Processing Gapped and Gapless Relative Clauses in Mandarin: Evidence from Event–Related Brain Potentials

Abstract

This paper reports an event–related potential (ERP) study examining the processing of “gapless” relative clauses (RCs), normal (“gapped”) RCs, attributive clauses, Subject–Verb–Object (SVO) forms and the morphological markers de and le in Mandarin Chinese. The objectives were to (1) determine whether a processing cost signature can be identified for gap–filler processing by comparing gapped RCs with SVO forms; and (2) determine whether processing gapless RCs is more similar to processing gapped RCs that they resemble on surface or attributive clauses that they resemble structurally by comparing gapped RCs, gapless RCs, and attributive clauses. ERP data was collected from 27 native speakers as they read 160 sentences containing gapped, gapless, attributive, and SVO sentences, which differed only in the second word of the string (e.g., 教授设计的/烧菜的/助手的/设计了/那个广告… The professor designed Ø/cooked food/assistant’s/designed that advertisement…). Results indicate that (1) the head noun evoked larger N400 and P600 components and the determiner elicited a larger sustained anterior negativity for the gapped RCs than for the SVO forms, and (2) the word immediately following the head noun evoked a larger P600 and the determiner elicited a larger sustained anterior negativity for both gapped and gapless RCs than for attributive clauses. These results indicate that (1) gap–filler integration in Mandarin is both semantic (hence the N400) and syntactic (hence the P600) in nature, and (2) gapless RCs are neurophysiologically processed more like gapped RCs than attributive clauses. The ERP signatures to verb+le, verb+de, verbs with verb–object (VO)+de morphological structure, and noun+de revealed that the processing costs for VO verbs and noun modifications were higher than for simple verbs, and the morphological markers de and le induced similar processing costs, even though de triggered the positing of a syntactic gap whereas le did not.

1. Introduction

One of the reasons for conducting event–related brain potential (ERP) research on sentence processing in psycholinguistics is to find neurophysiological evidence for entities that linguistic theory posits but may not be visible on the surface, and in doing so provide information on procedures that underlie the processing of sentences. In this regard, it is helpful to consider the status of syntactic “gaps” in linguistic theory. It is commonplace in linguistics to posit gaps, i.e., entities signified in print by the symbol Ø, presumed to be present in the structure of a phrase even though they are not present on the surface. A simple example of this in English is the occurrence of “Ø close the window”, in which the presence of the second–person pronoun you is posited in an imperative sentence, even though the pronoun you does not overtly appear on the surface.

A more complex example occurs in the case of relative clauses, in which a gap in the clause is posited that is coreferential with the clause head. For example, the English relative clause “the boy who chased the dog” is usually assumed to contain a gap representing the boy between who and chased in the subject position of the embedded sentence “X chased the dog” (i.e., “the boy_i who Ø_i chased the dog”).

While it is clear that the hearer in some sense “knows” that you is the subject of the “close the window” sentence and that the boy is the subject of the embedded “X chased the dog” sentence in the relative clause, the question is whether such gap knowledge can be shown to have an overt, observable manifestation as the hearer processes the sentence in real time. This can be done by comparing the processing of relative clauses with gaps to Subject–Verb–Object (SVO) forms in a language like Mandarin Chinese, which differ on the surface only in the function words attached to the verbs (de versus le), as illustrated in (1) and (2).

Example (1) is a relative clause with a syntactic gap triggered by the modification marker 的 de, a gap which is integrated downstream with the head 广告 advertisement. In Mandarin Chinese, it is a gap–filler integration, which is different from the filler–gap dependency common in left–headed Indo–European languages because the order of the gap and the filler is reversed. In gap–filler integration, what is held in memory for integration is the “gap” rather than the filler. In addition, gap–filler integration in gapped relatives in Mandarin Chinese involves two concurrent processes—the parser recognizes the head noun as the direct object of the RC verb (e.g., design the advertisement) and, at the same time, links the RC to the head noun as its modifier. The latter process is absent in the processing of the head noun in left–headed Indo–European languages because RCs appear after rather than before the head noun. Therefore, we expect that gap–filler processing and filler–gap processing, which has been extensively studied in Indo–European languages (e.g., Felser et al. 2003; Fiebach et al. 2002; Kaan et al. 2000; Phillips et al. 2005) are different in some way.

Example (2) is an SVO sentence, differing from example (1) only in the use of the aspect marker 了 le. The processing effort associated with the placement, holding, and integration of a syntactic gap with its filler (i.e., the head noun) in gapped relatives can be examined by comparing the brain waveforms in response to (1) and (2).

1.	教授	设计的		那个	广告
	professor	sheji de		na ge	guanggao
	the professor	design Øi	de–MOD	that–classifier	advertisement i
	that advertisementi that the professor designed Øi
2.	教授	设计了		那个	广告
	professor	sheji le		na ge	guanggao
	the professor	design	le–ASP	that–classifier	advertisement
	The professor designed that advertisement.
3.	教授	烧菜的		那个	广告
	professor	shao cai de		na ge	guanggao
	the professor	cook food	de–MOD	that–classifier	advertisement
	that advertisement in which the professor cooked food

Gapless RCs, such as (3), contain no missing arguments within the RC verb phrase, can stand alone as independent sentences (e.g., 教授烧菜 the professor cooked food) and express an “aboutness” relationship with the head noun. This type of RC is thus termed a “gapless” relative clause (Beavers and Bender 2004; Chen and Sybesma 2006; Zhang 2008) or “aboutness” relative (Chen and Sybesma 2006). There is no movement posited for an empty operator inside gapless RCs (Adger and Ramchand 2005; see Lin 2018, for more restrictions on the verbs and nouns that can be used in gapless RCs).

There are two ways of analyzing the processing of gapped (1) versus gapless (3) relative clauses. In the first case, processing costs are implied for the gapped form, because the gap (Ø) is held in memory until the head (filler) is reached at the end of the clause, when the gap and filler are integrated, as outlined in (4) (see, e.g., Gibson 1998, 2000). The straightforward implication is that filler–gap processing costs are incurred in the gapped form but not the gapless form, because processing the gapless form by definition does not require holding the contents of a gap in memory (Adger and Ramchand 2005). The second way of viewing potential processing cost differences between the two forms is that the content held in memory in the case of the gapless RCs is the SVO proposition jiaoshou shao cai “professor cook food”, as outlined in (5). Huang et al. (2009) analyze gapless relatives as noun complement clauses, and so the content held in memory in the case of the gapless relative would be the complement representing the SVO proposition, i.e., “professor cook food” in (5). The content held in memory in the case of the gapped RCs is SVØ (Ø is the gap), with that gap Ø serving temporarily as a placeholder until it is integrated with the head (i.e., the filler) at the end of the clause. Given that the subject is the same in (1) and (3), the processing costs involved in the holding and integration of the VØ with the head in gapped RCs versus holding and integrating the VO with the head in gapless RCs can be examined by comparing the brain waveforms in response to the determiner–classifier and the head noun in (1) and (3), where lexical items are identical. It is hard to predict which type of integration is more costly, VØ with the head noun in gapped RCs or VO with the head noun in gapless RCs. VØ–head noun integration may be more costly because the parser must posit a gap, a process that is not required in VO–head noun integration. On the other hand, VO–head noun integration may be more costly since the direct object in VO carries semantics, which is absent in Ø (the empty category). It is also possible that the two kinds of integrations incur similar processing costs.

4.	[[N V Øi] de Ni ]	“[[the professor design Øi] de advertisementi ]”
5.	[[N V O]i de Ni ]	“[[the professor cook food]i de advertisementi ]”

Two previous studies that are relevant to the current study were Zhang (2015) and Tsai (2008). According to Zhang (2015), the gapless structure in Mandarin resembles the regular gapped relative clause structure. Tsai (2008), on the other hand, argued that gapless relatives in Mandarin are not a type of RC, but rather are complex noun phrases that resemble simple noun phrases structurally, with gapless clauses having the RC as their modifiers and noun phrases having attributive phrases as their modifiers. Similar to Tsai (2008), Comrie (1996, 1998, 2002) posits that there are typological differences in noun–modifying clauses in world languages and that Asian languages such as Chinese have only attributive clauses that are structurally different from RC languages such as English.

In Mandarin, gapped RCs, gapless RCs and attributive modifying clauses are structurally similar, because they all are instances of the general modification structure A de B, modifier de head (see examples in Table 1). This property allows us to examine whether processing RCs containing gaps would differ from not only RCs without gaps, but in addition from attributive modification structures which also do not contain gaps. The question, then, is whether the gapless clause is processed like an RC or simply like an attributive clause modifying a head noun. We were able to address this question by comparing processing costs in gapped RCs, gapless RCs and attributive clauses in Mandarin.

The processing of normal (gapped) relative clauses in Mandarin has been extensively investigated but largely centered on the comparison of subject–extracted RCs (SRCs) and object–extracted RCs (ORCs). While the linear distance of the filler–gap dependency in SRCs is shorter than in ORCs in English (e.g., the dog_i that Ø_i chased the cat vs. the dog_i that the cat chased Ø_i), it is longer in Mandarin (e.g., Ø_i追猫的狗_i Ø_i chase–cat–de–dog_i vs. 狗追的猫 dog–chase Øi–de–cat_i). The pre–nominal RCs in Mandarin thus offer an intriguing case for investigating the relative importance of the linear and structural distances between fillers and gaps. Existing psycholinguistic studies on comprehension in Mandarin have yielded contradictory results, with the majority of studies reporting a processing advantage for ORCs over SRCs (Gibson and Wu 2013; Hsiao and Gibson 2003; Lin and Garnsey 2011; Packard et al. 2011; Qiao et al. 2012; Sung et al. 2016) and a smaller number of studies finding an advantage for SRCs over ORCs (Jäger et al. 2015; Lin and Bever 2006). Studies investigating the processing of gapless RCs in Mandarin have yet to be performed. The present study fills that void by comparing the processing of gapped RCs, gapless RCs, and attributive clauses, to shed light on whether gapless RCs are more like gapped RCs or attributive clauses from the perspective of processing cost.

Processing costs in ERP sentence processing research have been found to be associated with the P600 component, a positive waveform that emerges approximately 600 milliseconds after word onset. The P600 originally was associated with garden–path sentences and a range of various syntactic violations, and was believed to reflect the revision of syntactic incongruity (Hagoort et al. 1993; Neville et al. 1991; Osterhout and Holcomb 1992). It was subsequently discovered that the P600 accompanies not just syntactic incongruity, but also processing difficulty. For example, the P600 has been observed when the verb carrying the gap that completes a filler–gap dependency is integrated (Felser et al. 2003; Fiebach et al. 2002; Kaan et al. 2000; Phillips et al. 2005). To illustrate, Kaan et al. (2000) found a larger P600 indicating integration difficulty when processing a verb in a wh– complement (such as the word initiated in 6) compared to the same verb in a “whether” complement (as in 7). The authors argue that the observed P600 effect—the larger P600 in the wh– complement condition—reflects an increased cost of processing the verb in the wh– complement condition, because integrating the verb includes integrating its anaphoric gap with the antecedent filler who, which is held in memory throughout the length of the dependency. Similarly, Phillips et al. (2005) found a larger P600 component on the verb that signals the completion of a wh– dependency (i.e., the verb recognize in 8) than on the same verb (recognize in 9) in a sentence with no dependency, leading the authors to argue that the P600 effect reflected additional processing cost, as integrating the verb in the dependency condition also included integrating the gap associating the verb with its antecedent.

6.	Emily wondered whoi the performer in the concert had imitated Øi (… for the audience’s amusement)
7.	Emily wondered whether the performer in the concert had imitated (… a pop star for the audience’s amusement)
8.	The detective hoped that the lieutenant knew [which accomplice]i the shrewd witness would recognize Øi in the lineup.
9.	The detective hoped that the lieutenant knew that the shrewd witness would recognize the accomplice in the lineup.

The P600 component is sometimes preceded by an anterior negativity (AN), a negative–going deflection that is maximal at anterior scalp sites. Sometimes the negativity is larger over the left hemisphere, so it is sometimes referred to as the left anterior negativity (LAN; Kluender and Kutas 1993). Hagoort (2003) has interpreted the anterior negativity as reflecting automatic detection of morphosyntactic violations during first passes and the P600 as reflecting reanalysis or revision of such violations during later parses. Anterior negativity has been found to index memory costs (e.g., Fiebach et al. 2002; King and Kutas 1995; Kluender and Kutas 1993; Mueller et al. 2005) and predictive processing (Qian and Garnsey 2016).

Another index of processing cost is the N400 component—a negative–going deflection peaking approximately 400 milliseconds after word onset. The N400 has been found in response to semantic aspects of lexical processing, and is sensitive to how easy it is to process the meaning of a word depending on the context in which it appears (Kutas and Hillyard 1980; for a review, see Kutas and Federmeier 2011).

2. Aims and Predictions

The current study took advantage of the N400, P600, and Anterior Negativity ERP components to examine the processing of gapped relatives, gapless relatives, attributive clauses, and SVO forms in Mandarin. Our first goal was to determine if gapped relatives have a greater processing cost than SVO forms, and our second goal was to determine if gapless RCs have an ERP waveform signature that is more similar to the gapped RCs they resemble on the surface or to the attributive clauses they resemble structurally. We predict that gapped RCs are more costly to process than SVO forms because a syntactic gap is posited, held in memory, and integrated with the filler in gapped RCs but not in SVO forms, and expect this effect to be indexed by the P600 component. We also predict that gapped RCs may be more costly to process than attributive clauses for the same reason, and that such an effect may be indexed by the P600 component as well. It is, however, difficult to predict whether gapped RCs are more costly to process than non–gapped forms for the reasons stated earlier. Finally, the amplitude of the anterior negativity component may reflect the cost of holding linguistic elements in working memory, which may differ for gapped, gapless, and attributive clauses.

3. Methods

3.1. Participants

The participants were 27 native Mandarin speakers recruited from a university in Beijing. They had normal or corrected–to–normal vision and were right–handed according to the results of a handedness questionnaire (Li 1983). None of them reported neurological, psychiatric, or cognitive disorders. All of the participants gave their written informed consent and were compensated for their participation. One additional participant was run but excluded from analysis due to excessive motion artifacts.

3.2. Materials and Design

A total of 640 stimulus sentences were constructed, consisting of 160 items, with each item containing four conditions: gapped, gapless, attributive, and SVO, as illustrated in Table 1. All sentences comprised the same structure: Subject + Modifier 1 + Determiner–Classifier (e.g., 那个 that) + Modifier 2 + Head Noun + the rest of the sentence. The critical regions, where the ERP epochs were extracted and analyzed, were Modifier 1, the determiner–classifier phrase, Modifier 2, the head noun, and the segment immediately after the head noun. The segment immediately following the head noun was included to capture potential spill–over effects. All sentences within the same item had the same words in the determiner–classifier phrase, Modifier 2, and the head noun. Gapped, Gapless and Attributive conditions were also identical in all the words following the head noun. The words used in Modifier 1 were verbs carrying gaps in the Gapped condition, verbs carrying no gaps in the Gapless condition, nouns functioning as possessive attributives in the Attributive condition, and verbs with the aspect marker le in the SVO condition. A variety of classifiers were used in the determiner–classifier position to match the semantics of the head noun, but the classifiers were identical in the four conditions within each item. Four head nouns, “案件(case),” “方案(plan),” “精神(spirit),” and 事件(event)” were each used twice, and one head noun, “往事(past),” was used three times in the stimuli, as the construction of gapless sentences necessitated the use of nouns that contained “relatedness” or “aboutness” meanings and there were only a limited number of this type of nouns.

Table 1. Example sentences.

Condition	Subject	Modifier 1	Determiner–Classifier	Modifier 2	Head	Rest of the Sentence
Gapped RC	教授	设计的	那个	独特的	广告	挽救了公司。
	Professor	design Øi MOD	that–classifier	special	advertisementi	saved company
	That special advertisement that the professor designed saved the company.
Gapless RC	教授	烧菜的	那个	独特的	广告	挽救了公司。
	Professor	cook food MOD	that–classifier	special	advertisement	saved company
	That special advertisement about the professor’s food cooking saved the company.
Attributive Clause	教授	助手的	那个	独特的	广告	挽救了公司。
	Professor	assistant MOD	that–classifier	special	advertisement	saved company
	That special advertisement featuring the professor’s assistant saved the company.
SVO	教授	设计了	那个	独特的	广告	以后很高兴。
	Professor	design ASP	that–classifier	special	advertisement	afterwards happy
	The professor designed that special advertisement and afterwards he was happy.

Table 1. Example sentences.

Condition	Subject	Modifier 1	Determiner–Classifier	Modifier 2	Head	Rest of the Sentence
Gapped RC	教授	设计的	那个	独特的	广告	挽救了公司。
	Professor	design Øi MOD	that–classifier	special	advertisementi	saved company
	That special advertisement that the professor designed saved the company.
Gapless RC	教授	烧菜的	那个	独特的	广告	挽救了公司。
	Professor	cook food MOD	that–classifier	special	advertisement	saved company
	That special advertisement about the professor’s food cooking saved the company.
Attributive Clause	教授	助手的	那个	独特的	广告	挽救了公司。
	Professor	assistant MOD	that–classifier	special	advertisement	saved company
	That special advertisement featuring the professor’s assistant saved the company.
SVO	教授	设计了	那个	独特的	广告	以后很高兴。
	Professor	design ASP	that–classifier	special	advertisement	afterwards happy
	The professor designed that special advertisement and afterwards he was happy.

Each segment was presented to the participants at a speed of 800 ms per segment. The segments are shown in Table 1, except that the rest of the sentences were presented in two segments. The modification marker 的 de and the aspect marker 了 le were presented together with the verbs preceding them in Modifier 1, which made it possible to investigate the processing of gaps in the Gapped condition, as the parser could only be certain that a gap was posited when 的 de was encountered (e.g., 教授设计Ø的 the professor designed Ø). The length of each segment was two to three characters. The verbs used in Modifier 1 in the Gapped and SVO conditions were identical. The Gapped, Gapless, and Attributive conditions only differed in the verbs used at Modifier 1, and were identical in all other positions. All stimuli were distributed over four lists using the Latin Square design, such that each participant saw only one condition of each item and all participants saw an equal number of sentences in each condition. An additional 80 sentences of different lengths and syntactic structures were constructed as fillers and added to each list. Each sentence was followed by a yes/no probe question. All participants read 240 sentences in total, and all lists were presented to all participants in the same order.

A norming study was conducted prior to the experiment to check the degree of naturalness of the critical sentences. The 160 items were distributed into four lists according to the Latin Square design, randomized, and rated by native speakers for naturalness on a scale of 1 to 5, with 1 being the least natural and 5 being the most natural. Another 60 sentences with various syntactic structures were added as fillers. The norming study was completed by 40 native speakers of Mandarin, and each list was rated 10 times. The mean naturalness rating was 3.7 for all stimulus sentences, 3.9 for type Gapped, 3.8 for type Attributive, 3.7 for type Gapless and 3.5 for type SVO. The difference between the Gapped and SVO sentences was significant F(1,78) = 9.29, p < 0.01. There were no other statistically significant differences between the ratings of any other two types of sentences (all ps > 0.05). A close inspection of the SVO sentences revealed that they were rated the least natural because they necessarily included two clauses (e.g., Clause 1: 教授设计了那个独特的广告 The professor designed that special advertisement. Clause 2: 以后很高兴 afterwards he was happy) so that the head noun (e.g., 广告 advertisement) was not the final word of the sentence, whereas all other sentence types contained one clause. As shown in Table 1, the segments after the head nouns were identical for the Gapped, Gapless and Attributive conditions, but had to be different for the SVO sentences for them to be grammatical. As a result, the more complicated structure of the SVO sentences caused them to be rated as the least natural among the four. However, since the critical words of the SVO sentences were those that appeared before and included the head noun, the slight unnaturalness (SVO 3.5 vs. Gapped 3.9) should not affect the results.

3.3. Procedure

Participants were tested in a soundproof, electrically shielded booth, seated in a comfortable chair, with a distance of approximately 100 cm between their forehead and the computer screen. Sentences were presented segment–by–segment in the center of the screen in RSVP (Rapid Serial Visual Presentation) format at a rate of 800 milliseconds (ms) per segment using the Presentation software in 24–point–font simplified Chinese characters. All critical sentences were presented in seven segments. Each sentence began with a 400 ms fixation in the middle of the screen, followed by the sequential display of each segment within an 800 ms time frame. Each time frame began with the word appearing for 400 ms, followed by 400 ms of blank screen. The last word of each sentence was followed by 1000 ms of blank screen, followed by the display of the complete probing sentence, which was displayed on the screen for four seconds or until the participant responded with a yes/no. There was a 500 ms display of a blank screen until the fixation appeared, signifying the start of the following sentence.

The experiment began with instructions and a brief practice session, followed by the presentation of experimental sentences. The practice session contained 10 sentences. Participants were told to silently read the experimental sentences before pressing one of two buttons located side by side on a keypad to reply “yes” or “no” to a subsequent probe question that was based on the content of the experimental sentence. The probing questions were incorporated to make sure the participants paid attention to the experimental sentences. The 240 sentences of each list were divided into four blocks. Participants could take a rest between the blocks. The sentences were pseudo–randomized so that no participant saw more than three critical sentences consecutively. The EEG recording session lasted about an hour, and the whole experiment took about two hours.

3.4. EEG Recording and Data Analysis

Continuous EEG was recorded from 64 electrodes (AF7, AF3, FP1, FPz, FP2, AF4, AF8, F7, F5, F3, F1, Fz, F2, F4, F6, F8, FT7, FC5, FC3, FC1, FCz, FC2, FC4, FC6, FT8, T7, C3, C5, C1, Cz, C2, C4, C6, T8, TP9, TP7, CP5, CP3, CP1, CPz, CP2, CP4, CP6, TP8, TP10, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO5, PO3, POz, PO4, PO6, PO8, O1, Oz, O2) attached to an elastic cap (Electro Cap International) worn by the participants, referenced online to the nose tip and re–referenced offline to the average of the left and right mastoids (TP9 and TP10). The EEG signals were amplified using a BrainAmps amplifier with a band pass of 0.05 to 70 Hz that was digitized at 500 Hz. A 0.1 Hz high–pass digital filter was applied offline. The vertical electrooculogram (VEOG) was recorded using electrodes above and below the left eye, and the horizontal electrooculogram (HEOG) was recorded using electrodes located on the cap. The AFz electrode on the cap served as ground. The impedance of the electrodes was kept below 5 kOhm.

The continuous EEG files were epoched over sentence type to isolate 1200–millisecond epochs from 200 ms before the onset of the critical segments. The critical segments that entered statistical analysis were Modifier 1 (e.g., 设计了 designed), Determiner–Classifier (e.g., 那个 that), Modifier 2 (e.g., 独特的 special), Head Noun (广告 advertisement), and the segment immediately after the head noun (e.g., 挽救了 saved). The 200 ms pre–stimulus interval was used for baseline correction. Epochs contaminated with artifacts were discarded, leading to an average loss of 3.3% of the data. On average, the Attributive condition lost 0.5% of the data, the SVO condition lost 1.6%, the Gapless condition lost 4.6%, and the Gapped condition lost 6.3%. Pair–wise comparisons with Bonferroni correction revealed that all pairwise comparisons between any of the two conditions were significant (all ps < 0.05) except for the Gapped versus Gapless comparison.

Mean amplitudes for the four conditions were entered into statistical analysis. Mean amplitudes were calculated for each electrode in each condition for each participant for the time windows intended to capture the N400, P600, and sustained anterior negativity components. Window mean amplitudes were submitted to separate repeated–measures analyses of variance. Pairwise comparisons with Bonferroni correction were conducted when the main effect of condition reached significance. Two sets of analysis were carried out. The first set compared Gapped and SVO conditions, and the second set compared Gapped, Gapless, and Attributive conditions. In each set of analysis, the lateral analysis included all lateral electrodes to capture laterality effects and the midline analysis included just midline electrodes. The ANOVA including lateral electrodes had three within–participant factors: two or three levels of structure (Gapped and SVO in the first set of analysis; Gapped, Gapless, and Attributive in the second set of analysis), five levels of electrode site anteriority (anterior, anterior–central, central, central–posterior, posterior), and two levels of electrode site laterality (left, right). The ANOVA including just the midline electrodes (Fz, FCz, Cz, CPz, Pz, F1, F2, FC1, FC2, C1, C2, CP1, CP2, P1, P2) consisted of the same within–participant factors except that there was no laterality factor. When interactions with electrode site in the omnibus ANOVAs motivated further analysis, analyses were conducted on 15 regions of interest (ROIs), each comprising three electrodes: left anterior (F3, F5, F7), left anterior–central (FC3, FC5, FT7), left central (C3, C5, T7), left central–posterior (CP3, CP5, TP7), left posterior (P3, P5, P7), medial anterior (F1, FZ, F2), medial anterior–central (FC1, FCZ, FC2), medial central (C1, CZ, C2), medial central–posterior (CP1, CPZ, CP2), medial posterior (P1, PZ, P2), right anterior (F4, F6, F8), right anterior–central (FC4, FC6, FT8), right central (C4, C6, T8), right central–posterior (CP4, CP6, TP8), and right posterior (P4, P6, P8). The mean amplitudes over the three electrodes in each ROI were subjected to statistical analysis. The Greenhouse–Geisser correction was applied wherever necessary (Greenhouse and Geisser 1959) to correct for the violation of sphericity. Corrected p–values and original degrees of freedom are reported. For all pairwise comparisons, the p–values were adjusted using the Bonferroni method. Grand average ERPs were digitally low–pass filtered at 10 Hz to smooth the waveforms for display.

4. Results

4.1. Behavioral Data

Accuracy rates were computed for the 27 participants’ true/false answers to the comprehension questions for both the fillers and the stimulus sentences. On average the participants correctly answered 94% of the probes following filler sentences and 91% of the probes following the critical sentences, indicating that they paid attention to the sentences during the experiment. The accuracy rate was 92.9% to Gapped sentences, 91.5% to SVO sentences, 88.9% to Gapless sentences and 88.8% to Attributive sentences. The ANOVA analysis on arcsine–transformed accuracy of the four conditions revealed a main effect of condition (F(3,78) = 6.3, p < 0.001). Subsequent pair–wise comparison with Bonferroni correction showed that the accuracy rate to Gapped sentences was significantly higher than Attributive sentences (p < 0.01) and Gapless sentences (p < 0.05). These results suggested that Attributive sentences and Gapless sentences were harder to comprehend than Gapped structures, although the three structures were rated as equally natural by native speakers (3.9 for Gapped, 3.8 for Attributive, 3.7 for Gapless). The attributive sentences in the present experiment used possessive attributives (e.g., 教授助手的…广告 The professor’s assistant’s advertisement), which may be less common than the gapped relative clauses and therefore may have been harder to comprehend. The gapless structures were two–syllable “verb object” forms (e.g., 烧菜 cook food) and for this reason may have been harder to comprehend than the simple two–syllable verbs (e.g., 设计design) in the gapped structures.

4.2. EEG Data

Two sets of analyses were performed. The first set compared the Gapped and the SVO conditions to examine the processes of holding a gap until the head noun and the integration cost at the head noun. The only surface difference between the Gapped and the SVO conditions was that the verb was followed by the modification marker的 de (e.g., 设计的 designed Ø) in the Gapped condition while the verb was followed by the aspect marker 了 le (e.g., 设计了 designed) in the SVO condition. The second set of analyses compared three conditions—Gapped, Gapless, and Attributive—to determine if gapless RCs have an ERP waveform signature that is more similar to the gapped RCs they resemble on the surface or to the attributive clauses they resemble structurally.

4.2.1. Gapped vs. SVO

At the N400 (350–550 ms) time window of Modifier 1, Determiner–Classifier, and Modifier 2, the ANOVAs that included all lateral electrodes showed no main effect of condition, no interaction between condition and anteriority, and no interaction between condition and laterality (all ps > 0.05). The same results were obtained in the ANOVA for midline channels (all ps > 0.05).

At the head noun, visual inspection of the N400 time window revealed that the Gapped condition was more negative than the SVO condition in the right posterior region, as shown in Figure 1. Statistical analysis confirmed this observation (see Table 2). The ANOVA on lateral electrodes revealed no main effect of condition, no interaction of condition and anteriority, but a significant three–way interaction of condition, anteriority, and laterality F(4, 104) = 4.1, p < 0.01. This three–way interaction resulted because this effect was significant in the right–posterior region F(1, 26) = 4.3, p < 0.05, and marginally significant in the right central–posterior region F(1, 26) = 3.1, p < 0.1 (see the scalp distribution map in Figure 1). The ANOVA over midline electrodes revealed a marginally significant interaction between condition and anteriority F(4, 104) = 3.0, p < 0.1, but none of the midline ROIs showed significant effects of condition (Fs < 1, ps > 0.05). This effect was lateralized to the right, was larger towards to the centro–posterior area, and was significant in the right posterior region (p < 0.05). It had the centro–parietal scalp distribution that is typical for N400, indicating that more processing effort was involved at the head noun position in gapped structures than in SVO structures.

At the head noun, there was a P600–like effect that was more positive for the Gapped condition than the SVO condition, as shown in Figure 1 and Table 2. This effect had a more frontal distribution than the typical P600 effect, starting at around 700 ms and lasted throughout the entire epoch. Analysis of the 700–900 ms time window revealed that, at the midline channels, there was a significant interaction between condition and anteriority F(4, 104) = 3.7, p < 0.05, which resulted because the head noun elicited more positive responses in the Gapped condition relative to the SVO condition, and this effect was significant at the medial–anterior region F(1, 26) = 5.3, p < 0.05, marginally significant at the medial anterior–central region F(1, 26) = 3.0, p < 0.1, medial–central region F(1, 26) = 3.3, p < 0.1, and medial cento–posterior region F(1, 26) = 3.0, p < 0.1. The ANOVA over lateral electrodes revealed similar results. There was a significant interaction between condition and anteriority, F(4, 104) = 3.7, p <0.05, and a marginally significant three–way interaction of condition, anteriority and laterality, F(4, 104) = 2.6, p = 0.05. The significant two–way interaction appeared because this P600–like effect was frontally distributed. Further analysis combining ROIs into anterior, anterior–central, central, central–posterior and posterior regions revealed that this effect was significant at the anterior region F(1, 26) = 4.4, p < 0.05, marginally significant at the anterior–central region F(1, 26) = 3.0, p < 0.1, and central region F(1, 26) = 3.1, p < 0.1.

At the Determiner–Classifier, there was a sustained anterior negativity that was more negative for the Gapped sentences relative to the SVO sentences, as seen in Figure 2 and Table 2. This effect was confirmed by statistical analysis at the 400–800 ms time window. ANOVA at midline electrodes yielded a marginally significant main effect of condition F(1, 26) = 3.4, p < 0.1 and a marginally significant two–way interaction between condition and anteriority F(4, 104) = 2.1, p < 0.1. The effect of condition was significant at medial anterior ROI F(1, 26) = 5.3, p < 0.05 and medial anterior–central ROI F(1, 26) = 5.1, p < 0.05, indicating that this effect had a frontal scalp distribution. Analysis over lateral electrodes revealed that this effect was significant at right anterior–central region F(1, 26) = 5.2, p < 0.05 and marginally significant at left central region F(1, 26) = 3.3, p < 0.1. This anterior negativity effect suggested that the parser was making more effort to hold a verb with a gap in working memory in gapped relatives than a verb without a gap in the SVO sentences.

Table 2. ANOVA F–values for Gapped vs. SVO comparison.

	Determiner–Classifier	Head Noun
	Anterior Negativity	N400	P600
Lateral analysis
Cond (1, 26)	2.3	--	1.6
Cond × Ant (4, 104)	--	2.0	3.7 *
Cond × Lat (4, 104)	--	1.1	4.2 !
Cond × Ant × Lat (4, 104)	1.2	4.1 **	2.6 !
Midline analysis
Cond (1, 26)	3.4 !	--	3.1 !
Cond × Ant (4, 104)	2.1 !	3.0 !	3.5 *
ROI analysis (1, 26)
Left anterior	2.0	--	3.0 !
Medial anterior	5.3 *	--	5.3 *
Right anterior	1.2	--	3.6 !
Left anterior–central	--	--	3.2 !
Medial anterior–central	5.1 *	--	3.0 !
Right anterior–central	5.2 *	--	1.8
Left central	3.3 !	--	4.7 *
Medial central	3.2 !	--	3.3 !
Right central	2.3	--	--
Left central–posterior	--	--	3.2 !
Medial central–posterior	2.4	--	3.0 !
Right central–posterior	1.6	3.1 !	--
Left posterior	1.2	--	2.9 !
Medial posterior	--	--	--
Right posterior	--	4.3 *	--
Combined ROIs (1, 26)
Anterior region	3.1 !	--	4.4 *
Anterior–central region	4.1 !	--	3.0 !
Central region	3.3 !	--	3.1 !
Central–posterior region	1.7	--	1.1
Posterior region	1.2	--	--

Notes: Cond = condition; Ant = anteriority; Lat = laterality; Anterior Negativity = 400–800 ms; N400 = 350–550 ms; P600 = 700–900 ms. ** p < 0.01. * p < 0.05. ! 0.05 < p < 0.1.

Table 2. ANOVA F–values for Gapped vs. SVO comparison.

	Determiner–Classifier	Head Noun
	Anterior Negativity	N400	P600
Lateral analysis
Cond (1, 26)	2.3	--	1.6
Cond × Ant (4, 104)	--	2.0	3.7 *
Cond × Lat (4, 104)	--	1.1	4.2 !
Cond × Ant × Lat (4, 104)	1.2	4.1 **	2.6 !
Midline analysis
Cond (1, 26)	3.4 !	--	3.1 !
Cond × Ant (4, 104)	2.1 !	3.0 !	3.5 *
ROI analysis (1, 26)
Left anterior	2.0	--	3.0 !
Medial anterior	5.3 *	--	5.3 *
Right anterior	1.2	--	3.6 !
Left anterior–central	--	--	3.2 !
Medial anterior–central	5.1 *	--	3.0 !
Right anterior–central	5.2 *	--	1.8
Left central	3.3 !	--	4.7 *
Medial central	3.2 !	--	3.3 !
Right central	2.3	--	--
Left central–posterior	--	--	3.2 !
Medial central–posterior	2.4	--	3.0 !
Right central–posterior	1.6	3.1 !	--
Left posterior	1.2	--	2.9 !
Medial posterior	--	--	--
Right posterior	--	4.3 *	--
Combined ROIs (1, 26)
Anterior region	3.1 !	--	4.4 *
Anterior–central region	4.1 !	--	3.0 !
Central region	3.3 !	--	3.1 !
Central–posterior region	1.7	--	1.1
Posterior region	1.2	--	--

Notes: Cond = condition; Ant = anteriority; Lat = laterality; Anterior Negativity = 400–800 ms; N400 = 350–550 ms; P600 = 700–900 ms. ** p < 0.01. * p < 0.05. ! 0.05 < p < 0.1.

4.2.2. Gapped, Gapless and Attributive

The next set of comparisons focused on the Gapped, Gapless, and Attributive conditions to examine ERP differences between the three types of modifier–head dependency that were of equal distance from the head noun.

At Modifier 1, Modifier 2, and the head noun, no main effect of condition or any interaction involving the condition factor was found in either the midline or the lateral analysis in the N400 and P600 time windows (Fs < 1, ps > 0.05).

At the first segment after the head noun, which was identical across the three conditions, visual inspection revealed that the Gapped and Gapless conditions were more positive in the 600–900 ms time window than the Attributive condition, and this P600 effect had a broad distribution, as shown in Figure 3 and

Table 3. ANOVA F–values for Gapped, Gapless, and Attributive comparisons.

	Determiner–Classifier	First Word after the Head Noun
	Anterior Negativity	P600
Lateral analysis
Cond (2, 52)	4.4 *	4.0 *
Cond × Ant (8, 208)	2.5 *	2.0 !
Cond × Lat (2, 52)	--	--
Cond × Ant × Lat (8, 208)	--	1.2
Midline analysis
Cond (2, 52)	3.1 !	3.8 *
Cond × Ant (8, 208)	1.7	2.7 *
Gapped vs. Attributive (1, 26)
Anterior region	13.3 **	6.1 *
Anterior–central region	11.4 **	6.0 *
Central region	6.1 *	9.0 *
Central–posterior region	2.4	4.9 !
Posterior region	--	2.2
Gapless vs. Attributive (1, 26)
Anterior region	6.5 *	4.2
Anterior–central region	6.1 *	6.8 *
Central region	5.2 !	11.5 **
Central–posterior region	2.2	17.1 **
Posterior region	--	6.8 *
Gapped vs. Gapless (1, 26)
Anterior region	4.5	--
Anterior–central region	4.3	--
Central region	2.3	--
Central–posterior region	2.4	--
Posterior region	--	--

Notes: Cond = condition; Ant = anteriority; Lat = laterality; Anterior Negativity = 400–800 ms; P600 = 600–900 ms. ** p < 0.01. * p < 0.05. ! 0.05 < p < 0.1.

. This observation was confirmed by statistical analysis. There was a significant main effect of condition F(2, 52) = 3.8, p < 0.05 and a significant interaction of condition and anteriority F(8, 208) = 2.7, p < 0.05 in the midline analysis. This effect was significant in medial anterior–central region F(2, 52) = 3.3, p < 0.05, medial–central region F(2, 52) = 5.3, p < 0.01, and medial central–posterior region F(2, 52) = 4.9, p < 0.05. Post hoc pairwise comparisons with Bonferroni correction revealed that the Gapped condition was significantly more positive than Attributive condition at the medial central region F(1, 26) = 7.2, p < 0.05 and the medial central–posterior region F(1, 26) = 5.9, p < 0.05. Similarly, the Gapless condition was significantly more positive than the Attributive condition in medial central region F(1, 26) = 13.9, p < 0.001, medial central–posterior region F(1, 26) = 17.5, p < 0.001, and medial posterior region F(1, 26) = 6.8, p < 0.05. There was no difference between the Gapped and Gapless conditions in any of the midline ROIs (all Fs < 1, all ps > 0.05).

The ANOVAs over lateral electrodes yielded the same results. The main effect of condition was significant F(2, 52) = 4.0, p < 0.05 and the interaction between condition and anteriority was marginally significant F(8, 208) = 2.0, p = 0.05. Subsequent ROIs revealed that there were significant effects of condition in the left anterior–central region F(2, 52) = 4.8, p < 0.05, left central region F(2, 52) = 3.8, p < 0.05, right central region F(2, 52) = 5.5, p < 0.01, and right central–posterior region F(2, 52) = 3.8, p < 0.05. Because there were no significant effects involving the laterality factor, pairwise comparisons were performed on the anterior, anterior central, central, central–posterior, and posterior regions, with each region comprising three ROIs (left, medial right) that included both lateral and midline electrodes. The Gapped condition was significantly more positive than the Attributive condition in the anterior sites F(1, 26) = 6.1, p < 0.05, anterior central sites F(1, 26) = 6.0, p < 0.05, central sites F(1, 26) = 9.0, p < 0.05, and the effect was marginally significant at central–posterior sites F(1, 26) = 4.9, p < 0.1. The Gapless condition was significantly more positive than the Attributive condition at anterior–central sites F(1, 26) = 6.8, p < 0.05, central sites, F(1, 26) = 11.5, p < 0.01, central–posterior sites F(1, 26) = 17.1, p < 0.001, and posterior sites F(1, 26) = 6.8, p < 0.05. There was no difference between the Gapped and Gapless conditions in any of these regions (Fs < 1, ps > 0.05). These results indicated that gapless relatives patterned with gapped relatives, requiring more processing effort than attributive structures.

At Determiner–Classifier, there was a sustained anterior negativity that was more negative for the Gapped and Gapless conditions than for the Attributive condition, as shown in Figure 4 and Table 3. The midline ANOVA over the 400–800 ms time window showed a marginally significant effect of condition F(2, 52) = 3.1, p < 0.1. The ANOVA over lateral electrodes showed a significant effect of condition F(2, 52) = 4.4, p < 0.05 and a significant interaction between condition and anteriority F(8, 208) = 2.5, p < 0.05. The Gapped and Gapless conditions patterned together and were more negative than the Attributive condition. There was no laterality effect. The Gapless condition was significantly more negative than the Attributive condition in the anterior region F(1, 26) = 6.5, p < 0.05 and the anterior–central region F(1, 26) = 6.1, p < 0.05, and was marginally significant in the central region F(1, 26) = 5.2, p < 0.1. Similarly, the Gapped condition was significantly more negative than the Attributive condition in the anterior F(1, 26) = 13.3, p < 0.01, the anterior–central F(1, 26) = 11.4, p < 0.01, and the central regions F(1, 26) = 6.1, p < 0.05. There was no difference between the Gapped and Gapless structures (ps > 0.05).

5. Discussion and Conclusions

The present study conducted an ERP experiment to examine the processing of gapped RCs, gapless RCs, attributive clauses and SVO sentences. Our first objective was to determine whether we could detect an ERP waveform signature reflecting the placement and holding of the gap in memory and the gap–filler integration cost at the head noun by comparing the waveforms of gapped relatives and SVO sentences. An additional objective was to determine if gapless relatives generate an ERP waveform signature that is more similar to the gapped structures that they resemble on surface (Zhang 2015) or to the attributive structures that they resemble structurally (Tsai 2008).

Our findings indicate that there is an ERP waveform signature that reflects both the holding of the gap in memory and the gap–filler integration cost. That waveform was primarily seen in comparing the waveforms of gapped relatives to SVO sentences. Our findings also indicate that gapless relatives generate an ERP waveform signature that is more similar to the gapped structures than to the attributive structures, which we interpret as indicating that the modifier in the gapless relative is a verb–object (VO) structure (the verbs used in the gapless relatives must have a VO compound internal structure) that is held in memory until it is integrated with the head, just as the verb gap is held in memory until it is integrated with the head in the case of the gapped relative.

Our result related to the first objective was a sustained anterior negativity (AN) at the determiner–classifier, and an N400, followed by a frontally distributed P600–like effect at the head noun for the Gapped versus SVO clauses, reflecting the higher storage and integration costs when gaps need to be placed and then filled. Since the only surface difference between the Gapped and SVO conditions was the appearance of 的 de in the Gapped condition and 了 le in the SVO condition, the larger N400 and P600 amplitudes at the head noun illustrate the neurophysiological signatures of the gap–filler integration process, which differs from the filler–gap integration process in left–headed languages.

The N400 effect indicates that in Mandarin, more semantic processing effort is involved in gap–filler integration in gapped relatives than verb–object integration in SVO forms, whereas no such N400 effect is found in filler–gap integration in left–headed languages (e.g., Kaan et al. 2000; Phillips et al. 2005). Gap–filler integration in gapped relatives involves two distinct and concurrent processes—the parser recognizes the head noun as the object of the relative verb (e.g., design the advertisement) and, at the same time, links the RC to the head noun as its modifier. The SVO condition, on the other hand, also involves verb–object integration but not modifier–noun integration, resulting in a smaller N400 for SVOs than for gapped relatives. Because RCs appear before the head noun in right–headed Mandarin Chinese but after the head noun in left–headed languages, filler–gap integration does not occur together with modifier–noun integration at the head noun (see examples 5–8) in English, and thus no N400 effect has been found in the processing of gapped relatives relative to control sentences. This result shows that modifier–noun integration in Mandarin is semantic in nature, as shown by the N400 effect, traditionally an indication of semantic processing.

The larger AN and P600–like effects for the Gapped vs. SVO structures occur because in the case of SVO, the AN reflects the parser waiting for a noun to complete a VO integration, while in the gapped form, the AN reflects the parser waiting for a noun to complete the gap–filler integration, and when the noun appears, the P600–like component reflects the processing cost of completing the VO integration and the gap–filler integration. In the SVO form, it is not a “missing” element but rather an element that is expected based on lexical verb subcategorization, while in the gapped form, the DP was “built” but incomplete (Ø) with the appearance of 的 de. For the gapped form, the speaker must hold in working memory a dependency containing an empty category (Ø), but with the SVO form, the speaker is waiting for the object without holding an empty category in memory. Holding a missing element (i.e., trace) in memory results in a larger AN in the gapped structure than not holding a missing element in memory in the SVO structure. The AN effect occurred at determiner–classifier, which was the word immediately following the relative clause verb that triggered the placement of a gap in gapped relatives. As a result, the larger AN and P600–like effects in the gapped relatives reflect the storage and integration of a gap with the head in the processing of the gapped relatives.

At Modifier 1, which was verb+de (设计Ø的 designed Ø) in the Gapped condition and verb+le (设计了 designed) in the SVO condition, neither the N400 nor the P600 time windows showed any effect of condition, indicating that placing a gap in gapped relatives did not incur more processing cost than not placing a gap in SVO sentences. However, holding the gap in memory was more costly than not holding a gap, as evidenced by the sustained AN effect that was larger for Gapped than SVO forms at the determiner–classifier position.

In summary, the comparison of gapped relatives to SVO sentences reveals that (1) placing a gap after a verb incurs no additional processing cost than processing the verb without having placed a gap at the position of the verb; (2) the modifier–noun integration in gapped RCs is semantic in nature, as evidenced by the N400 effect found in the processing of gapped RCs versus SVO clauses in Mandarin; (3) the sustained anterior negativity effect reflects the cost of holding an empty category in memory; and (4) the P600–like effect indicates that gap–filler integration with an empty category requires more syntactic processing effort than VO integration in SVO sentences.

Our result related to the second objective (to determine if gapless relatives generate an ERP waveform signature that is more similar to gapped structures or attributive structures) is an ERP waveform for the gapless clauses that more closely resembles that of the gapped than the attributive clauses. The gapped, gapless and attributive clauses all complete the modifier–noun integration at the head noun, which is integrated with a verb carrying a gap in the gapped condition, a verb carrying a direct object (the direct object was included in the VO internal structure of the verbs used in Modifier 1) in the gapless condition, and a possessive attributive in the attributive condition. The similarity in the ERP waveforms between gapped and gapless relatives indicates that integrating the VO (the verb and the direct object) in the gapless relatives is as costly as integrating a verb carrying a gap in the gapped relatives, and both are harder to process than the integration of a noun modifying a noun in the attributive clauses. With regard to the typology of noun–modifying clauses, our results indicate that gapped and gapless RCs are more costly to process than attributive clauses, but the elevated processing cost does not necessarily mean that gapped and gapless RCs are typologically different from other noun–modifying attributive clauses.

The ERP waveform signatures are a more negative sustained anterior negativity (AN) at determiner–classifier and a more positive P600 effect at the first word following the head noun for the Gapped and Gapless than for the Attributive conditions. We interpret the similarity in the AN effect as indicating that holding a VO (e.g., cook food) in memory until it is integrated with the head (e.g., advertisement) in the case of gapless relatives is as costly as holding a V+gap (e.g., design + Ø) in memory until gap–filler integration in the case of gapped relatives, and we interpret the similarity in the P600 effect as indicating that V–O–HN (head noun) integration in the case of gapless relatives is as costly as V–gap–HN integration in the case of gapped relatives, with both being more costly than attributive–head noun integration in the case of attributive clauses.

In addition to indexing the holding of linguistic elements in working memory (e.g., Mueller et al. 2005), sustained anterior negativity (AN) has been found to indicate anticipatory processing, with a larger AN being associated with greater effort in trying to predict upcoming words (e.g., Qian and Garnsey 2016). The AN effect in our findings we believe likely reflects both the storage cost and the prediction effort. The AN is more negative at the determiner–classifier for gapped and gapless RCs than for the attributive and SVO sentences, indicating that it is more effortful to predict the upcoming noun in the gapped and gapless RCs than the SVO and attributive sentences. It is possible that gapped and gapless RCs are less frequent than other types of attributives, such as nouns and adjectives, and are certainly less frequent than SVO structures, and thus it is more effortful to predict the upcoming noun in RCs than in attributive and SVO structures.

At Modifier 1, the ANOVAs including all of the four conditions over midline channels and lateral channels revealed a significant main effect of condition at the N400 time window. Pairwise comparisons revealed that the SVO condition was significantly less negative than the attributive condition and the gapless condition.

At the position of Modifier 1, which was verb+le (设计了 designed) in the SVO condition, verb+de (设计Ø的 designed Ø) in the Gapped condition, VO verbs (烧菜的 cook food) in the Gapless condition, and nouns (助手的 assistant’s) in the Attributive condition, the left–lateralized N400 effect for the Attributive and the Gapless conditions versus the SVO condition indicates that the integration costs of VO verbs and noun modifications were higher than for simple verbs, and that this effect was primarily semantic. In the case of the VO verbs, the higher integration costs stem from the fact that VO verbs inherently carry the meanings of both the verb and the direct object, which is semantically denser than the simple verbs in the SVO condition. As for the nouns used at Modifier 1 in the Attributive condition, we suggest the N400 effect was caused by a “stacking” effect of multiple noun modifiers. In the Attributive condition, the parser encounters a noun–noun combination (教授助手的 the professor’s assistant’s) at the position of Modifier 1, whereas in the SVO condition, it was a subject–verb combination (教授设计了 the professor designed), which is a more common structure at the beginning of a sentence than the stacked structure of noun–noun modification. There was a tendency for the Gapped condition to be less negative than the Attributive and the Gapless conditions but more negative than the SVO condition, suggesting that processing verbs that carry gaps is more costly than processing simple verbs but less costly than processing verbs carrying direct objects, with the difference not reaching statistical significance. Note that it was not necessarily that VO verbs were more costly to process than verbs carrying gaps or simple verbs. It was the semantic density of VO verbs that led them to be more costly to process than simple verbs and verbs carrying gaps. There was no P600 effect associated with any of the conditions, suggesting that the difference in the processing effort for these lexical items (simple verb + Ø, VO verb, noun, simple verb) was primarily in the semantic domain. In terms of syntactic processing, there was no additional effort needed to integrate these lexical items into the preceding context.

It appears that the words’ internal—or, morphological, structures—played a role in the amplitude of the N400 component. The effects of the four morphological structures of the lexical items used in the present study suggest that verbs with internal VO structures are more costly to process than simple verbs containing gaps, followed by simple verbs with no gaps. Given that it was the N400 amplitude rather than the P600 amplitude that was affected by the words’ morphological structures, what mattered was semantic density rather than the morphological structure of lexical items that modulated the effort needed to process them. It appeared that there was no influence on the P600 amplitude, which is usually seen as an indication of syntactic processing effort. Therefore, simple verbs with gaps, simple verbs without gaps, VO verbs, and noun modification require a similar amount of syntactic processing effort as long as they are grammatical. In the present study, words used in Modifier 1 were not controlled for frequency and number of character strokes. Future research is needed to further explore this effect while controlling for those lexical properties.

Taken together, we interpret our findings as neurophysiological evidence that gap–filler integration in gapped RCs in Mandarin Chinese is both semantic (hence the N400) and syntactic (hence the P600) in nature, with the P600 reflecting the integration of the gap with the head noun and the N400 reflecting the processing of modifier–noun integration. Placing a gap after a verb incurs no additional processing cost compared to processing a verb without having placed a gap. We interpret our findings that gapped and gapless relatives elicit similar waveforms at the determiner–classifier and the word immediately following the head noun as evidence that gapless RCs are neurophysiologically processed more like gapped RCs, and both are more costly to process than attributives. The determiner–classifier elicits a sustained anterior negativity that is more negative for gapped and gapless RCs than for attributive clauses, indicating that holding a verb and a gap in working memory in gapped relatives is as costly as holding a verb and a direct object in working memory in gapless relatives, and both are more costly than holding a noun–modifier in attributive clauses. The word immediately following the head noun elicits a larger P600 for gapped and gapless relatives than for attributive clauses, suggesting that gap–filler integration in gapped relatives incurs similar processing costs as verb–object integration with the head noun in gapless relatives, and both require more processing effort than the integration of a noun modifying a noun in the attributive clauses.