Time course of Chinese compound word recognition as revealed by ERP data

ABSTRACT Previous studies have yielded conflicting results regarding the onset of semantic processing in compound word recognition. This study examined the role of semantics in morphological processing using event-related potentials (ERP) recorded for Chinese compound targets primed by W+M+, W−M+, W−M− (W = whole-word semantics, M = morpheme meaning, + = congruent, and − = incongruent), semantically related and unrelated primes. Two experiments were conducted. In Experiment 1 of a masked priming lexical-decision task (SOA = 50 ms), EEG results demonstrated that the brain was sensitive to semantic information as early as between 100 and 250 ms. In Experiment 2 of an unmasked priming lexical-decision task (SOA = 200 ms), data confirmed early semantic access. The two EEG experiments also showed that the semantics of constituent morphemes may have little bearing on compound recognition. Overall, these results seem to converge with a form-and-meaning account of compound recognition.


Introduction
A common assumption in word recognition models is that the orthographic form of a word must be processed before its meaning.Regarding morphology, form-thenmeaning models of complex words assume that words with complex morphology are always processed by fast morpheme segmentation based only on orthographic information and that their semantic attributes are activated during the later stages of word recognition (Beyersmann et al., 2016;Beyersmann et al., 2019;Lavric et al., 2012;Rastle & Davis, 2008;Solomyak & Marantz, 2010;Taft, 2004;Taft & Forster, 1975).In contrast, form-and-meaning accounts point that after the orthography of complex words is visually absorbed, the semantic attributes of the whole word may be accessed without the intervention of morphological decomposition and recombination (Feldman et al., 2015;Feldman et al., 2009;Kuperman, 2013;Marelli & Luzzatti, 2012;Schmidtke & Kuperman, 2019).
Although the number of studies examining the neural basis of morphological processing has been on the rise, its time course and underlying brain network remain to be clearly defined (Crepaldi et al., 2019;Leminen et al., 2019;Zwitserlood, 2018).Compared with inflectional and derivative words, the morphological processing mechanism of compounds requires further attention and development given the scarcity and volatility of the results (Leminen et al., 2019).
Many studies have adopted the semantic priming paradigm to explore whether the meaning of individual components is accessed during compound processing.Evidence for the morphological decomposition of compounds comes from the phenomenon that compound words can be facilitated by a certain component (e.g.man-milkman) (Dunabeitia et al., 2009).However, it has been argued that the semantic priming effect between constituents and compounds is not decisive evidence for combinatorial processing, because the priming effect may be driven by pure semantic relevance (Koester et al., 2007).
By strictly controlling the semantic, morphological, and semantic transparency between primes and targets, Marslen-Wilson and Zhou (1996) provided evidence for a non-hierarchical semantically based model of compound word processing.Specifically, they first discovered that transparent compounds promote the processing of their constituents, and vice versa (e.g."bathroom" primed "bath" and "room").However, no priming effect was found between opaque or pseudocompound words and their components (e.g."blackmail" did not prime "black" or "mail", "shamrock" did not prime "sham" or "rock").The above priming result was interpreted as the transparent compound word representation being decomposed and connected by two independent morphemes, while the opaque word was represented as a whole.However, it can also be interpreted that these findings on the priming effect result from the semantic correlation between primes and targets as whole-whole representations.In other words, repetitive processing of the semantic features shared between transparent compounds and their morphemes may lead to a priming effect (Zhou & Marslen-Wilson, 2000a).
Further speculations supporting a semantic rather than a decompositional view come from the discovery that transparent compound words that share morphemes do not prime each other unless the compounds as whole words are semantically related.Thus, although prime-target pairs are all transparent compound words, "teacup" primes "teapot", whereas "headache" does not prime "headscarf".If the compound word is represented by constituent morphemes connected by a morphological relationship, the morpheme representation obtained repeatedly should have a facilitated priming effect on the target word, regardless of the whole-word semantic relationship between primes and targets.In contrast, it is more likely that the semantic correlation between primes and targets as a whole determines the priming effect (Zhou & Marslen-Wilson, 2000a).
Compared with inflection and derivative morphology, compound words comprise a relatively small proportion of complex words in English and other alphabetic languages.However, more than 70% of words in modern Chinese are compounds, and compounding is the only morphological processing mechanism in the Chinese language.Therefore, Chinese provides an opportunity for in-depth study of the morphological processing mechanisms of compounds.For example, the compound word "花粉" (hua(1) fen(3) pollen) is composed of two morphemes "花" and "粉".However, the single morpheme "花" can mean "cost" or "a flower", and "粉" can mean "pink" or "dust".Usually, a Chinese character is used as the constituent morpheme of a compound word, but most characters carry multiple meanings.Some of these meanings are related to each other, and some have no connection (e.g."花").In addition, although compound words can be divided into verbs and nouns according to lexical categories, Chinese compounds adopt more complicated morphological relationships to connect two morphemes.These characteristics may indicate that the meaning of a compound may be difficult to predict using constituent morphemes, as there is no simple superposition between the two morphemes.
Regarding the morphological processing mechanism of Chinese compound words, Zhou and his colleagues have done a lot of related research using masked and unmasked priming lexical decision tasks (Zhou & Marslen-Wilson, 2000a;Zhou et al., 1999;Zhou et al., 2009).Specifically, through Chinese ambiguous morphemes, morphologically and orthographically related primes were set.For example, a target word "华贵" (hua(2) gui(4) luxurious) was preceded by four types of primes.The first type was a morphological prime "华 丽" (hua(2) li(4) magnificent), sharing the same morpheme and whole-word level semantic with the target.The second type of primes "华侨" (hua(2) qiao(2) overseas Chinese) is orthographically and phonologically related to the target, that is, it is irrelevant regardless of the morphemic or whole-word semantics.The third type of homophone "滑翔" (hua(2) xiang(2) glide) is the same only as the phonology of the initial morpheme of the target word.Finally, in the control condition, the same target was primed by unrelated compound "完 整" (wan(2) zheng(3) intact).Participants were asked to perform a lexical judgment task for the target word, and the primes were either forward-pattern masked (SOA = 57 ms) or unmasked (SOA = 200 ms).Zhou et al. (1999) found consistent morphologicalrelated priming effects, regardless of whether the conditions were masked or unmasked.However, there is no priming effect under the homophonic condition, which indicates the limited role of pure phonological information in Chinese lexical access (Wang, Jiang, Huang, et al., 2021;Wang, Jiang, Xu, et al., 2021;Zhou & Marslen-Wilson, 2000b).In addition, the priming effect was found under the homographic-and-homophonic condition of masked priming but disappeared at the long SOA.The author attempted to explain the above results in terms of active and competing morpho-semantics.If the priming word is masked, the ambiguous morphemes contained therein may activate different meanings, thereby facilitating the target word regardless of the morpheme contained in the interpretation.When the prime is presented for a longer time, the correct morpheme in the priming word is strongly activated, whereas the correct morpho-semantic in the target word is inhibited, resulting in no priming effect at the long SOA (Zhou et al., 1999).To further explore the morpho-semantics, the authors also set up a comparative experiment of morphologically and purely semantically related priming conditions, and the results showed that both produced significant priming effects regardless of the masked or unmasked conditions, but a greater priming effect occurred under morphologically related conditions.Repeated activation of the morphemic form and semantic attributes may lead to a greater priming effect of the morphological primes (Zhou et al., 1999).
Based on the above research on English and Chinese compound words, Zhou and Marslen-Wilson (2000a) proposed a lexical representation framework for compound words that is neither purely based on morphemes nor on whole words.The framework assumes that at the form and semantic levels, both for semantically transparent and opaque compounds, the lexical representation of the whole word is based on the representation of its components, overlapping one another in different ways in orthographic and phonological levels, as opposed to the semantic level.At the semantic level, both compound words and their morphemes have representations composed of semantic features, there are many overlaps in the semantic representations of whole words and morphemes (Zhou & Marslen-Wilson, 2000a).According to this framework, the activation of semantic representations of whole words and constituent morphemes may be parallel, rather than hierarchical.The interaction of semantic, orthographic and phonological activation may promote the morphological priming effect of compound words.
Using electrophysiology, the processing of complex words can be tracked before any obvious behavioural response, thereby providing new clues for compound word recognition.This technique provides a high level of temporal resolution, which makes it particularly suitable for tasks that investigate time-course problems, such as the study of the time flow of information in lexical processing (Schmidtke & Kuperman, 2019).There are fewer ERP studies on the morphological processing mechanism of compounds than derivative or inflectional words (Davis et al., 2019;Leminen et al., 2019;Vergara-Martinez et al., 2009).Critically, Davis et al. (2019) suggested that some level of semantic access occurs as early as 100 ms.Specifically, different types of compounds ranging from fully composed semantic transparency (e.g."grapeseed") to partially transparent (e.g."grapefruit") to completely opaque (e.g."hogwash") were set separately.Their findings suggested transparency and word frequency of both the first and second constituents, each uniquely predicting the P100 amplitude.Overall, their results support models that emphasise the simultaneous processing of form and meaning, rather than serial or hierarchical accounts (Davis et al., 2019).Interestingly, electrophysiological research investigating another complex word processingderived word (e.g."dreamer")has also reported the onset of semantic effects as early as 100-250 ms (Cavalli et al., 2016;Jared et al., 2017;Morris et al., 2007;Morris et al., 2013).The aforementioned evidence supports the view that semantic activation may occur from the very beginning of processing complex words and, importantly, has set a baseline for how long we should expect this effect to appear in the brain.
However, some EEG studies appeared to uncover evidence in favour of the form-then-meaning view of compound word recognition.For example, Koester and colleagues found that gender-incongruent first constituents in German compounds recognition induced a LAN effect, independent of the compound's transparency.This seems to indicate that both transparent and opaque compounds are identified by decomposition and that both the first component and head are accessed morpho-syntactically (Koester et al., 2007).In addition, a study in English revealed more negativegoing components elicited by existing and novel compounds compared to monomorphemic words.The authors explain that these findings demonstrate the process of decomposition and combination of existing and novel compounds (Fiorentino et al., 2014).In a next study, using masked priming, Fiorentino et al. (2015) found comparable and significant priming effect for both novel English compounds (drugrack → RACK) and novel pseudo-embedded words (slegrack → RACK).Using overt priming, however, their results showed that the priming effect (in terms of reaction time and N400 amplitude) was greater for the new compounds compared to the novel pseudo-embedded word.All these results seem to support the view that semantic access of whole compounds requires the processes of morphological decomposition and recombination.
The evidence for the form-then-meaning view of morphologically complex word identification seems to come more from derived word studies.Specifically, some studies observed statistically equal priming effects for farmer-farm (semantically transparent morphological relationships) and corner-corn (semantically opaque or pseudo-morphological relationships) in masked lexical decision tasks, which supports the view that the initial morphological processing stages are semantically blind (Beyersmann et al., 2016;Beyersmann et al., 2019;Lehtonen et al., 2011;Longtin et al., 2003;Rastle et al., 2004).For example, a masked-priming EEG study showed a more pronounced N400 attenuation by transparent and opaque primes relative to form primes, but equivalent N400 reduction in the former two cases, implicating a purely structural morphemic segmentation procedure operating in the early stages of morphological processing (Lavric et al., 2007).Even studies using long-term priming methods still did not find semantic priming effects in the early time window 100-250 ms or even the early N400 window (300-380 ms) (Beyersmann et al., 2014;Lavric et al., 2011).Thus, these results seem to be consistent with the twostage model: orthography-based morphological decomposition in the first stage and validation using semantic information in the later stage (Fruchter & Marantz, 2015).
The present study aimed to further explore the time course of semantic processing of Chinese compounds using event-related potentials.Specifically, as in Zhou et al. (1999), Chinese ambiguous morphemes and lexical decision tasks (masked and unmasked priming) were adopted.In addition, like Marslen-Wilson and Zhou (1996), the semantics between the primes and targets are strictly controlled at both the whole-word and morpheme levels.Specifically, the targets (e.g."干 旱/drought") were primed by W+M+ ("干燥/arid", W = whole word semantics, M = morpheme meaning, + = congruent, and − = incongruent), W−M+ ("干洗/drycleaning"), W−M− ("干涉/interfere"), semantically related ("枯萎/withered", semantically but not morphologically or orthographically related) and unrelated primes ("歌曲/song").Thus, the five priming conditions were strictly controlled by the semantic relationship between the whole word and morphemes, and to the best of our knowledge, this is the first study to use these five priming conditions within a single ERP experiment.
Regarding the morphological processing mechanism, most studies adopted the masked priming paradigm (Diependaele et al., 2011;Diependaele et al., 2005;Jared et al., 2017;Lavric et al., 2007;Morris et al., 2013;Morris et al., 2008); however, some studies have used the unmasked long SOA priming method (Beyersmann et al., 2014;Lavric et al., 2011;Rueckl & Aicher, 2008;Smolka & Libben, 2017).Most researchers believe that masked and unmasked priming are qualitatively different (Dehaene et al., 2001;Forster, 2009;Gomez et al., 2013;Grainger, 2008).Using the diffusion model, one experiment directly revealed that masked and unmasked priming involve different cognitive processes: masked priming effects reflect savings in the encoding of the target stimulus, whereas unmasked priming effects reflect the familiarity of the primer-target compound cue (Gomez et al., 2013).In other words, it provides support for examining encoding mechanisms in the early stages of visual word recognition using masked priming techniques.Therefore, masked and unmasked long-term priming techniques should be jointly adopted to explore the morphological processing mechanisms of Chinese compound words.
In the masked priming experiment (SOA = 50 ms), we hypothesised that if semantics affect early morphological processing, then W+M+, W−M+, and W −M− primes would have a significant difference in the priming effect in the 100-250 ms time window, especially the comparison between W+M+ and W−M− in this early time window, significant priming effects may also be observed in purely semantically related conditions.If we follow form-then-meaning accounts, this early semantic influence will not appear.Following these early potentials, we hypothesised that W+M+ will produce the most significant priming effect in the N400 time window, whether because of the morphosemantic effect or the higher similarity in semantics and orthography between whole words (Jared et al., 2017;Kuperman, 2013;Morris et al., 2013).
In the unmasked priming experiment (SOA = 200 ms), if it was indeed semantically blind in the early morphological processing stage, we would expect a significant effect of W+M+ priming but no semantic effects in the early time window.However, if rapid access to semantic information can affect early morphological processing (Davis et al., 2019;Schmidtke & Kuperman, 2019), W+M + and pure semantic priming may equally occur in the early time window.In addition, we will compare the priming effects between the W−M+ and W−M− conditions to verify whether the semantics of morphemes had little bearing on the recognition of compounds the way the findings in English compounds in Marslen-Wilson and Zhou (1996) and Kuperman (2013) did.If the compound words are represented by constituent morphemes that are connected by morphological links, then the repetitive access of morpheme semantics will facilitate target word recognition regardless of the semantic relevance between the whole words (Kuperman, 2013;Zhou & Marslen-Wilson, 2000a).Similarly, in the N400 time window, we predicted that the most significant priming effect would occur in the W+M+ condition.

Experiment 1
In Experiment 1, we investigated whether W+M+, W−M +, and W−M− (the three types of primes are consistent in orthographic and phonological aspects, but only different in semantics) produce equivalent or different priming effects in the early time window.Experiment 1 used masked priming (SOA = 50 ms) in a lexical decision task.

Participants
Thirty native Chinese students (mean age = 23.13years; range = 19-26 years; 13 men) from Tsinghua University participated in this experiment.All participants were right-handed according to the Edinburgh handedness test, and all reported normal or corrected normal vision before the experiment.None of the participants reported a history of neurological or psychiatric impairment.Informed consent was obtained from all the participants in accordance with the Declaration of Helsinki.This study adhered to ethical procedures for the protection of human participants in research and was approved by the ethics committee of the School of Psychology at Tsinghua University.

Materials and design
The experiment contained 110 Chinese compound targets, half of which were real words and the other half pseudo-words used to balance the reaction.The first character of each target was an ambiguous morpheme, for example, "公/gong(1)", which means "public" in "公园/public park" but "male" in "公鸡/ rooster".All ambiguous morphemes in our experiment were selected based strictly on the modern Chinese dictionary to ensure that the different meanings of one morpheme were not related in any way.In addition, like Marslen-Wilson and Zhou (1996) and Zhou et al. (1999), the semantics between the primes and targets are strictly controlled at both the whole-word and morpheme levels.For example, the targets (e.g."面颊 (mian (4) jia(2)/cheek")) were primed by W+M+ (e.g."面庞mian (4) pang(2)/face") (W = whole word semantics, M = morpheme meaning, + = congruent, and − = incongruent), W−M+ (e.g."面试mian(4) shi(4)/interview"), W−M− (e.g."面粉mian(4) fen(3)/flour"), semantically related (e.g."容貌rong(2) mao(4)/appearance") (semantically but not morphologically or orthographically related) and unrelated primes ("环境huan(2) jing(4)/environment").In addition, all ambiguous morphemes in the target words have a dominant meaning, that is, the homomorphic priming condition (W−M−) has a subordinate meaning.All the materials in this experiment were high-frequency words included in the high-frequency vocabulary of the modern Chinese frequency dictionary.In addition, a separate group of 26 participants was paid to evaluate the degree of semantic relevance between all pairs used in this study on a 6-point scale, with 1 reflecting the lowest and 6 the highest relevance.The average score of semantically related word pairs was 4.59, and the average score of W+M+ pairs was 4.57, with no significant difference between them (p > .1).In addition, the semantic relevance of unrelated word pairs, W−M+, and W−M− pairs were 1.07, 1.18, and 1.14, respectively.All the stimuli used in this study are shown in the Appendix.
The entire experiment was divided into five blocks, and each block contained 110 trials, with half being true words and half pseudo-words (filler trials).That is, each block contained 55 pairs of true targets, and each type of prime contained 11 pairs.Consequently, 550 trials were included in this experiment.No target repetition occurred within a block and the order of the blocks and trials in each block was random.

Procedure
All participants sat comfortably in a dimly lit ERP laboratory.During the formal experiment, they were instructed to focus their eyes on the middle of the screen, avoid any physical movement, and relax.All stimuli were presented word by word on an LCD computer screen 80 cm away from the participants.Each trial began with a forward mask (constructed by overlaying five random Chinese characters) appearing for 500 ms, and then the priming word in the PMingLiU font was presented for 50 ms, followed by the target word in the LiSung font presented for 2000 ms until the subject made a key press judgement.Participants were required to judge whether the target was a true word as quickly and accurately as possible.Finally, "-" was presented for 2000 ms to signify the end of a judgment, during which the subject could blink.
Before the formal experiment, a practice session containing 30 priming target pairs was conducted.Participants with an accuracy rate >80% directly participated in the formal experiment.The participants were allowed to rest for 3-5 min after each block.The entire experiment (including electrode preparation) lasted 1.5-2 h in total.

ERP recordings and analyses
EEGs were recorded from 62 Ag/AgCl electrodes in an elastic cap equipped with an International 10-20 electrode placement system (Easycap; Brain Products GmbH, Gilching, Germany).The electrode under the left eye was recorded for the vertical electrooculogram (VEOG), and the electrode at the outer canthus of the right eye was recorded for the horizontal electrooculogram (HEOG).The impedance of all electrodes was kept <10 kΩ.The EEG signal was amplified using the BrainAmpDC amplifier system (Brain Products GmbH), which has a bandpass 0.01-100 Hz and continuous sampling at 500 Hz.
Averaged ERPs time-locked to target words were formed offline from trials free of ocular and muscular artefacts using Brain Vision Analyzer 2.0 (Brain Products GmbH, Gilching, Germany).The records were re-referenced offline to the mean values of the left and right mastoids.The epoch of interest spans from −200 to 600 ms relative to the onset of the target word.Prestimulus 200 ms data were averaged and used as baseline.The analysis focused on the 600 ms epoch after the onset of the target word.All data were filtered with a 30 Hz low-pass filter and a 15 Hz high-cut filter for graph plotting only.The ocular artefacts were first detected by an algorithm implemented in Brain Vision Analyzer 2.0 and then manually rechecked.Epochs exceeding ±80 μV were automatically discarded by artefact rejection.Trials with incorrect answers were excluded from analysis.The overall rejection rate of the 30 participants kept for calculation was 14.09% (W +M+, 13.71%; W-M+, 14.24%; W−M−, 14.39%; semantically related, 14.02%; control, 14.09%).ERP signals (100 −500 ms) were decomposed into 50-ms windows (Wu et al., 2017), which provided finer details on the time course of semantic and morphological activation (Lavric et al., 2011).For each time window, a repeatedmeasures analysis of variance (ANOVA) was conducted with Prime Type (W+M+, W−M+, W−M−, semantically related, and control), hemisphere (left and right), and region (anterior, central, and posterior) as independent variables.Region and hemisphere variables were crossed, forming six lateral regions of interest (ROI), and each region had six representative electrodes (see Figure 1): left anterior (F5, F3, F1, FC5, FC3, and FC1), right anterior (F6, F4, F2, FC6, FC4, and FC2), left central (C5, C3, C1, CP5, CP3, and CP1), right central (C6, C4, C2, CP6, CP4, and CP2), left posterior (P5, P3, P1, PO7, PO3, and O1), and right posterior (P6, P4, P2, PO8, PO4, and O2).The mean ERP magnitude for each ROI was averaged over the electrodes in each region.
To reduce Type I errors caused by violation of sphericity, the Greenhouse-Geisser correction was used to evaluate the effect with more than one degree of freedom in the numerator.In these cases, the original degrees of freedom and corrected probability levels were reported.ANOVA was also used for behavioural data analysis, and the prime type was used as an independent factor (Wu et al., 2017).

Behavioural data
The error rate and reaction time data for the five priming conditions are listed in Table 1.The data were submitted to two ANOVAs with prime type (W+M+, W−M+, W−M−, semantic-related, unrelated) as a within-subject factor.
For reaction times, there was a significant main effect of prime type [F(4,116) = 18.578,P < .001,h 2 p = .390].Further pairwise comparisons indicated that the priming effect of W+M+ was the most significant, and its reaction time was significantly faster than that of the other four conditions (ps < .001).In addition, compared with the unrelated condition, all related priming conditions were significantly faster (W+M+: P < .001;W −M+: P = .014;W−M−: P = .027;semantic related: P = .002).All other comparisons were not statistically significant (p > .05).
On error rates, the ANOVA also yielded a significant effect of prime type [F(4,116) = 4.038, P = .015,h 2 p = .122].Compared with the unrelated control condition, the error rates of the participants under W+M+ (P = .001)and the semantically related condition (P = .019)were significantly lower.In addition, compared with the W−M+ and W−M− conditions, the reaction time under the W +M+ and semantically related conditions both had a significant priming effect (W+M+ vs. W−M+ = .003;W+M+ vs. W−M− = .013;semantic related vs. W−M+ = .019;semantic related vs. W−M− = .032).All other comparisons were not statistically significant (p > .05).

ERP data
The EEG data from 0 ms to 500 ms of the target word were analysed in detail, with an interval of 50 ms, to  detect the earliest time period of semantic activation.In addition, all EEG data were displayed in two stages.
Compared with the unrelated control conditions, the priming effects of the four related conditions were mainly shown in Stage 1.The content of Stage 2 is mainly composed of the difference in the priming effect between the four related conditions.
Stage 1: priming within conditions.As shown in the waveform and topographical maps in Figure 1, the four relevant conditions seem to have a priming effect both in the early and late periods.The detailed data are presented in Table 2. First, in the earliest time period of 100-150 ms, the earliest simultaneous priming effect of purely semantically related primes and morphologically related primes (W+M+ and W−M+) was found.Specifically, in 100−150 ms, the ANOVA yielded a significant main effect of prime type [F(4,116) = 2.685, P = .035,h 2 p = .085],region [F(2,58) = 63.677,P < .001,h 2 p = .687],and hemisphere [F(1,29) = 4.822, P = .036,h 2 p = .143].In addition, the interaction between prime type, hemisphere, and region was significant [F(8,232) = 3.875, P = .002,h 2 p = .118].More detailed calculations found that the purely semantic-related priming effect mainly occurred in the left anterior (P = .037)and left central (P = .032).As shown in Table 2, this purely semantic priming effect lasted for 450 ms.
In 150-200 ms, the four related conditions maintained the priming effect.In 200-250 ms and 250-300 ms, the priming effects of the four relevant conditions begin to differ, and the detailed data will be discussed in Stage 2.
In addition, as shown in Table 3 and Figure 4, the difference in the priming effect between the W−M+ and W−M− conditions never reached significance during the period 100−500 ms.

Discussion
Experiment 1 explored whether semantics affected early morphological processing through a masked priming lexical-decision task (SOA = 50 ms).We will derive this question from behavioural and EEG data.First, in terms of response time, all four priming conditions showed significant facilitating effects.The most significant priming effect was triggered by W +M+, which was even more significant than all other conditions.In terms of the error rate, the W+M+ and purely semantically related priming conditions were significantly lower than all other conditions.In  general, behavioural data showed that W+M+ and purely semantically related primes had advantages over other conditions in terms of reaction time and accuracy.In addition, comparing the two, W+M+ had an absolutely significant priming effect.These data are consistent with those of Zhou et al. (1999), who showed that the response times of MORPH primes (W+M+) and CHAR primes (W−M−) were significantly  shorter than those of unrelated primes in Chinese compound recognition.
EEG data echo behavioural results.All related priming conditions showed priming effects, among which W+M+ was the most significant, followed by the pure semantic condition.More importantly, the EEG data seem to show that there is access to semantics as early as 100 ms, or no more than 200 ms at the latest.Evidence 1 comes from purely semantically related conditions in which the priming effect emerged as early as 100 ms and persisted to 450 ms.Evidence 2 is derived from the difference in priming effect between W+M+ and W−M −, which emerged as early as 200 ms.As there was no difference in orthographic or phonological information between W+M+ and W−M−, this difference in priming effect can only come from semantics.In other words, the effect of semantics on morphological processing may be as early as 200 ms.Thus, the data supported the notion that semantic information may be available at the earliest points of processing (Davis et al., 2019;Kim & Lai, 2012;Lewis & Poeppel, 2014;Schmidtke et al., 2017;Williams et al., 2017;Zauner et al., 2014).
In addition, compared with pure semantic-related conditions, W+M+ had a greater priming effect especially in the 300-450 ms period.Although W+M+ and SEM prime matched very well in terms of their whole-word semantic relevance to the target, the shared morpheme only exists in W+M+ conditions.It can be speculated that repeated access to the formal and semantic representations of key morphemes in W +M+ primes and targets leads to higher activation of the semantic representation of the target word (Zhou et al., 1999).However, there was no significant difference between the W−M+ and W−M− conditions, regardless of the early or late time window, indicating that it may be a whole-word semantic correlation between the prime-target pairs rather than the shared morphosemantics determining the priming effect (Zhou & Marslen-Wilson, 2000a).There is also a similar view that the semantics of morphemes have little bearing on the recognition of compounds (Kuperman, 2013).
Regarding Chinese compound words, Tsang and colleagues have also done masked priming studies to explore early morphological processing mechanism by using ambiguous morphemes (Wu et al., 2017;Wu et al., 2020).There were four conditions.The compound targets (e.g."公园-public garden/park") were primed by morpheme ("公众-public people/the public"), homograph (e.g."公鸡-male chicken/ cock"), semantic-sharing (e.g."草 地-lawn") and unrelated (e.g."嗅觉-olfaction") conditions respectively.First, in their results, semantic-sharing failed to produce reliable masked priming effects regardless of behavioural or EEG data.Most importantly, morpheme and homograph primes produced statistically comparable effects in the 150-250 ms time window (Wu et al., 2017).Additionally, in another masked study (Wu et al., 2020), the results again revealed similar priming strengths for the morpheme and homomorphic conditions in the early time window (200-300 ms).In addition, in the latter study, pure semantic correlation groups were not set, and the authors defaulted that lexical semantic priming would not appear in the masked priming studies.One early behavioural experiments also showed that the facilitation by the morpheme and homomorphic primes was statistically identical in the masked priming procedure (Tsang et al., 2013).These results seem to indicate that morpho-orthographic structure will trigger morphological decomposition in the early recognition process of Chinese compounds.
Overall, EEG studies on morphological processing mechanisms of Chinese compound words are rare and controversial.Moreover, given that the prime-target correlations in Experiment 1 were controlled for the first time at both the whole word and morpheme levels, we intend to explore it again using unmasked experiments.In contrast to masked priming, unmasked or overt priming reflects a later stage of lexical processing (Milin et al., 2018).In addition, to our knowledge, there are no relevant unmasked EEG studies for Chinese morphological processing mechanisms.However, some unmasked EEG studies for alphabetic languages showed that initial morphological processing stages were blind semantically (Beyersmann et al., 2014;Lavric et al., 2011).In Experiment 2, we sought to explore whether semantics affected early morphological processing and whether the morpho-semantic priming effect was significant when primes were present for 200 ms.

Experiment 2
The goal of experiment 2 was to further test whether semantics affected early morphological processing using an unmasked priming paradigm (SOA = 200 ms) (Zhou et al., 1999).Simultaneously, another purpose is to monitor whether constituent morpho-semantics exert an influence on the whole-word recognition of compounds.Some scholars argue that sharing morphemes does not prime each other unless the compounds as whole words are semantically related (Zhou & Marslen-Wilson, 2000a).More importantly, Kuperman (2013) interpreted data on the weak effect of morphosemantics on compound word recognition as evidence against the view of obligatory decomposition.

Participants
Twenty-eight native Chinese students (mean age = 22.86 years; range = 20-26 years; 18 men) from Tsinghua University participated in this experiment.They were native speakers of Mandarin Chinese and were not tested in Experiment 1.

Materials and design
Same as in Experiment 1.

Procedure
Each core structure began with a "+" sign that lasted for 500 ms, and then the priming word in the PMingLiU font was presented for 200 ms, followed by the target word in the LiSung font presented for 2000ms until the subject made a key press judgement.The other procedures were the same as those used in Experiment 1.

ERP recordings and analyses
ERP was recorded from 400 ms before to 600 ms after the onset of the target word (plus a 200 ms pre-prime baseline).Since the primes in this experiment are presented for 200 ms, followed by the targets, the 200 ms before the primes was used as the baseline (Lavric et al., 2011).Other ERP recordings and analyses were the same as those in Experiment 1.

Behavioural data
The error rate and reaction time data for the five priming conditions are listed in Table 4.The data were submitted to two ANOVAs with prime type (W+M+, W−M+, W−M−, semantic-related, unrelated) as a within-subject factor.
For reaction times, there was a significant main effect of prime type [F(4,108) = 29.406,P < .001,h 2 p = .521].Further pairwise comparisons indicated that the priming effect of W+M+ was the most significant, and its reaction time was significantly faster than that of the other four conditions (ps < .001).In addition, the priming effect of the semantically related condition was significant, and its response time was significantly faster than that of the W−M+, W−M−, and unrelated conditions (ps < .005).Compared to the unrelated control group, the subjects' reaction time under the W −M+ condition was also significantly faster (P = .016).All other comparisons were not statistically significant (P > .05).On error rates, the ANOVA also yielded a significant effect of prime type [F(4,108) = 4.657, P = .002,h 2 p = .147].Compared with the unrelated control condition, the error rates of the participants under W+M+ (P = .006)and semantically related conditions (P = .002)were significantly lower.In addition, compared with the W−M+ and W−M− conditions, the reaction time under the W+M+ and semantically related conditions also had a significant priming effect (W+M+ vs. W−M+ = .094;W+M+ vs. W−M− = .026;semantic-related vs. W −M+ = .011;semantic-related vs. W−M− = .010).
Stage 2: comparing the magnitude of priming across conditions.First, it can be seen from Table 6 that the difference between the priming effects of W+M+ and the other three related conditions all reached significance, especially in the 300-450 ms period.
Finally, as shown in Table 6 and Figure 7, the priming difference between W−M+ and W−M− did not reach significance in either the early or late time periods.

Discussion
Experiment 2 examined whether semantics affected early morphological processing through the unmasked priming lexical-decision task (SOA = 200 ms).The results were congruent with Experiment 1, providing further evidence that semantic effects may emerge simultaneously with, or even earlier than, morphological effects (Davis et al., 2019;Jared et al., 2017).
In the behavioural data, there were significant priming effects of the W+M+ and semantically related conditions.In addition, the amplitude of W+M+ primes was significantly greater than that of purely semanticrelated primes, indicating that morphological priming cannot simply be attributed to the semantic primetarget relationship (Zhou et al., 1999).However, no priming effect was observed under orthographic conditions (W−M−), indicating that morphological priming cannot be explained by the orthographic prime-target relationship alone.In addition, the difference between W−M+ and W−M− was not significant, indicating that the morpho-semantics alone seem to have little effect on whole-compound recognition.
First, ERP results showed that compared with unrelated conditions, all related primes elicited a promoting effect, especially in the N400 time window, among the W +M+ condition was still the best.In the early time window of 100-150 ms, pure semantic-related and W +M+, W−M+, and W−M− primes all produced significant priming effects, suggesting simultaneous effects of visual form and semantics (Feldman et al., 2015;Jared et al., 2017).In addition, at 150-200 ms, the influence of semantics on morphological processing was the first to appear.Specifically, the difference between W+M+ and W−M− priming effects becomes significant as early as 150−200 ms and persists to 500 ms.In general, the priming effect of whole-word semantics and the effect of semantics on morphological processing observed in the early time window all support theories of complex word recognition, which posits early simultaneous access of form and meaning (Schmidtke et al., 2017;Schmidtke & Kuperman, 2019).
Another purpose of Experiment 2 was to verify the influence of constituent morpho-semantics on the recognition of the entire compound word.The differences between the priming effects of W−M+ and W−M− were not significant, either in the early (100-300 ms) or late time window (300-500 ms).This result echoed Experiment 1, that is, regardless of whether the priming word appears 50 ms (masked) or 200 ms (unmasked), the separate morpho-semantic seems to have a limited impact on whole compound recognition (Kuperman, 2013;Zhou & Marslen-Wilson, 2000a).

General discussion
Our goal was to survey the hotly contested issue of the time-course of compound recognition.Current visual recognition models of complex words can be divided into two camps based on their assumptions regarding the time course of morphological processing.First, form-then-meaning models describe a process in which the meaning of a complex word is only accessed after its orthographic form is decomposed into constituent morphemes, and the isolated morpheme representations are then recombined (Beyersmann et al., 2016;Lavric et al., 2012;Lehtonen et al., 2011;Rastle & Davis, 2008;Solomyak & Marantz, 2010;Taft & Forster, 1975).In contrast, form-and-meaning accounts assume that the semantic properties of the whole word can be accessed without the intervention of the morphological  decomposition and reorganisation process (Feldman et al., 2015;Grainger & Ziegler, 2011;Jared et al., 2017;Schmidtke & Kuperman, 2019).By strictly controlling the morpho-semantic and whole-word semantic relationships between the prime and target pairs, we evaluated the relative order of formal and semantic effects in the time course of compound processing.Our data support the notion that semantic information may also be available at the earliest points of processing (Davis et al., 2019;Jared et al., 2017;Schmidtke & Kuperman, 2019).
The key to exploring the influence of morphological factors in complex word recognition lies in whether experimental studies involving morphological contrasts can successfully rule out confusion due to the possible overlapping of forms and meanings.In English, this is complicated because morphologically related words almost always overlap in form and meaning (Marslen-Wilson et al., 2008).In this experiment, ambiguous morphemes that are commonly found in Chinese were used to rule out confusion caused by possible formal and meaningful factors.To the best of our knowledge, this is the first study to explore the semantic and morphological processing mechanisms of Chinese compound words by simultaneously controlling semantics at both the whole-word and morpheme levels.The results of our behavioural data are consistent with early research on Chinese and English compounds (Zhou et al., 1999;Zhou & Marslen-Wilson, 2000a).As Zhou et al. (1999) reported, our behavioural data also showed that both W+M+ and W−M− primes had a significant priming effect in the masked experiment (Experiment 1), and W  +M+ continued to maintain this significant priming effect in the unmasked experiment (Experiment 2), whereas W−M− did not.Zhou et al. (1999) explained that when W−M− primes are masked, an alternative semantic representation of homographic morphemes is also activated, thereby facilitating the recognition of the target word.The semantic representations of homographic-homophonic morphemes also compete with each other, and their activation interacts with the semantic activation of the whole word.With an increase in SOA, this kind of competition becomes more intense, and the degree of the priming effect also decreases.Therefore, W−M− induced a facilitatory priming effect in the masked experiment (SOA = 50 ms), but not in the unmasked condition (SOA = 200 ms).Another important point is that we also found that purely semantically related conditions without morphemes showed a significant priming effect in both the masked and unmasked experiments, indicating the early activation of semantics.In addition, W+M+ was more pronounced than pure semantic priming, indicating that morphological priming cannot simply be attributed to semantic priming.
Another important finding in the behavioural data was that the difference between W−M+ and W−M− primes was not significant in either masked or unmasked studies, indicating that morpho-semantics alone may have a weak effect on the recognition of compound words.This result is consistent with the discovery of English compound words, which showed "teacup" primes "teapot" while "headache" does not prime "headscarf" (Zhou & Marslen-Wilson, 2000a).The crucial implication of these results is that the semantic relevance between compound words as a whole may determine the priming effect.Therefore, the hierarchical claim that compound words are represented as individual morphemes connected by morphological relations can be excluded (Zhou & Marslen-Wilson, 2000a).
EEG data reveal the timing of semantic access more accurately than behavioural results.In the masked experiment (SOA = 50 ms), the EEG results demonstrated that in processing complex words, the brain is sensitive to semantic information as early as, or earlier than, 100-250 ms.Specifically, the priming effect of the purely semantically related condition was found to be as early as 100 ms and then maintained at 450 ms.In addition, the difference in priming effect between W +M+ and W−M− reached significance at the earliest 200−250 ms, indicating the early influence of semantics on morphological processing.Importantly, lexical semantic factors have also been found to exert influence within 100−250 ms in other studies (Dien et al., 2003;Hauk & Pulvermuller, 2004;Hauk et al., 2006;Morris et al., 2007;Morris et al., 2013;Penolazzi et al., 2007;Pulvermuller et al., 2001;Segalowitz & Zheng, 2009).Similarly, Davis et al. (2019) also pointed out that there has been a body of work against the form-then-meaning approach, which suggests that the effect of visual form appears in the first 100-250 ms following stimulus presentation, and the semantic effect emerges later, usually in the window of 250-500 ms post-stimulus.
The data of unmasked Experiment 2 (SOA = 200 ms) also supported the notion that semantic information is available at the earliest points of processing.Specifically, the purely semantically related primes and the other three morphologically related primes simultaneously showed significant priming effects in the early time period (100-150 ms, 150-200 ms, and 200-250 ms).In addition, the difference in the priming effect between the W+M+ and W−M− conditions reached significance at the earliest 150-200 ms, which directly indicated the early impact of semantics on morphological processing.Using the same longer SOA paradigms (SOA = 226 ms), Lavric et al. (2011) found that, in the early stage of the N400 range (300-380 ms), equal priming effects were observed under opaque (corner-CORN) and transparent conditions (hunter-HUNT), and a significant difference between the two occurred after 380 ms.Beyersmann et al. (2014) directly compared morphological and semantic priming and observed an obvious morphological but no semantic priming effect within 100-250 ms, indicating that the initial morphological processing stages were blind semantically.Considering that our research goal is to identify Chinese compound words, the time sequence of semantic access may vary according to different languages or morphologies.
Another important finding of the two EEG experiments was that there was no difference in the priming effect between W−M+ and W−M− conditions in the early (100−300 ms) or late (300−500 ms) time windows.In other words, morpho-semantics alone seems to have little bearing on the recognition of compounds.In contrast, the semantic correlation between the prime and target as a whole may play a major role in the priming effect.Similarly, one ERP study did not found any difference in the processing costs produced by between and within-morpheme letter transpositions in English compounds, providing evidence that English compounds may be comprehended via whole-word recognition (Stites et al., 2016).Kuperman (2013) interpreted the weak role of morphemic semantics as evidence against obligatory decomposition and dual-route accounts of morphological processing and in favour of the naive discriminative learning account (Baayen et al., 2011;Pham & Baayen, 2013).In this method, orthographic (and phonological) cues are directly mapped onto meanings, without using morphemes as representation mediators.The weight of the orthographic-semantic connection is the result of statistical learning obtained from the experience of recognising words in context.The model proposes that orthographic clues activate multiple meanings simultaneously, including the meaning of complex words, the meaning of their morphemes, and the meaning of non-morpheme strings that constitute lexical units (Bowers et al., 2005).In addition, selective attention focuses mainly on the meaning of the entire word, and only marginal attention resources are allocated to other co-activated meanings (Pham & Baayen, 2013).The theories of parallel distributed processing hold a similar view, which advocate that morphological information is encoded via patterns of statistical cooccurrences between form and meaning (Gonnerman et al., 2007;Jared et al., 2017;Plaut & Gonnerman, 2000;Rueckl & Raveh, 1999;Seidenberg & Gonnerman, 2000).According to this view, morphology reflects the super-additive effect of shared form and meaning, beyond the effect predicted by words that overlap only with words that are isolated in form or meaning.Because the morpheme is not explicitly represented, there is no need for a discrete morpho-orthographic parsing process (Jared et al., 2017).
By combining our masked and unmasked priming experiments, we repeatedly validated the early activation of semantic information. 1To decide on a string of letters, participants had to encode it first, and only after the process was over were they able to match the acquired perceptual representation with lexical knowledge (Gomez et al., 2013).Masked priming was shown to be compatible with "savings" accounts, where priming occurred primarily because of changes in encoding timing, whereas unmasked priming was consistent with compound cueing accounts, where priming-target correlations improved the quality of lexical information driving the decision-making process (Gomez et al., 2013).Therefore, the semantic priming effect of the early time window found under our masked priming experiment can already tend toward the form-and-meaning view, and the unmasked priming results confirmed this speculation again.Some scholars also use masked and unmasked EEG techniques to explore morphological processing, but semantic effects are only found in the late time window of the unmasked experiment, indicating that semantic constraints are only considered at later stages during morphological processing (Lavric et al., 2007;Lavric et al., 2011;Lavric et al., 2012).Whether the early stages of morphological processing are semantically modulated may also depend on the nature of the morphological analysis system in a particular language (Beyersmann et al., 2019).Most EEG studies on compound word processing mechanisms have focused on Indo-European languages, such as German, English, Italian and Dutch (Arcara et al., 2014;Fiorentino et al., 2014;Kaczer et al., 2015;Koester et al., 2007;Schmidtke & Kuperman, 2019;Stites et al., 2016).These experimental tasks involve lexical decisions, word and picture naming, and familiarity or grammaticality judgments, and the paradigms include long-lag repetition priming, sentence or single word reading, violation (gender, affix or plural), passive-listening oddball, and associative recognition (Leminen et al., 2019).In addition, other factors, such as experimental materials, proportion of related trials, the exposure duration of the prime, type of masking, etc., may also lead to different results (Baayen & Smolka, 2020;Feldman & Milin, 2018).
Regarding Chinese compound word recognition, although there are some inconsistencies in the studies, Tsang and colleagues also found an early unconscious activation of semantic.Specifically, by introducing frequency factors, their behaviour showed that the recognition of target words containing the dominant meaning of ambiguous morphemes was primed by all morpheme-sharing primes, whether it had the same interpretation.In contrast, the target word containing the secondary meaning of ambiguous morphemes was only promoted by primes with the same interpretation (Tsang & Chen, 2013a, 2013b).More precisely, their EEG data suggested that the early N250 was sensitive to morpheme meanings (Wu et al., 2020).In addition, their fMRI study further revealed that the bilateral MTG and left SFG are responsible for morpho-semantic analyses (Zhao et al., 2021).Our current behavioural and ERP data have clearly demonstrated that while morphological overlap alone did not yield any clear and robust priming effect, the effect of semantic plus morphological (W+M+) priming surpassed purely semantic priming effects in early time windows (200-250 ms), suggesting that semantic and morphological information may interact early and jointly exert an additional effect.Thus, the results of our early semantic access may also be consistent with parallel-processing theories, which posit that access to complex words is possible through direct access to orthography and the meaning of the whole word or through a process of morpho-orthographic decomposition (Grainger & Ziegler, 2011;Kuperman et al., 2009;MacGregor & Shtyrov, 2013;Schmidtke & Kuperman, 2019).The present results also supported the multiple-route interactive model, which showed that in the early stages of visual word processing, multiple sources of information regarding words were accessed simultaneously to maximise opportunities for meaning creation (Libben et al., 2020).
In general, our data showed that the semantic information of Chinese compound words may be available at the earliest points of processing. 2 These results run counter to the premises of form-then-meaning accounts but are compatible with a host of theories of form-andmeaning processing.Although these theories differ in some precise details, they all assume that morphological activation is conditioned simultaneously by the form and meaning characteristics of complex words (Davis et al., 2019;Diependaele et al., 2005Diependaele et al., , 2011;;Grainger & Ziegler, 2011;Kuperman et al., 2009;Schmidtke & Kuperman, 2019), or that the activation of complex meanings is not aided by morphological level representations at all (Baayen et al., 2011(Baayen et al., , 2016;;Jared et al., 2017;Kuperman, 2013;Plaut & Gonnerman, 2000).

Conclusion
The present findings address core issues in the time course of semantic access and the nature of morphological processing of Chinese compound recognition.Through masked and unmasked ERP experiments and strictly controlled semantic relevance at both the whole-word and morpheme levels, we found that semantic information is available at the earliest points of processing.Overall, these results support models that emphasise the simultaneous processing of form and meaning rather than serial or hierarchical approaches.
Notes 1.Given that EEG studies related to morphological processing mechanisms in alphabetic languages and Chinese have largely been calculated from 150 ms or 200 ms, this led us to initially ignore the period before 100 ms.We followed a reviewer's suggestion and explored whether significant priming differences would be obtained in the 0-100 ms time window of Exp 1 and Exp 2. In Exp 1, no significant effect was found in the 0-50 ms time window.In the 70-80 ms time window, the results of Exp 1 showed that W−M− released significantly more negative waves compared to W+M+ and W −M+ in the central-anterior regions.For experiment 2, in 0-50 ms, all four related conditions were found to release significantly more negative waves compared to the unrelated condition in the central-anterior regions.
In the left posterior region only, the results showed a significant difference between W−M− and W−M+ (p = .014).In the 50-100 ms time window, the results revealed that the irrelevant condition released significantly more negative waves than the purely semanticrelated condition in the central-anterior regions.
Overall, in both Exp1 and Exp2, the morpho-semantic alone still seems to be weak in 0-100 ms. 2. It is important to note that we are unsure about whether the pre-100-ms effects are real or artifact.So, for future studies we might consider analyzing the entire time course to determine if priming effects in EEG signals occur earlier than previously thought.

Table 1 .
Mean reaction times and error rates for each prime condition (Experiment 1).

Table 2 .
Significance values (ps) of each priming condition against the unrelated baseline in a sequence of 50 ms time windows (Experiment 1).

Table 3 .
Significance values (ps)of comparison between two related priming conditions in a sequence of 50 ms time windows (Experiment 1).

Table 4 .
Mean reaction times and error rates for each prime condition (Experiment 2).

Table 5 .
Significance values (ps) of each priming condition against the unrelated baseline in a sequence of 50 ms time windows (Experiment 2).

Table 6 .
Significance values (ps)of comparison between two related priming conditions in a sequence of 50 ms time windows (Experiment 2).