.Four different pictures were chosen to have different representations of the same concept in the assessment and learning sessions.

Test sessions
Participants were tested individually with a picture naming task in a soundproof, dimly lit room, sitting 60 cm in front of a computer screen.Items (colored pictures) were presented twice on a gray screen in a pseudo-randomized order (no more than two consecutive items from the same semantic category).The two presentations were presented in separate blocks, preceded by two warm-up filler trials.Four different orders were created for each task in order to counterbalance them across participants and assessment sessions.This was done to avoid any effect of the order of stimulus presentation.The E-prime 2 (E-studio) software presented the trials and recorded the vocal responses.Vocal responses were digitized and recorded for later verification of their latencies and accuracy using a speech analysis software (CheckVocal 2.2.6).Each trial began with a fixation cross in black Arial font, size 24, presented for 500 msec at the center of the screen.Afterwards, a gray screen was presented for 100 msec.The stimulus then appeared for 3000 msec, followed by a gray screen for 500 msec, before the presentation of the next trial.Participants were informed about the presentation of images of unfamiliar objects and were asked to name them aloud as accurately and quickly as possible.If they did not know the name of the object, they were asked to say "no" aloud.Each picture was presented twice in two separate pseudorandomized blocks, separated by a self-paced break, for a total of 160 stimuli.

Learning sessions
Each participant learnt only one list of rare words on Day 1 and 2 either with immersive virtual reality or with a digital static learning method in the other half.The virtual reality application, developed in Unity (2019.4.1) and Blender for the creation of 3D objects, consisted in a market where participants were asked to buy the items of the shopping list.Participants were equipped with a virtual reality headset (Oculus Rift S).The other digital learning application was on a tablet; it was developed using Unity (2019.4.1) and Android Studio.It included a typical picture naming task (see Fig. 2). 1 No a priori effect size was available to perform power analysis and determine the number of participants (and items), as to our knowledge there are no studies that compared the electrophysiological correlates of rare word production before and after learning on a learned and non-learned list.In addition, as we aimed to perform mixed model analysis, the number of items should also be considered (Brysbaert & Stevens, 2018).Therefore, the number of participants and items was based on previous studies: Fargier & Laganaro, 2020 included 16 participants and 50 items and Cornelissen et al., 2004 had 5 participants and 20 items.In order to obtain sufficient statistical power, we decided on at least 25 participants with 40 items per list.

Data analysis
Exclusion criteria were applied before analysis, including not having completed the study, not having a high enough learning score (less than 10%) and too few epochs for EEG analysis (less than 25 epochs per day or outlier ERP signal identified by Ragu software (Koenig, et al., 2011)).24 participants were included in the final analyses.While one participant did not complete the study, two other participants were excluded due to insufficient learning rate.Eight additional subjects were removed due to too much contaminated EEG signal.

Behavioral analysis
Participants responses were checked using the CheckVocal software (CheckVocal 2.2.6, Protopapas, 2007) and were categorized as correct, incorrect or 'no' responses.Correct responses corresponded to the word associated with the picture/object to be named, with a tolerance of phonemic transformations on a maximum of two phonemes of the word (i.e., /eʃənilɯɑʁ/ for /eʃənjɯɑʁ/).The "no" response corresponded to a clear and unambiguous "no" during the presentation interval of the picture.For correct and 'no' responses, the vocal onset was manually defined from the waveform and spectrogram using the mentioned software.Incorrect responses were all other responses that were not among the accepted/correct or "no" responses.Accuracy (correct responses corresponding to correct answers and errors to "no" responses and incorrect responses) was analyzed with generalized linear mixed effect models (Baayen et al., 2008) using the R software (Venables & Smith, 2003).Each list of words ('learned list' and 'non-learned list') was analyzed separately with participants and items as random effects and test sessions (Day 1 and Day 3) as a fixed factor.

Pre-analyses
During the pre-and post-learning test sessions (Day 1 and Day 3, respectively), EEG signal was continuously recorded with the 128-channel Active-Two Biosemi EEG system (Biosemi V.O.F Amsterdam, The Netherlands).Signals were sampled at Finally, contaminated electrodes (up to 20% of the 128 electrodes) were interpolated using 3-D splines interpolation (Perrin et al., 1987).ERPs were analyzed from 0 msec, at the onset of the picture presentation, to 800 msec after stimulus onset on epochs averaged per participant, per session and list.The 800 msec time window was selected based on the mean reaction times of participants (see results) in order to cover word encoding/planning time far from articulation.To assess the effect of learning, including the difference between correct and incorrect responses, single epochs were also analyzed after being interpolated with the same approach.

ERP analyses
Analyses on waveform amplitudes were performed at each time point on the 128 electrodes over the entire data set using a classer-massed statistical approach run with the permu-co4brain R package by Frossard and Renaud (2021).Nonparametric Threshold-Free Cluster Enhancement (TFCE) permutations were computed with 5000 randomizations and alpha set to 5% to identify the time windows in which significant divergences in ERP amplitudes were observed.Analyses were performed to compare Day 1 and Day 3 on each list separately.
For the spatio-temporal/topographic analysis, we first tested whether the different conditions (test sessions and lists) showed consistency in Global Map Dissimilarity (GMD) Fig. 2 e Example of an immersive virtual reality trial for the item "sapotille" (a) and of a tablet trial for the item "tabla" (b).
c o r t e x 1 7 8 ( 2 0 2 4 ) 1 7 4 e1 8 9 using nonparametric randomization tests (L2 normalization and 5000 runs) in the RAGU software (topographic consistency tests; Koenig et al., 2011) (see Appendix Fig. B.1).Then, we computed the topographic analyses on the entire time windows of the analyzed ERPs.Significant differences in map topographies between test sessions for each list were first tested using a topographic analysis called TANOVA, which consists of nonparametric randomization tests based on GMD (Koenig et al., 2011).We only considered topographic differences that lasted longer than 50 ms, with an alpha criterion of .05 for each time point.For the spatio-temporal segmentation analyses, we performed an analysis in a 2 Â 2 design (two lists and two test sessions e Day 1 and 3) over all the ERP time period.Spatio-temporal clustering of stable microstate maps was performed on grand averages with the RAGU software.This procedure segments ERPs into periods of quasi-stable global electrophysiological patterns at the scalp (i.e., ERP microstates or topographic maps) by compressing the variability of ERPs into a set of template maps that summarize the data and serve to determine which topographic template best explains participants' ERP responses in each list and period (Koenig et al., 2011;Michel & Murray, 2012).The optimal number of ERP maps was selected according to the one that best explained group-averaged data across test sessions.Statistical validation of this analysis was obtained using a backfitting procedure, which consisted of comparing each of the microstates observed in the grand averages with the momentto-moment scalp topography of individual subjects' ERPs in each list and session.This procedure allowed us to extract the topographic index of interest, being the duration of each topography (number of TFs).
2.4.2.1.SOURCE LOCALIZATION.Source localization analysis was performed using the Cartool software.Magnetic Resonance Imaging (MRI) from the MNI-152 adult symmetric atlas (Fonov et al., 2011) was used.The whole brain was extracted from the MRI (combining gray and white matter) and the electrode position was coregistered on the MRI head of the adult template.The solution space was estimated using 6000 solution points and the lead field was calculated using the LSMAC 6shell isotropic method.Inverse solutions were estimated using an inverse matrix (LORETA) based on the grand mean of each condition.Before evaluating the source localization of the time of interest, an estimation of the sources underlying the visual evoked potential (VEP) P100 was performed to ensure an effective localization of the occipital areas.The time period of interest was determined by the results of the microstate analysis and will therefore be further described in the result section.

Behavioral results
During the first test session ('Day 1'), very few items were named correctly (mean: 0.1%, range: 0e4).These items were removed for each participant from the computation of learning accuracy on Day 3 (defined as correct production of rare words not known on Day 1).Learning accuracy at Day 3 reached 46% for the learned list (range: 15e86%), which represented a significant improvement compared to Day 1 (z ¼ 14.78, p < .001,dl ¼ 1915, b ¼ 7.37, SE ¼ .49).The mean reaction time for correctly produced words at Day 3 was 1487 msec (SD: 453 msec).For the non-learned list on Day 3, only 1.45% of responses were correct across all participants.

Waveform analysis
Waveform analyses were computed on the two test sessions (Day 1 and Day 3) for each of the two wordlists (see Fig. 3a).For the learned list, statistical analyses of ERP amplitudes between Day 1 and 3 revealed significant differences on a cluster of central-posterior electrodes in two time windows: from 260 msec to 460 msec and from 480 msec to 800 msec on parietal and posterior electrodes, with amplitudes increasing from Day 1e3 (Fig. 3a, b).The unlearned list showed differences in amplitudes between Day 1 and Day 3 on parietal electrodes from 570 msec to 740 msec after picture onset (Fig. 3a, b), reflecting an increase in amplitude on Day 3 compared to Day 1 (Fig. 3b).Given that amplitudes increased in both lists across test sessions, to assess whether the increase in amplitude was greater for the learned list, we computed the difference in amplitude between the two lists.The median of absolute values of amplitudes was computed for each participant on each time point across electrodes on each day and list in the time window leading to the largest differences (570e740 msec).Then, we calculated the difference between the median at Day 1 and Day 3 separately for each list on each time point.Finally, we used the median amplitude difference in absolute value across all time frames for each participant to carry out a mixed model analysis (amplitude difference ~list þ (1|Subject)).This analysis was used to assess if the differences in median amplitudes were greater for one of the lists.The model revealed a significant effect of list with a higher amplitude difference between Day 1 and Day 3 for the learned list than for the nonlearned list (b ¼ À.43, SE ¼ .2,t(23) ¼ À2.13, p ¼ .04).

Global dissimilarity analysis
For the global dissimilarity analysis, topographical analyses of variance (TANOVA) were conducted separately for each list across test sessions (Day 1 and 3).See Fig. 3a, c and b, c.For the learned list, the Day 1e3 comparison showed a large period of topographic difference between 290 msec and 590 msec, whereas for the unlearned list, the TANOVA differences were limited to a 50 msec period, between 350 msec and 400 msec.

Microstate analysis
The analysis was computed on a 2 Â 2 design with the two learned lists and the two test sessions (Day 1 and 3).Results revealed 10 different periods of quasi-stable electrophysiological activity at the scalp during 800 msec, accounting for 97.5% of the variance.A change in topographic stability on the grand averages is observed between Day 1 (map 5 in Fig. 3d) and Day 3 (map 6 in Fig. 3d) for both the learned and the unlearned list.
In order to further evaluate significant differences between days, we applied the topographic fitting procedure to the participants' ERPs in the time window covering the period of c o r t e x 1 7 8 ( 2 0 2 4 ) 1 7 4 e1 8 9 significant differences observed in the TANOVA, including the entire P2 component.Thus, a time-window spanning from 200 msec to 800 msec was chosen, with the maps observed on the grand averages (maps 4, 5, 6, 7, 8, 9 and 10, see Fig. 3).The map duration in each individual ERPs was used for the statistical analysis to compare each session per list.Results on the learned list showed significant differences between Day 1 and Day 3 on Maps 5 to 10 (duration increased for Maps 6, 8 and 10, while it decreased for the remaining Maps from Day 1 to Day 3).Differences between the two Days for the unlearned list were significant on Maps 6 to 10 (duration increased for the Maps 6, 8 and 10 during the Day 1 in comparison to the Day 3).The detailed results are presented in Appendix Tables C.1 and C.2 for the learned and unlearned lists, respectively.3.2.3.1.DIFFERENCES BETWEEN 300 MSEC AND 400 MSEC.In order to determine whether the two microstates in the 300e400 msec (Maps 5 and 6) were indeed different across test sessions on both lists, a further fitting procedure was performed.The statistical analyses were carried out with mixed models, run separately for each list, with test sessions (Day 1 and 3) as fixed effects, participants as random effects and duration of Maps 5 and 6 as dependent variable (Negative Binomial).The model showed a significant interaction effect between maps and test sessions for both lists (learned list: (X 2 (1) ¼ 27.95, p < .001),non-learned list: X 2 (1) ¼ 11.17, p ¼ .001).This reveals an inverse dynamic between Map 5, which tended to decrease in presence in terms of number of TFs from Day 1 to Day 3 (À24 msec for the learned list and À15 msec for the unlearned list), and Map 6, which increased in duration at Day 3 (þ24 msec for the learned list and þ15 msec for the unlearned list) (see Appendix Table D.1).
3.2.3.2.LEARNED VERSUS UNLEARNED WORDS.Since differences between Day 1 and 3 were observed also on the unlearned list and that the learned list at Day 3 included both correct and incorrect productions, we ran analyses on single epochs/trials by separating correct (learned) words from incorrect (unlearned) productions of the learned list only.Thus, Maps 4 to 10 were fitted from 200 msec to 800 msec in all epochs according to whether they corresponded to learned words or not.The statistical analyses were run with a mixed-model including the leaned/unlearned status of the items and the maps as fixed factors, the participants and the items as random effect and the map duration as dependent variable (Negative Binomial).The model showed a main effect of learning (X 2 (1) ¼ 8.82, p ¼ .002), a main effect of map (X 2 (6) ¼ 1536, p < .001)and an interaction effect between those two variables (X 2 (6) ¼ 29.13, p < .001).Further analyses presented in Appendix Table E.1 indicate that Maps 6, 7, 8 and 10 increase in correctly produced (learned) words in comparison to non-correctly produced (non-learned) words of the learned list.

Source localization
For the 300e400 msec time windows, which showed consistent changes in the previous analyses, source localization was performed on the grand average for each test session.We combined the learned and unlearned lists because of the same process changes.In this time window, Day 1 is characterized by strong activation of the occipital lobe, including the middle and inferior occipital gyrus, in addition to some activations in the right parietal lobe, the superior temporal lobe on both sides (see Fig. 3f) and some prefrontal activations (ventrolateral prefrontal cortex).Day 3 showed similar responses in the occipital lobe, apart from activations in the left and right superior and inferior parietal lobes.Pronounced activations in the superior and middle temporal lobes were also observed, especially in the left hemisphere.

Discussion
The aim of the present study was to investigate changes in the electrophysiological correlates of word production after learning.Specifically, we compared the ERP changes related to word production that occur before and after learning rare words in the mother tongue.Morevoer, we investigated whether those changes are specific to the learned words or extend to an unlearned list of rare words from the same semantic categories.After two learning sessions of 40 rare words from four semantic categories, the participants showed on average 45% improvement and no behavioral changes on the unlearned word list.At the neurophysiological level, ERP changes in word production after the learning phase were observed on waveform amplitudes, TANOVA and topographies between 300 and 800 msec, but particularly between 300 and 400 msec.This time window is underpinned by a change in processes before and after learning, with a switch from one microstate (Map 5 in Fig. 3) to another (Map 6) and with a larger involvement of left temporal and parietal regions after learning.Crucially, the neurophysiological changes between Day 1 and 3 were not restricted to the learned list, but were also observed on the unlearned list.Overall, the ERP changes are in the same direction, although they are larger and temporally more extended for the learned list than for the unlearned one.
In the following, we will discuss the changes in the electrophysiological correlates of word production before and after word learning, before hypothesising why ERP changes were also observed on the unlearned list.

ERP changes related to learning
The results of the present study report robust and converging electrophysiological effects on the learned list when comparing word production before and after learning.These changes occur in a time window of word production encoding extending between 300 and 800 msec, but with the most prominent and converging changes between 300 and 400 msec.First, waveform amplitude analysis showed that learning induced an increase in amplitude from 260 to 460 and from 480 msec to 800 msec in the production of words from the learned list, suggesting changes in the neuron firing to produce learned words when comparing before and after learning.An increase in amplitude has previously been interpreted as a differential ERP effect due to memory (Korochkina et al., 2023;Paller & Wagner, 2002) or an episodic training effect in word recognition/understanding tasks (Balass et al., 2010;Howseman et al., 1999;Perfetti et al., 2005).However, the amplitude increase in the present study was not restricted to the learned list, but also extended to the unlearned list (with only <1% of words correctly produced).Subsequent analyses showed that the effects on the learned list were larger than those of the unlearned list.This suggests that neural firing overall increases due to learning, but it increases more for the learned items.This result is in line with previous observations showing that increased neural discharge is associated with better learning (Balass et al., 2010).
Changes in amplitudes are observed over a large time window during the encoding processes underlying word production, starting from the end of the P2 component and extending over a large part of the word planning time (for mean RTs of 1487 msec).Meta-analyses of the time course of encoding in word production (Indefrey, 2011) suggest that production encoding involves multiple cognitive processes, including lexical selection associated to the P2 component, phonological encoding and phonetic (motor speech encoding) processes.Here, the observed changes in amplitude can be interpreted as larger network/firing in lexicalephonological encoding processes.However, they may also be related to changes in the underlying brain processes.This issue can be further explored with high-density topography analyses.Topographic analysis has the valuable advantage of providing information about whether microstates (that have been associated with mental processes, (see Michel & Koenig, 2018 for a review), tend to change in terms of the source of activations (i.e., qualitative difference) or in terms of duration (i.e., quantitative difference).
In fact, changes in microstates occur between 300 and 600 msec, with TANOVA differences and qualitative and quantitative differences between before and after learning.The most important ERP changes occur between 300 and 400 msec.On the topographies, a shift in the median axis is observed on the two maps that characterize pre-and postlearning.This is consistent with the source location analysis, which showed predominantly right activations prior to learning in occipital and temporal cortex, followed by a bilateral shift, with more pronounced left parietal and left temporal activities after learning.However, no activations were found in hippocampal regions, as predicted by the Complementary Learning System model.This may be partly due to the low spatial resolution of EEG, which makes it difficult to detect subcortical activations from scalp data.However, it may also be related to the present study design, which did not capture episodic memory effects and only measured ERPs of already consolidated words.
Before being able to further interpret these results, a few surprising results have to be addressed.First, learning does not involve changes on the P2 component peaking here at around 220 msec (see Fig. 3b).This component is generally associated with lexical selection (Indefrey, 2011).Actually, the microstate of the P2 (map 4 in Fig. 3) seems to last longer after learning on the grand-average.This result was not significant in the fitting analysis with the individual ERPs.This is quite surprising because producing newly learned words has largely been associated with an increase in the time needed for lexical selection (Fargier & Laganaro, 2020).This result has generally been associated to the addition of competitors (Strijkers et al., 2010(Strijkers et al., , 2011) ) or a search in the mental lexicon that does not lead to concrete production (failed trial).However, the absence of differences in P2 before-after learning does not necessarily mean that lexical processes do not change with learning.Indeed, firstly, models of word production tend to describe processes as interactive rather than serial (Roelofs, 1997).This implies that lexical processes continue beyond the 200 msec.Some studies have previously shown a P2 peak between 160 and 290 msec (Fairs et al., 2021;Strijkers et al., 2017), with also an extension of lexical selection beyond the P2 in others (Conner et al., 2019).Secondly, the production latencies in the present study are considerably longer than in studies on the temporal signature of word production.The production latency of already known words is generally 800 msec (Fargier & Laganaro, 2020;Indefrey, 2011;Levelt et al., 1999) compared to 1487 msec in the present study.Therefore, the duration of encoding processes may be delayed for newly learned words.Taken together, these two arguments suggest that the changes observed from 300 msec onwards may involve both lexical selection (lexicalesemantic processes) and lexicalephonological word encoding processes.
However, as stressed above, it is noteworthy that the observed changes were not restricted to the learned word list but extended to the unlearned word list.Indeed, the same electrophysiological changes are present on the list of unlearned words, albeit to a lesser extent.In fact, the TANOVA of the unlearned list showed significant differences over a limited period of time between 300 and 400 msec, revealing exactly the same process shift over this time period (Map 5 to Map 6) for the unlearned list.The fitting in the individual ERPs revealed that the changes are less important on the unlearned items than on the learned one.In the following, we will discuss some possible interpretations of the present result.

ERP changes on the unlearned list
Firstly, changes observed on the unlearned list may be the consequence of a simply testeretest situation.Indeed, words from the unlearned list have been seen on Day 1 and on Day 3 and this repetition may lead to electrophysiological changes.An increase in amplitude and topographic changes two days later on the non-learned list could have been the result of the second exposure to the same items, i.e., as a familiarization effect, initiating the onset of learning (Razorenova et al., 2020).This interpretation would fit with the well documented effect that repeated presentation of the same stimuli leads changes in brain activity (Guillaume et al., 2009;Guo et al., 2007;Kelly & Garavan, 2005;Penney et al., 2001;Razorenova et al., 2020).However, previous studies have identified that the testeretest effects seem to disappear when the tests are separated by several days.A recent review of the literature showed that EEG/ERP (and also MEG) measures have good test/ retest reproducibility when administered several days apart in language tasks (Cocquyt et al., 2023).This is actually the case in the present study reducing the plausibility of a test-retest effect.Nevertheless, it is important to note that studies of the testeretest effect have always used frequent and familiar words, whereas the present study used infrequent words.To the best of our knowledge, no study has ever investigated the ERP changes of the testeretest effect with unfamiliar words.Moreover, it is less likely that testeretest changes are limited to specific time windows.But this is something that needs to be the topic of further research.
A plausible second interpretation of these results could be that these effects are related to attentional factors.Attentional factors predominate not only in learning but also in word production (Roelofs & Piai, 2011;Roelofs, 2008).Therefore, it is possible that both the presentation and learning phases changed the way participants directed their attention during the word production task, leading to significant differences between Day 1 and Day 3. One way to investigate this issue is to analyze the differences in pre-stimulus ERPs between Day 1 and Day 3 for the two lists.This allows to determine whether differences exist prior to stimulus presentation, a period that generally involves the recruitment of attentional skills.In fact, both the waveform amplitude analyses and the TANOVA did not highlight different ERPs between Day 1 and 3 (see Appendix Fig. B.2).This suggests that attentional orientation was the same between Day 1 and 3 and c o r t e x 1 7 8 ( 2 0 2 4 ) 1 7 4 e1 8 9 therefore the changes observed on the unlearned list were not necessarily due to attentional factors.
Finally, a last possible interpretation of the differences on the non-learned list may be related to learning effects.Changes are actually induced by learning which lead to modifications in the interconnected network of the language system.The lexicon consists in semantic nodes corresponding to different word features and/or to associate words connected to networked lexical items which are themselves linked to phonological information (Howard et al., 2006;Rogers and McClelland, 2004).Learning a word needs the intergration of a new word in the mental lexicon by making new connections in this system and modifying existing connections (Oppenheim et al., 2010;Rosenzweig et al., 2002).Learning may lead to changes in the semantic network, in particular within the semantic categories concerned.
Surprisingly, in the present study, no behavioral effects were observed between Day 1 and 3 on the unlearned list.In the two sessions, the majority of the productions were noresponses.However, between these two days, words of the same semantic space have been learned.It is therefore likely to hypothesize that the observed effects were caused by a lexicalesemanticephonological reorganization captured with EEG.Even if behaviorally the production corresponded to failed trials, after learning, the search in the mental lexicon involved different activation due to the fact that new words from the same semantic category have been added (Strijkers et al., 2010(Strijkers et al., , 2011)).The integration of new words led to increased complexity in the lexical search and activations within the network, which in turn led to changes in the underlying mental processes when trying (and failing) to produce unlearned rare words.As the lexical selection and consecutive encoding were unsuccessful, the ERP changes observed were consequently smaller than those observed for the learned word list.These results show that word learning involves changes in a wide range of language processes, from lexical to phonological, and affects not only the learned words but also related rare words that have not undergone learning.This effect has been observed here with a word production task and it is worth seeking it in future studies with word recognition/comprehension as well as on frequent words from the same semantic categories and then tested if the changes are the real consequence of a reorganization of the network.

Conclusion
The aim of the present study was to investigate the electrophysiological correlates of word production in a referential task before and after learning on a learned list and the ERP changes on an unlearned list containing words of the same semantic categories.The results show that learning occurs both qualitatively and quantitatively in a large time window between 300 msec and 800 msec following the presentation of the referent.An increase in neuronal firing after learning was observed as well as a processing shift between 300 and 400 msec involving neural activation changes over left temporo-parietal regions.Importantly, these ERP changes are not restricted to the learned list, but were also observed, to a lesser extent, on the unlearned list.This leads to the hypothesis that learning implies spread effects on the lexicalesemanticephonological system and may have an influence on the production encoding of the unlearned word list of the same semantic categories captured by the ERP.This is the first observation of changes in brain processes spilling over to semantically related rare words that were not learned, a result that should be sought in future studies with other language paradigms, such as word recognition and/or comprehension and on semantically related known/frequent words.

Funding
The study was funded by the National Center of Competence in Research "Evolving Language" (grant number 51NF40_180888) of the Swiss National Science Foundation.

Institutional review board statement
The study was approved by the local ethics committee (N PSE.PSE20201003.06).

Informed consent statement
Informed consent was obtained from all subjects involved in the study.

Data and code analysis availability statement
The data are available to download on this link: https://doi.org/10.26037/yareta:qxwracwpefe5pb5ll32wvh7ecq.

Declaration of AI-assisted technologies
During the preparation of this work, the authors used Deepl, pro version in order to correct English grammar and vocabulary.After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Pre-registration
No part of the study procedures and analyses was preregistered prior to the research being conducted.

Open practices
The study in this article has earned Open Data and Open Materials badges for transparent practices.The data and materials are available at: https://doi.org/10.26037/yareta:qxwracwpefe5pb5ll32wvh7ecq.

Fig. 1 e
Fig. 1 e Design representation and variable recoding.

512
Hz with band-pass filters set between .16 and 100 Hz.ERP extraction, averaging and cleaning were performed using the Cartool software(version 3.91;Brunet et al., 2011).First, the EEG signal was filtered with Butterworth filters set at .2e30Hz and notch-filtered at 50 hz.Epochs of 1024 msec aligned to the onset of the picture (100 msec pre-stimulus and 925 msec post-stimulus) were extracted and averaged for each participant and per days separately for the two lists of stimuli.All epochs corresponding to correct, and no-responses were visually inspected and validated only in the absence of motor artefacts (e.g., eyeblinks).The number of epochs for each participant was balanced across sessions to have exactly the same number of epochs in each test session (Day 1; Day 3).The first test session ('Day 1') consisted mainly of no-responses, whereas the second session ('Day 3') contained both noresponses and correct responses (see behavioral results).

Fig. 3 e
Fig. 3 e Results of the ERP analysis on the learned list (a.) and on the non-learned list (b.).Note.(a) Significant differences (TFCE non-parametric p values) on ERP waveform amplitudes analyses on each electrode (Y axis) and time point (X axis).(b) Example of group averaged ERP waveforms for each list on anterior (Cz) and central (Oz) electrodes.(c) Significant differences on GMD values thresholded at 0,05 (TANOVA).(d) Grand-average ERPs (128 electrodes) for each test session from picture onset to 800 ms and temporal distribution of the topographic patterns revealed by the spatiotemporal segmentation analysis.Stable electrophysiological configurations are color-coded.(e) Map templates for the ten stable topographies.Positive (red) and negative (blue) values are displayed.(f) Source localization analysis from 300 to 400 ms.
Table D.1 e contrasts on the Map of the fitting on the Map 5 and 6 for the learned list (Number of TFs ~Test*Map þ (1| Subject), family ¼ nbinom1(link ¼ "log")).Table D.2 e contrasts on the Test of the fitting on the Map 5 and 6 for the learned list (Number of TFs ~Test*Map þ (1| Subject), family ¼ nbinom1(link ¼ "log")).TableD.3 e contrasts on the Map of the fitting on the Map 5 and 6 for the non-learned list (Number of TFs ~Test*Map þ (1|Subject), family ¼ nbinom1(link ¼ "log")).TableD.4 e contrasts on the test of the fitting on the Map 5 and 6 for the non-learned list (Number of TFs ~Test*Map þ (1|Subject), family ¼ nbinom1(link ¼ "log")).TableE.1 e Contrasts of the fitting on the single trials between correctly produced and incorrectly produced words during the Day 3 of the first learned list for each map on the map duration (Number of TFs ~Learning*Map þ (1| Subject) þ (1|Items), family ¼ nbinom1(link ¼ "log")).