Processing of non-canonical word orders in (in)felicitous contexts: evidence from event-related brain potentials

ABSTRACT

In many languages with flexible word orders, canonical word order has a processing advantage over non-canonical word orders. This observation suggests that it is more costly for the parser to represent syntactically complex sentences. Alternatively, this phenomenon may relate to pragmatic factors because most previous studies have presented non-canonical word orders without felicitous context, which violates participants’ expectations regarding the information structure. The present study conducted an event-related potential experiment to examine the locus of the processing difficulty associated with non-canonical word orders in Japanese by manipulating word order (SOV vs. OSV) and the givenness of arguments. The results showed that OSV elicited a sustained left anterior negativity from O to S and a P600 effect at S compared to that of SOV in the infelicitous but not in the felicitous context. This result suggests that the processing difficulty of non-canonical word orders in Japanese is alleviated by discourse factors.

KEYWORDS:

• Word order
• givenness
• filler-gap dependency
• Japanese
• event-related potentials

1. Introduction

In real-time sentence comprehension, the parser incrementally constructs various structural dependencies from a string of successive inputs. Among such dependencies, the processing of filler-gap dependency has been extensively examined. Behavioural experiments of many languages with flexible word orders have repeatedly reported that canonical word order has a processing advantage over other possible derived word orders with filler-gap dependency (Bader & Meng, 1999; Kaiser & Trueswell, 2004; Kim, 2012; Koizumi et al., 2014; Mazuka, Itoh, & Kondo, 2002; Sekerina, 1997; Tamaoka et al., 2005; Tamaoka, Kanduboda, & Sakai, 2011). For example, Tamaoka et al. (2005) found that it took more time to judge whether a sentence makes sense in non-canonical object-subject-verb (OSV) sentences than in canonical subject-object-verb (SOV) sentences in Japanese. The processing advantage for canonical word order has also been attested by neurolinguistic evidence, such as fMRI and event-related brain potentials (ERPs) (Fiebach, Schlesewsky, & Friederici, 2001, 2002; Fiebach, Schlesewsky, Lohmann, von Gramon, & Friederici, 2005; Hagiwara, Soshi, Ishihara, & Imanaka, 2007; Kim et al., 2009; Rösler, Pechmann, Streb, Röder, & Hennighausen, 1998; Ueno & Kluender, 2003).

These observations raise the question of why canonical is preferred over non-canonical word orders in sentence comprehension. One possible factor is conceptual accessibility (“the ease with which the mental representation of some potential referent can be activated in or retrieved from memory”, Bock & Warren, 1985, p. 50) (Bornkessel-Schlesewsky & Schlesewsky, 2009a, 2009b; Kemmerer, 2012; Tanaka, Branigan, McLean, & Pickering, 2011). In the languages in which an S precedes an O, a conceptually more accessible agent precedes a conceptually less accessible patient in canonical word orders, whereas the opposite order occurs in non-canonical word orders. Several studies have reported that prominent entities such as an agent, animates, concretes, and prototypicals tend to appear as sentence-initial subjects (cf. Bock & Warren, 1985; Bornkessel-Schlesewsky & Schlesewsky, 2009a; Branigan, Pickering, & Tanaka, 2008; Hirsh-Pasek & Golinkoff, 1996; Primus, 1999; Slobin & Bever, 1982). Accordingly, the preference for canonical SO order may derive from the preference for agent-patient order. However, this hypothesis cannot explain the preference for canonical word orders in languages in which an S follows an O, such as Kaqchikel (a Mayan language spoken in Guatemala) and Truku Seediq (an Austronesian language spoken in Taiwan). Previous behavioural and ERP experiments have found that canonical VOS order incurred a lower processing cost compared to that of non-canonical word orders, such as SVO and VSO in Kaqchikel (Koizumi et al., 2014; Koizumi & Kim, 2016; Yano, Yasunaga, & Koizumi, 2017; Yasunaga, Yano, Yasugi, & Koizumi, 2015). Moreover, an ERP experiment found a larger P600 effect for the non-canonical SVO than the canonical VOS, irrespective of agent-patient order (i.e. voice alternation) in Truku Seediq (Yano, Niikuni, et al., 2017). Therefore, the conceptual accessibility hypothesis does not seem plausible for explaining canonical word-order preference.

Another possibility concerns the syntactic complexities of non-canonical sentences. Since the filler must be associated with its gap position (Frazier & Clifton, 1989), the storage and integration cost should increase in non-canonical sentences (Gibson, 1998; 2000). This hypothesis has been supported by ERP experiments. For example, Ueno and Kluender (2003) compared the canonical SOV sentences in (1a) and the non-canonical OSV sentences in (1b) in Japanese. They found a sustained (bilateral) anterior negativity for “the reckless adventurer-nom” in OSV compared to SOV. Furthermore, OSV elicited a phasic P600 effect at the S (adventurer-NOM).¹ They interpreted their results by assuming that the parser needed to actively maintain an O in the working memory and syntactically integrate it with its original position, reflected by sustained anterior negativity and P600 effects, respectively. An fMRI study conducted by Kim et al. (2009) revealed greater activity at the left inferior frontal gyrus (LIFG) in OVS than in SVO in Japanese. They also took this increased activity as evidence that syntactic complexity due to filler-gap dependency induces a processing load associated with OSV.

Table

(1)	Ano jimotono shinbun-ni yoruto …
	the local newspaper-to according …
	“According to the local newspaper ...”
	a. SOV:
		sono	inochishirazuno	bokenka-ga	toto	sore-o	mitsuketa-ndesu-ka.
		the	reckless	adventurer-nom	finally	that-acc	discovered-pol-q
		“did the reckless adventurer finally discover that?”
	b. OSV:
	sore-oi	sono	inochishirazuno	bokenka-ga	toto	_____i	mitsuketa-ndesu-ka.
	that-acc	the	reckless	adventurer-nom	finally		discovered-pol-q

However, most previous studies did not take discourse factors into account. Canonical word order can be used in a variety of contexts, while non-canonical order is used in limited contexts in which discourse requirements are satisfied. Because previous studies presented non-canonical sentences in isolation, which violated participants’ expectations regarding their information structure, the extent to which the increased processing difficulty can be explained by discourse factors remains unclear.

2. Processing of non-canonical word order in context

The felicitous use of non-canonical word orders has been suggested to correlate with discourse factors, such as givenness, as well as sentence-internal, non-syntactic factors, such as the heaviness of displaced constituents (e.g. Aissen, 1992; Birner & Ward, 2009; Kuno, 1987, inter alia). In other words, canonical word order is a default option for describing an event and occurs in a wide range of contexts, whereas non-canonical word order is a marked choice, and its use must be well motivated. Kuno (1978) claimed that scrambling is motivated by what he called the Information Flow Principle. According to the Information Flow principle, OSV in Japanese is used felicitously when O refers to discourse-older information than S does. Otherwise, SOV should be preferred over OSV. Consistent with this view, a corpus analysis conducted by Imamura (2014) demonstrated that the O of OSV in Japanese was discourse-old information in 81% of OSV occurrences in the corpus (see also Imamura, 2015; Imamura & Koizumi, 2011). Furthermore, in a sentence recall task, Ferreira and Yoshita (2003) observed that native Japanese speakers tend to produce ditransitive sentences in the given-new order when asked to recall those that originally had the new-given order. These observations support that scrambling in Japanese is motivated to create a given-new order in a sentence.

Despite this close correlation, most previous studies on sentence comprehension examined the processing of non-canonical structures without felicitous context, which leads to a confounding of the difficulty of syntactically complex structures and the accommodating of an unsatisfied discourse requirement. This problem was discussed in Kaiser and Trueswell (2004), who conducted a self-paced reading experiment to examine whether the processing difficulty with non-canonical word order relates to discourse factors rather than syntactic complexities in Finnish (see also Clifton C & Frazier, 2004; Grodner, Gibson, & Watson, 2005; Meng, Bader, & Bayer, 1999; Sekerina, 2003). They presented two context types, as shown in (2). The supportive context in (2a) referred to an O of the target sentences in (3b) to license a felicitous use of OVS, in which O must be discourse-old information in Finnish, whereas the non-supportive context in (2b) did not. The result showed no interaction of context and word order at the V (seurasi “followed”). However, they found a significant two-way interaction at the NP2 (“hare-part” and “mouse-nom”), due to a longer reading time in OVS than in SVO only in the non-supportive context.

Table

(2)	Preceding context
	Lotta	etsi	eilen	sieniä	metsässä.	Hän	huomasi
	Lotta	looked-for	yesterday	mushrooms	forest-in	She-nom	noticed
	heinikossa	(a)jäniksen /(b)hiiren	joka	liikkui	varovasti	eteenpäin.
	grass-in	hare-acc/mouse-acc	that	was.moving	carefully	forward.
	“Lotta looked for mushrooms in the forest yesterday. She noticed {(a) a hare /(b) a mouse} moving forward carefully in the grass.”

Table

(3)	a. SVO
	Hiiri	seurasi	jänistä	ja	linnut	lauloivat.
	mouse-nom	followed	hare-part	and	birds	were.singing.
	b. OVS
	Jänistä	seurasi	hiiri	ja	linnut	lauloivat.
	hare-part	followed	mouse-nom	and	birds	were.singing.
	“The mouse followed the hare and birds were singing.”

A similar interaction of context by word order has been observed in Japanese. In Japanese, the canonical word order is SOV. According to transformational syntactic theories, non-canonical OSV sentences involve a filler-gap dependency between the fronted O and an associated gap between S and V (i.e. [O_i [S gap_i V]]). This filler-gap dependency is lacking in canonical SOV sentences. Koizumi and Imamura (2017) ran a self-paced reading experiment using the same factorial manipulation as Kaiser and Trueswell (2004) (i.e. supportive/non-supportive × canonical/non-canonical word order). They observed a significant interaction between word order and context at the NP2. This interaction showed a larger word-order effect in the unsupportive than in the supportive context. At the V, only the main word-order effect was significant, reflecting a longer reading time for OSV.

Table

(4)	Gaimusyoo-no	zikan-wa	(a) Kaneda-da. / (b) Kuroki-da.
	Ministry.of.Foreign.Affairs-gen	vice.minister-top	Kaneda-cop / Kaneda-cop
	“It is (a) Kaneda/(b) Kuroki who is the vice minister of the Ministry of Foreign Affairs.

Table

(5)	a. SOV:
	Kuroki-ga	Kaneda-o	mukaeta	rashii.
	Kuroki-nom	Kaneda-acc	welcomed	is.likely
	“It is likely that Kuroki welcomed Kaneda.”
	b. OSV:
	Kaneda-o	Kuroki-ga	mukaeta	rashii.
	Kaneda-acc	Kuroki-nom	welcomed	is.likely.

These results suggest that the processing difficulty of non-canonical sentences decreased when their discourse requirement was satisfied. However, it is not clear from behavioural experiments how this context effect pertains to the processing difficulty that has been claimed to be associated with long-distance dependency formation, such as the filler storage cost indexed by sustained anterior negativity and the syntactic integration cost indexed by P600 in ERP experiments. If these ERP effects are related to the cost of syntactic dependency formation as has been suggested, we expect them not to be modulated by contextual factors. On the other hand, if they reflect discourse-level processing difficulty, we expect that they would attenuate or disappear in felicitous contexts.

3. Experiment

3.1. Stimuli

The sentences in (6) and (7) show a sample set of experimental context and target sentences, in which two factors are manipulated: Word Order (SOV/OSV) × Givenness (New-Given/Given-New).

Table

(6)	Kooban-ni	(a) Yoshida-san-ga /(b) Kimura-san-ga	imasu.
	police.box-in	Yoshida-Mr-nom / Kimura-Mr-nom	be
	“(a) Mr. Yoshida /(b) Mr. Kimura is in the police box.”

Table

(7)		NP1	ADV1	ADV2	NP2	V	AUX
	a. SOV:	Yoshida-san-ga	kinoo-no	yoru	Kimura-san-o	yurushita	rashii.
		Yoshida-Mr-nom	yesterday-gen	night	Kimura-Mr-acc	forgave	seem
		“It seems that Mr. Yoshida forgave Mr. Kimura last night.”
	b. OSV:	Kimura-san-o	kinoo-no	yoru	Yoshida-san-ga	yurushita	rashii.
		Kimura-Mr-acc	yesterday-gen	night	Yoshida-Mr-nom	forgave	seem

Givenness of arguments was manipulated by presenting an existential sentence, such as in (6), which referred to either the S or O of the target sentences. SOV can be used in a wider range of contexts, allowing given-new or new-given orders. Thus, (7a) did not violate an information order requirement. OSV, on the other hand, is a marked word order; thus, it is used felicitously when O is discourse-given information. The lead-in sentence in (6b) made the OSV in (7b) felicitous because it mentioned O’s referent; thus, the OS order corresponded to the given-new order. In contrast, the lead-in sentence in (6a) did not establish a supportive context for an appropriate use of OSV in (7b).²

The NPs of the target sentences were common family names with no bias for particular thematic roles. Temporal adverbs intervened between the S and the O to increase the memory cost. These NPs and temporal adverbs were used four times across items, so any ERP difference was not due to lexical differences between conditions. Verbs were followed by the modal auxiliary “rashii” (seem) to avoid the wrap-up effect at the V. One hundred twenty sets of experimental stimuli were distributed into four lists, according to a Latin square design, so no participant read more than one sentence from the same set. The lists were counterbalanced across the participants.

3.2. Prediction

The present study is interested in two types of ERP effects, namely, sustained left anterior negativity (SLAN) and P600. SLAN has been observed between the filler and its original position in scrambled sentences (Hagiwara et al., 2007; Matzke, Mai, Nager, Rüsseler, & Münte, 2002; Ueno & Kluender, 2003), wh-questions (Fiebach et al., 2001; Phillips, Kazanina, & Abada, 2005), and post-nominal relative clauses (King & Kutas, 1995; Müller, King, & Kutas, 1997). Previous studies have proposed that it is an index of the working memory load to actively maintain the filler in working memory (Hagiwara et al., 2007; King & Kutas, 1995; Kluender & Kutas, 1993; Matzke et al., 2002; Müller et al., 1997; Phillips et al., 2005). If this is the case, we expect that SLAN would not be modulated by givenness (see Discussion). On the other hand, if it reflects a discourse-level processing cost to accommodate the discourse requirement encoded by non-canonical word orders, we predict that a felicitous context ameliorates it, leading to SLAN’s lack.

P600 has been observed at the gap position of filler-gap dependency and proposed to reflect syntactic processing difficulty to associate a filler with its original position (Kaan, Harris, Gibson, & Holcomb, 2000). We expect no givenness effect if the P600 reflects a syntactic integration difficulty. However, if it relates to a discourse-level processing cost, we predict a P600 in the infelicitous but not in the felicitous context.

In addition to these ERP effects, we expect a reduced N400 effect for given NPs at NP1 (S of SOV and O of OSV) and NP2 (O of SOV and S of OSV) since N400 has been known to be sensitive to priming effect (Kutas & Federmeier, 2011; Kutas & Van Petten, 1988). Although it forms part of the result of statistical analyses reported below, this effect is of no interest for the present purpose.

3.3. Procedure

Stimuli were presented in the centre of the monitor in random order, using Presentation ver. 17.0. (Neurobehavioral Systems). At the beginning of each trial, a fixation was presented for 1000 ms, followed by a blank screen for 300 ms. A lead-in context in (6) was presented in its entirety for 2000 ms with an inter-stimulus interval (ISI) of 200 ms. After that, each phrase of the target sentences was presented for 700 ms with 200 ms ISI. A comprehension task was administered at the end of each trial to check whether our participants understood sentences correctly. Participants were required to answer questions (e.g. Is it Mr. Yoshida who forgave Mr. Kimura?), by pressing the “YES” or “NO” button on the response pad (Cedrus, RB-740). Prior to the main experiment, twelve practice trials were completed to familiarise participants with the experimental procedure.

3.4. Electrophysiological recording

EEGs were recorded from 19 Ag electrodes (QuickAmp, Brain Products) located at Fp1/2, F3/4, C3/4, P3/4, O1/2, F7/8, T7/8, P7/8, Fz, Cz, and Pz according to the international 10–20 system (Jasper, 1958). Additional electrodes were placed below and to the left of the left eye to monitor horizontal and vertical eye movements. The online reference was set to the average of all electrodes and EEGs were re-referenced offline to the average value of the earlobes. The impedances of all electrodes were maintained at less than 10 kΩ throughout the experiment. The EEGs were amplified with a bandpass of DC to 200 Hz, digitised at 1000 Hz.

3.5. Electrophysiological data analysis

Trials with large artefacts (exceeding ±80 µV) were automatically removed from the analysis. Two types of analyses were conducted following previous studies on the processing of filler-gap dependency, namely, cumulative multi-word and single-word analyses (Fiebach et al., 2001; King & Kutas, 1995; Phillips et al., 2005; Ueno & Kluender, 2003, 2009). The cumulative multi-word analysis examined SLAN from NP1 to NP2. The baseline was set to 100 ms prior to the onset of NP1. The SLAN was expected to appear after lexical access to NP1 was completed (i.e. approximately 300–500 ms). Hence, for NP1, SLAN’s presence was examined to compare the mean amplitude of 500–900 ms. The time-window of 300–500 ms was also tested to examine a priming effect, although it is not of interest. For the following two adverbs, the SLAN was assessed by calculating the mean amplitude from 100 ms after the onset of each region to the end of the epoch (100–900 ms) (cf. Lau & Liao, 2018; Phillips et al., 2005).

The single-word analysis examined a P600 at NP2, which has been associated with the integration cost. The V region was also examined because some previous studies reported a P600 for non-canonical word orders. The baseline was set to 100 ms prior to the onset of each phrase. The ERPs were quantified by calculating the mean amplitude for each participant relative to the baseline using three time windows: 300–500 ms, 500–700 ms, and 700–900 ms. All EEGs were filtered offline using a 10 Hz low-pass filter for presentation purposes.

All statistical analyses were conducted separately at the midline (Fz, Cz, and Pz), lateral (F3/4, C3/4, and P3/4), and temporal (Fp1/2, F7/8, T7/8, P7/8, and O1/2) arrays. The midline analysis consisted of repeated measures ANOVAs with three within-group factors: WORD ORDER (WO) (SOV/OSV) × GIVENNESS (Given-New/New-Given) × ANTERIORITY. The lateral and temporal analyses involved four within-group factors: WO × GIVENNESS × HEMISPHERE (left/right) × ANTERIORITY. When an interaction occurred between WO × GIVENNESS, post hoc analyses were conducted to examine the effect of WO at each level of GIVENNESS and that of GIVENNESS at each level of WO. When WO and/or GIVENNESS interacted with topographic factors (ANTERIORITY/HEMISPHERE), post hoc analyses were conducted at each level of topographic factors (e.g. front, central, and posterior). The Greenhouse-Geisser correction was applied for all effects involving more than one degree of freedom (Greenhouse & Geisser, 1959). In these cases, the original degrees of freedom and the corrected p-value were reported.

3.6. Participants

Sixteen native Japanese speakers were recruited from Tohoku University (five females and 11 males, M = 20.6, SD = 1.6, range: 19.2–24.3). All participants were classified as right-handed based on the Edinburgh handedness inventory (Oldfield, 1971), and three of them had a left-handed family member. All participants had normal or corrected-to-normal vision and no history of reading disability or neurological disorders. This study was approved by the Ethics Committee of the Graduate School of Arts and Letters, Tohoku University. Written informed consent was obtained from all participants prior to the experiment, and they were paid for their participation.

3.7. Results

3.7.1. Behavioural data

The mean accuracy of the comprehension question task was 87% (S_NEWO_GIVENV: 87.9%, S_GIVENO_NEWV: 87.5%, O_NEWS_GIVENV: 85.2%, O_GIVENS_NEWV: 87.5%). The repeated-measures ANOVA showed no significant main effect or interaction in subject and item analyses (all ps > 0.10).

3.7.2. Electrophysiological data

Multi-word cumulative analysis

Figure 1 shows the grand average ERP from the onset of NP1 to that of NP2 of the target sentence. Visual inspection of the graph suggests a striking difference between OSV at the left frontal sites, with a larger negativity for O_NEWS_GIVENV.

Figure 1. Grand average ERPs from NP1 to NP2. (Boldface in the legend indicates discourse-given NPs).

In the 300–500 ms time-window of NP1, the main effect of GIVENNESS was significant at all arrays, showing a larger N400 for the new NPs compared to the given NPs, due to priming effect (Table 1).

Table 1. Statistical results of the cumulative analysis.

	NP1: 300–500 ms			NP1: 500–900 ms			Adv1: 100–900 ms			Adv2: 100–900 ms
	Midline	Lateral	Temporal	Midline	Lateral	Temporal	Midline	Lateral	Temporal	Midline	Lateral	Temporal
Word Order (WO)	1.45	1.11	0.66	0.82	0.83	0.01	0.52	1.14	0.65	0.36	0.16	0.02
Givenness (G)	32.31***	33.56***	50.74***	2.67	3.69	3.35^+	7.35*	6.95*	5.65*	6.64*	6.12*	3.28^+
WO × Anteriority (Ant)	0.02	0.26	0.34	0.30	1.81	1.55	0.06	1.11	0.74	0.04	1.43	0.86
G × Ant	0.14	0.21	0.53	4.03^+	10.88***	3.48^+	1.34	2.37	0.31	1.45	2.46	0.45
WO × G	1.25	0.98	1.68	0.40	0.29	1.79	0.55	0.75	0.30	1.01	1.32	0.01
WO × G × Ant	0.67	0.03	0.23	2.19	0.56	2.79^+	1.25	0.65	5.38*	0.28	0.54	1.58
WO × Hemisphere (Hem)		0.88	0.88		0.06	1.37		<0.01	3.09^+		0.01	3.57
G × Hem		0.04	1.50		0.21	1.45		0.73	1.28		1.15	0.49
WO × Ant × Hem		2.34	1.35		1.20	1.05		1.19	0.39		0.89	0.54
G × Ant × Hem		0.06	0.56		1.44	0.47		2.48	0.63		2.08	1.08
WO × G × Hem		0.27	0.61		0.26	<0.01		1.07	0.03		0.11	<0.01
WO × G × Ant × Hem		0.81	0.18		0.50	0.11		0.51	1.08		0.75	0.77

⁺p < .10, *p < .05, **p < .01, ***p < .005.

In the 500–900 ms time-window of NP1, the interaction of WO × GIVENNESS × ANTERIORITY was marginally significant at the temporal array. The post hoc analyses revealed that O_NEWS_GIVENV showed a larger anterior negativity than S_NEWO_GIVENV (Fp1/2: F(1, 15) = 15.41, p < 0.01; F7/8: F(1, 15) = 6.82, p < 0.05), whereas O_GIVENS_NEWV did not show a negativity compared to S_GIVENO_NEWV. Furthermore, O_NEWS_GIVENV showed a larger anterior negativity than O_GIVENS_NEWV (Fp1/2: F(1, 15) = 20.91, p < 0.01; F7/8: F(1, 15) = 15.36, p < 0.01), whereas the two SOV conditions did not differ. The interaction of GIVENNESS × ANTERIORITY was also significant at the lateral array and marginally significant at the midline and temporal arrays, due to a GIVENNESS effect at the frontal sites (Fz: F(1, 15) = 10.97, p < 0.01; F3/4: F(1, 15) = 9.83, p < 0.01; C3/4: F(1, 15) = 4.29, p = 0.05; Fp1/2: F(1, 15) = 6.97, p < 0.05; Fp7/8: F(1, 15) = 10.19, p < 0.01).

At ADV1, the three-way interaction of WO × GIVENNESS × ANTERIORITY was significant at the temporal array. Post hoc analyses revealed that the anterior negativity continued to ADV1 in the O_NEWS_GIVENV; O_NEWS_GIVENV showed a larger anterior negativity than S_NEWO_GIVENV (Fp1/2: F(1, 15) = 3.73, p = 0.07), whereas O_GIVENS_NEWV did not. Furthermore, O_NEWS_GIVENV showed a larger anterior negativity than O_GIVENS_NEWV (Fp1/2: F(1, 15) = 11.57, p < 0.01). The main effect of GIVENNESS was also significant, with a larger negativity for the new-given conditions compared to that of the given-new conditions.

At ADV2, the only significant effect was the main effect of GIVENNESS, showing a larger negativity for the new-given conditions than the given-new conditions.

3.7.3. Single-word analyses

The NP2

Figure 2 shows the grand average ERP of NP2 of the target sentence. Visual inspection of the graph suggests that O_NEWS_GIVENV showed a posterior positivity, but O_GIVENS_NEWV did not.

Figure 2. Grand average ERPs at NP2.

In 300–500 ms, the main effect of GIVENNESS was significant in all arrays, due to attenuated N400 for the given NPs compared to the new NPs (Table 2).³ The effect of GIVENNESS interacted with ANTERIORITY at the temporal array, due to a significant GIVENNESS effect except Fp1/2 (F7/8: F(1, 15) = 6.66, p < 0.05; T7/8: F(1, 15) = 9.95, p < 0.01; P7/8: F(1, 15) = 13.16, p < 0.01; O1/2: F(1, 15) = 11.10, p < 0.01). The effect of WO × GIVENNESS × ANTERIORITY was marginal, which reflected a significant GIVENNESS effect only at OSV (P7/8: F(1, 15) = 16.14, p < 0.01; O1/2: F(1, 15) = 13.68, p < 0.01) and a significant WO effect only at the new-given condition (P7/8: F(1, 15) = 6.20, p < 0.05; O1/2: F(1, 15) = 6.83, p < 0.05). These results suggest O_NEWS_GIVENV showed an early positivity compared to O_GIVENS_NEWV and S_NEWO_GIVENV.

Table 2. Statistical results for NP2.

	300–500 ms			500–700 ms			700–900 ms
	Midline	Lateral	Temporal	Midline	Lateral	Temporal	Midline	Lateral	Temporal
Word Order (WO)	0.12	0.15	0.28	0.38	0.25	0.51	0.31	0.53	0.12
Givenness (G)	25.48***	25.25***	9.25**	1.39	1.33	0.01	0.46	0.30	1.11
WO × Anteriority (Ant)	3.58	5.12*	3.19^+	1.07	1.41	0.53	0.10	0.28	0.15
G × Ant	0.65	1.07	3.61*	0.16	0.20	0.20	0.21	0.37	0.40
WO × G	2.18	2.10	1.72	3.31^+	2.95	1.80	1.21	1.18	<0.01
WO × G × Ant	2.00	3.23^+	2.96^+	3.61^+	4.41*	5.37*	0.48	1.69	3.58*
WO × Hemisphere (Hem)		0.01	0.90		1.64	2.33		0.02	0.20
G × Hem		0.22	1.62		0.36	2.65		0.83	2.43
WO × Ant × Hem		0.08	1.14		0.30	1.33		0.13	0.19
G × Ant × Hem		0.37	1.88		0.22	1.87		0.17	1.60
WO × G × Hem		1.24	0.53		1.18	<0.01		1.10	1.11
WO × G × Ant × Hem		0.02	0.57		0.07	0.17		0.35	0.76

⁺p < .10, *p < .05, **p < .01, ***p < .005.

In 500–700 ms, the interaction of WO × GIVENNESS × ANTERIORITY was significant at the lateral and temporal arrays and marginally significant at the midline array. Post hoc analyses showed a significant WO effect only at the new-given condition at the posterior sites (Pz: F(1, 15) = 5.67, p < 0.05; P3/4: F(1, 15) = 5.26, p < 0.05; P7/8: F(1, 15) = 5.30, p < 0.05; O1/2: F(1, 15) = 7.19, p < 0.05). At OSV, the effect of GIVENNESS was significant or marginally significant at the posterior sites (Pz: F(1, 15) = 5.44, p < 0.05; P3/4: F(1, 15) = 5.62, p < 0.05; O1/2: F(1, 15) = 4.02, p = 0.06). These results indicate that O_NEWS_GIVENV showed a posterior positivity compared to S_NEWO_GIVENV and O_GIVENS_NEWV.

In sum, a robust N400 reduction was observed for the given NPs at 300–500 ms. Importantly, O_NEWS_GIVENV elicited an early larger positivity than S_NEWO_GIVENV, whereas O_GIVENS_NEWV did not elicit any effect compared to S_GIVENO_NEWV.

The verb

Figure 3 shows the grand average ERP of V of the target sentence.

Figure 3. Grand average ERPs at V.

In 300–500 ms, a significant main effect of GIVENNESS was observed in the midline and lateral arrays, showing that the new-given order enhanced an N400 amplitude compared to the given-new order (Table 3). GIVENNESS interacted with ANTERIORITY in all arrays, reflecting that the negativity for the new-given conditions was distributed at the fronto-central sites (Fz: F(1, 15) = 9.90, p < 0.01; Cz: F(1, 15) = 7.40, p < 0.05; F3/4: F(1, 15) = 9.04, p < 0.01; C3/4: F(1, 15) = 4.46, p < 0.05; F7/8: F(1, 15) = 7.08, p < 0.05; T7/8: F(1, 15) = 6.91, p < 0.05).

Table 3. Statistical results for V.

	300–500 ms			500–700 ms			700–900 ms
	Midline	Lateral	Temporal	Midline	Lateral	Temporal	Midline	Lateral	Temporal
Word Order (WO)	0.99	0.97	0.69	0.96	1.11	0.65	3.77^+	4.45^+	3.40^+
Givenness (G)	6.16*	4.80*	2.98	4.04^+	2.96	2.13	2.40	1.15	1.50
WO × Anteriority (Ant)	0.45	0.80	0.20	1.00	2.62	0.84	0.99	4.06*	1.55
G × Ant	6.08*	6.36**	3.19*	3.50^+	4.12*	0.92	4.95*	8.52**	2.27
WO × G	0.05	0.02	<0.01	0.11	0.02	<0.01	1.65	1.29	1.04
WO × G × Ant	1.00	0.43	0.02	0.75	0.02	0.21	0.85	0.27	0.15
WO × Hemisphere (Hem)		<0.01	0.93		0.09	1.05		0.88	0.80
G × Hem		0.81	0.91		0.02	0.68		0.10	0.17
WO × Ant × Hem		1.19	0.53		1.38	0.68		1.07	1.36
G × Ant × Hem		0.04	2.70*		0.07	1.87		0.03	1.50
WO × G × Hem		0.13	1.09		1.29	<0.01		0.59	<0.01
WO × G × Ant × Hem		0.19	1.67		0.21	1.36		0.05	1.12

⁺p < .10, *p < .05, **p < .01, ***p < .005.

The interaction of GIVENNESS × ANTERIORITY was significant at the lateral array and marginally significant at the midline array at the 500–700 time-window, due to a GIVENNESS effect at the fronto-central sites (Fz: F(1, 15) = 3.76, p = 0.07; Cz: F(1, 15) = 6.07, p < 0.05; F3/4: F(1, 15) = 3.91, p = 0.06).

In 700–900 ms, the main effect of WO was marginally significant, showing that OSV elicited a positivity compared to SOV. For the same reason, the interaction of WO and ANTERIORITY was significant at the lateral array, with a significant effect at P3/4 and a marginally significant effect at C3/4 (P3/4: F(1, 15) = 7.67, p < 0.05; C3/4: F(1, 15) = 4.06, p = 0.06). WO did not interact with GIVENNESS in any array. However, the planned comparison between SOV and OSV conditions showed a positivity in the new-given condition (Midline: F(1, 15) = 4.80, p < 0.05; Lateral: F(1, 15) = 6.03, p < 0.05; Temporal: Lateral: F(1, 15) = 4.72, p < 0.05), but not in the given-new condition (Midline: F(1, 15) = 0.41, p > 0.10; Lateral: F(1, 15) = 0.45, p > 0.10; Temporal: F(1, 15) = 0.15, p > 0.10).

Overall, the new-given conditions (S_NEWO_GIVENV and O_NEWS_GIVENV) elicited a larger N400 effect compared to the given-new conditions (S_GIVENO_NEWV and O_GIVENS_NEWV). O_NEWS_GIVENV exhibited a posterior positivity compared to S_NEWO_GIVENV at the late time-window.

4. Discussion

The present ERP study aimed to elucidate the processing difficulties of syntactic complexity and an infelicitous use of OSV. The result showed an interaction of word order and givenness of arguments. O_NEWS_GIVENV elicited SLAN from O to S compared to S_NEWO_GIVENV. Importantly, however, O_GIVENS_NEWV did not exhibit SLAN compared to S_GIVENO_NEWV. At NP2, O_NEWS_GIVENV elicited a significant P600 effect compared to S_NEWO_GIVENV. O_GIVENS_NEWV, on the other hand, did not show a P600 effect compared to S_GIVENO_NEWV. These results are discussed in the following sections.

4.1. SLAN effect

O_NEWS_GIVENV showed a SLAN effect from O to S, in keeping with the results of Ueno and Kluender (2003), which presented OSV without context. However, O_GIVENS_NEWV did not show a comparable SLAN effect. One may view this difference between O_NEWS_GIVENV and O_GIVENS_NEWV as attributable to the number of referents presented to participants by the point of the NP1. In the new-given order, the participants read two NPs (i.e. an NP in the context and an NP1) by the NP1. In contrast, only an NP was presented at this position in the given-new order. If this difference affects SLAN’s amplitude, then we expect only the main effect of GIVENNESS. This prediction cannot explain an interaction between WO and GIVENNESS.

The question is why the processing reflected by the SLAN was costly in the infelicitous but not in the felicitous context. Previous studies on long-distance dependency formation have argued that the SLAN effect reflects a working memory load of actively storing a filler, such as a head noun of a relative clause and a fronted O (Hagiwara et al., 2007; King & Kutas, 1995; Kluender & Kutas, 1993; Matzke et al., 2002; Müller et al., 1997; Phillips et al., 2005). If we maintain this functional interpretation, SLAN’s lack in the felicitous context implies that a discourse representation ameliorates the cost of holding a filler in the working memory in some way. However, this interpretation is challenged when we examine more carefully SLAN’s functional significance. SLAN has also been observed for lexico-semantically vacuous displaced constituents, such as “wer” (who-ACC) in German (Thomas asks himself, who-acc on Tuesday afternoon after the accident the doctor ___ called has, Fiebach et al., 2002). Kluender and Kutas (1993) observed a larger SLAN effect for a matrix wh-question (Who have you forgotten … ?) than a yes-no question (Have you forgotten … ?), although they did not conduct a cumulative analysis. Furthermore, Wagers and Phillips (2009, 2014) argued that the parser actively maintains a syntactic category of a filler (e.g. NP) until encountering a gap, while it releases semantically detailed information from the working memory and reactivates it after receiving direct evidence for the location of a gap. Given these findings, a straightforward interpretation of what SLAN reflects is the active maintenance of a filler’s syntactic category rather than lexico-semantic information. In this interpretation, it is not clear how discourse-level information alleviates the memory cost of holding a filler’s syntactic category unless a linking hypothesis exists.

Alternatively, SLAN may reflect the process of manipulating a discourse representation in memory. In O_NEWS_GIVENV, the processing cost should increase when encountering a discourse-new O because it has no referent in the preceding context, even though scrambling presupposes a shared referent in a discourse that directly or implicitly refers to an O. Accordingly, participants had to accommodate an unsatisfied presupposition to build a coherent discourse representation, probably by inventing an additional implicit context that linked the context and the target sentence. This process may be reflected by a SLAN observed in O_NEWS_GIVENV. The SLAN effect could be similar to the S(L)AN effect elicited by definite NPs and pronoun in the context with two salient referents, because the use of these NPs also presupposes a salient referent to which a speaker intends to refer (“David had asked the two girls to clean up their room before lunchtime. But one of the girls had stayed in bed all morning, and the other had been on the phone all the time. David told the girl  … ”, Nieuwland & Van Berkum, 2006; Van Berkum, 2004; Van Berkum, Brown, & Hagoort, 1999; van Berkum et al.2003; Van Berkum, Zwitserlood, Bastiaansen, Brown, & Hagoort, 2004; Van Berkum, Koornneef, Otten, & Nieuwland, 2007).⁴ In O_GIVENS_NEWV, on the other hand, the manipulating process is not necessary because the discourse has already introduced a referent that refers to an O of OSV, leading to SLAN’s lack in the felicitous context.

To date, the extent to which this view can account for previous findings of SLAN effects remains unclear. To our knowledge, no ERP experiment examines the interaction of the processing of post-nominal relative clauses by context. However, given that a behavioural study showed that their processing was largely alleviated by the felicitous context (Roland, Mauner, O’Meara, & Yun, 2012), SLAN could decrease in amplitude due to a felicitous context. Another observation related to the present view is an LAN effect in response to the matrix question compared to the yes-no question (Kluender & Kutas, 1993). Since wh-questions, such as “What did Mary read?”, presuppose the existence of an entity that Mary read (e.g. Postal, 1971; Karttunen & Peters, 1976), a wh-question introduced without context requires receivers to accommodate this presupposition. Accordingly, the LAN effect in response to a wh-question (Kluender & Kutas, 1993) could be accounted for in terms of contextual factors. However, more empirical data needs to be accumulated to test this view.

4.2. P600 effect at the NP2

At NP2, O_NEWS_GIVENV elicited a larger posterior positivity compared to S_NEWO_GIVEN, whereas no indication exists of a positivity in response to O_GIVENS_NEWV. The former result is consistent with previous studies on scrambling in Japanese (Hagiwara et al., 2007; Koso, Hagiwara, & Soshi, 2007; Ueno & Kluender, 2003), as well as those on filler-gap dependency in other languages (Fiebach et al., 2001; Fiebach et al., 2005; Kaan et al., 2000; Phillips et al., 2005; Rösler et al., 1998; Yasunaga et al., 2015; Yano et al., 2017, Yano, iikuni, et al., 2017). The positivity for O_NEWS_GIVENV began to diverge from S_NEWO_GIVENV earlier than typical P600 effects. This early peak latency may be due to the priming effect. When comparing O_NEWS_GIVENV and S_NEWO_GIVENV, the critical noun was a discourse-given NP. Accordingly, the repetition priming facilitated lexico-semantic processing and the subsequent process started earlier. This possibility was supported by the observation that the peak latency of P600 was significantly earlier for the given NPs than the new NPs (Midline: F (1, 15) = 5.24, p < 0.05; Lateral: F (1, 15) = 6.01, p < 0.05; Temporal: F (1, 15) = 2.36, p = 0.14).⁵ Importantly, O_GIVENS_NEWV did not show a larger positivity even in the typical P600 time-window. This result of the WO by context interaction is consistent with the behavioural result by Koizumi and Imamura (2017).

First, it is less likely that the P600 reflects the revision cost of a syntactic structure. Because an accusative case was attached to the displaced O of experimental sentences, no S-O ambiguity existed in the present experiment. Furthermore, SLAN’s presence for OSV in the infelicitous context suggests that the parser noticed a scrambled structure while processing an adverbial phrase under either interpretation of SLAN discussed above. Thus, the P600 is not the same as the P600 that has been observed for garden-path sentences (Kaan & Swaab, 2003a, 2003b).

Second, Kaan et al. (2000) proposed that P600 reflects a syntactic integration difficulty. If we hold this view, the lack of the P600 effect suggests the facilitated syntactic integration of a filler and its gap. However, a question arises as to why the discourse givenness of an O reduces a P600 effect at the NP2. One possibility is that the syntactic integration difficulty might correlate with memory cost, because several studies failed to observe a P600 effect for scrambled sentences in cases in which the distance between a filler and its gap is relatively short. For instance, Hagiwara et al. (2007) compared the processing of middle scrambling with that of long scrambling, using the sentences in (8) below. They found that long scrambling elicited a P600 effect compared to the canonical condition, whereas middle scrambling did not (see also Koizumi & Yasunaga, 2017). Presumably, when the parser consumes less memory cost, it needs the less syntactic integration cost. In the present case, since a discourse-given O of OSV did not impose a memory pressure on the parser, the parser can integrate the filler with the gap relatively easily. On the other hand, a discourse-new O was difficult to integrate with the gap because it incurred a memory cost evidenced by SLAN. However, this conjecture is not consistent with the result of Fiebach et al. (2002) and Phillips et al. (2005), who reported a P600 effect, irrespective of the length between a filler and a gap.

Table

(8)	a. Canonical condition:
	Kaiken-de	shacho-wa	hisho-ga	bengoshi-o		sagashiteiru-to	itta.
	meeting-at	president-top	secretary-nom	lawyer-acc		was.looking.for-c	said.
	“At the meeting, the president said that the secretary was looking for the lawyer.”
	b. Middle-scrambled condition:
	Kaiken-de	shacho-wa	bengoshi-oi	hisho-ga	ti	sagashiteiru-to	itta.
	meeting-at	president-top	lawyer-acc	secretary-nom		was.looking.for-c	said.
	c. Long-scrambled condition:
	Kaiken-de	bengoshi-oi	shacho-wa	hisho-ga	ti	sagashiteiru-to	itta.
	meeting-at	lawyer-acc	president-top	secretary-nom		was.looking.for	said.

Alternatively, the P600 may reflect a conflict resolution of different information types. When scrambled sentences were not presented with a supportive context, the parser should have faced a conflict between syntactic and information structures. That is, syntactic information signals that the parser must reconstruct a filler at the derived position into the original position to receive a thematic role. However, the information structure does not validate that the filler is located at the topic position because scrambling yields an ill-motivated new-given order. An increasing number of recent ERP studies have argued that the P600 is not a manifestation of pure syntactic processing difficulty (e.g. Bornkessel-Schlesewsky & Schlesewsky, 2008; Brouwer, Crocker, Venhuizen, & Hoeks, 2016; Brouwer, Fitz, & Hoeks, 2012; Kuperberg, 2007; Vissers, Chwilla, & Kolk, 2006). Instead, it indexes a process of integrating several types of information, such as syntax and semantics. Thus, a P600 likely reflects the resolution of a conflict between syntactic structure and information structure encoded by OSV. However, the issue of whether the same process underlies a P600 effect in other languages is not clear, since, for example, the pre-gap region of relative clauses in English is a verb, where different types of processes should also be performed, as discussed below.

4.3. P600 effect at the V

At the V, a larger P600 effect was observed for OSV compared to SOV in the new-given condition. The P600 at the V has been observed in previous studies (Hagiwara et al., 2007; Weckerly & Kutas, 1999). Phillips et al. (2005) proposed that the P600 at the V reflects “the syntactic and semantic operations involved in confirming the compatibility of the filler and the verb for thematic role assignment, and compositionally interpreting the verb and its arguments” (Phillips et al., 2005, p. 425). Because this process should be necessary at the V of the OSV in Japanese, it explains a P600 effect for OSV in the new-given context. However, it remains unclear why OSV in the given-new context did not elicit a similar P600 at V.

The P600 effect in the present experiment contradicts with the previous ERP experiment of Japanese scrambling that reported a larger anterior negativity for OSV than SOV at the V (Ueno & Kluender, 2003). However, this AN effect reflects a wrap-up process that has often reported in ERP experiments (Friederici & Frisch, 2000; Osterhout, 1997). In the present study, an auxiliary verb was placed at the sentence’s end to avoid the wrap-up negativity at the V. As expected, the wrap-up negativity was observed at the sentence’s final region, although no significant effect existed between conditions.

5. Conclusion

The present study explored the interaction of syntactic complexity of scrambling and discourse factors during Japanese sentence comprehension. The result of the ERP experiment clearly demonstrated that the felicitous use of scrambling alleviated filler-gap dependency formation, as evidenced by a significant reduction of SLAN and P600 effects. This finding suggests that the processing difficulty that has been observed for non-canonical word orders is largely associated with discourse factors, such as the alignment of discourse-old and discourse-new NPs. Nevertheless, further investigation is necessary to clarify the functional significance of SLAN and P600.

Acknowledgements

We thank anonymous reviewers and the editor for their insightful comments. We are also grateful to Mineharu Nakayama, Ellen Lau, Hajime Ono, and Shin Fukuda for their helpful comments. This study was supported by JSPS KAKENHI (#15H02603, PI: Masatoshi Koizumi), a Grant-in-Aid for JSPS Research Fellows (#13J04854, PI: Masataka Yano) and Kyushu University Wakaba Project (#30203, PI: Masataka Yano).

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Masataka Yano http://orcid.org/0000-0003-4465-8456

Additional information

Funding

This study was supported by the JSPS KAKENHI (#15H02603, PI: Masatoshi Koizumi), a Grant-in-Aid for JSPS Research Fellows (#13J04854, PI: Masataka Yano) and Kyushu University Wakaba Project (#30203, PI: Masataka Yano).

Notes

1. They also observed a P600 again at the adverb (“toto” finally). The successive P600 effects may be due to a temporal ambiguity of an original position of the filler. If the parser actively attempts to fill a gap (Active Filler Strategy, Frazier & Clifton, 1989), it should perform a gap-filling parsing at S first and do so again at the adverb after detecting a final gap position.

2. One might wonder whether repeating proper names in the context and the target sentences sounds unnatural because they are discourse-old information. However, Tsuchiya, Yoshimura, and Nakayama (2015) reported that native Japanese speakers overwhelmingly preferred the use of referential nouns (e.g. definite NPs) to pronouns in the narrative telling task, unlike in English, in which pronouns are preferred to refer to a discourse-old referent. Hence, the repeated use of proper names is not problematic in Japanese. However, because the preceding context renders an NP a topic of a discourse, marking it with a nominative or accusative case instead of a topic marker (-wa) is not frequent in Japanese. However, Hirotani and Schumacher (2011) did not observe any difference between the nominative and the topic S when it was mentioned in the preceding context. Thus, it is unlikely that this affected our results. Furthermore, the use of a topic marker “-wa” for discourse-given NPs is problematic for the purpose of the current experiment because this induces an S-O ambiguity. Assuming that the native Japanese speakers disambiguate ambiguous sentences into canonical sentences, O-wa_givenS-ga_newV (O-top S-nom V) should be temporarily analysed as a canonical SOV sentence until encountering S. Accordingly, such a sentence should not elicit a SLAN effect, making it impossible to examine how it is affected by discourse-level information.

3. In single-word analyses, the ERP of the baseline time window was analysed to ensure that ERP difference was not induced as a result of the baseline correction procedure. The results reveal no significant main effect or interaction at the baseline time window when the ERPs were time-locked to the onset of the previous region.

4. An important caveat is that Nieuwland, Petersson, and Van Berkum (2007) found an increased activation for a referentially ambiguous pronoun at the medial prefrontal region, which is different from the left inferior frontal gyrus that activates during the processing of non-canonical word orders (Grewe et al., 2007; Kim et al., 2009; Kinno, Kawamura, Shioda, & Sakai, 2008).

5. The mean peak latency of the positivity was calculated for each channel with the ERP Measurement Tool of ERPLAB (Lopez-Calderon & Luck, 2014) by finding a latency in which the greatest positivity was observed between 300 and 900 ms of NP2. The statistical analyses were conducted in the same way as reported in Section 3.5.