Glossa Psycholinguistics

1.1 Processing models

As mentioned earlier, major processing models commonly assume that when a verb is encountered, preceding arguments have to be retrieved from working memory. Models differ as to what determines the difficulty of such retrieval.

The dependency locality theory (DLT) of Gibson (2000) assumes that the activation of a syntactic head decays as more words are processed, since the amount of activation available to the computational system is limited. In particular, processing of NPs and VPs that introduce a new discourse object is assumed to consume substantial computational resources. Therefore, the cost of retrieval of an argument at a verb is contingent on the number of intervening nominal and verbal heads that introduce a new discourse object. Take (1) as an example. The matrix verb disliked is read more slowly in (1b) than in (1a), because retrieval of the matrix subject head reporter is more difficult as two new discourse referents (senator and attacked) intervene. Some recent applications of DLT simply use the number of intervening words to measure the distance (Futrell, Gibson, et al., 2020; Futrell, Levy, et al., 2020). Thus, the core prediction of DLT is that distant dependencies are more difficult to process.

The activation model (or ACT-R model) of R. L. Lewis and Vasishth (2005) (also see R. L. Lewis et al., 2006) also predicts locality effects. The model assumes that linguistic structures are stored in our memory as a set of feature bundles, or chunks. Each chunk corresponds to a maximal projection. When necessary, chunks are retrieved by the cue-based retrieval mechanism (McElree, 2000; McElree et al., 2003; Van Dyke & Lewis, 2003). In this mechanism, chunks are directly accessed by a set of features, or cues, without serially searching through the memory. The mechanism is evidenced by the presence of retrieval interference in various configurations, many of which are independent from locality effects (Dillon et al., 2013; Van Dyke, 2002; Van Dyke & Lewis, 2003; Van Dyke & McElree, 2006, 2011; Wagers et al., 2009, among many others; see Jäger et al., 2017, for a review and meta-analysis).

Retrieval interference predicts a locality effect in configurations like (1). When disliked is encountered, the parser attempts to retrieve the matrix subject using cues such as [+subject] and [+animate] (Van Dyke, 2002). The embedded subject the senator interferes with the correct target the reporter, since both have features that match these cues. Retrieval interference leads to a slower reading time, either because the access to the correct target becomes slower, or because an incorrect item is sometimes retrieved, in which case the parser has to try another time to retrieve the correct one (Vasishth et al., 2019).²

In the activation model, a locality effect can also arise from the degradation of chunks due to time-based decay. Each chunk has a fluctuating activation level, which determines the retrieval speed.³ Since the activation decays as time passes, older chunks are more difficult to retrieve. Note that time-based decay makes predictions similar to the word-count-based version of DLT, with the trivial assumption that processing more words consumes more time.

Besides decay, encoding interference is another possible source of memory degradation, although it is not included in the activation model. Encoding interference refers to the damage to earlier items due to the encoding of new elements, especially those similar to the earlier ones. For example, in (1), encoding senator can damage the representation of reporter, making its later retrieval more difficult. Empirical evidence for encoding interference includes slowdown due to the similarity of features that are not used as retrieval cues (e.g., Gordon et al., 2001; Villata et al., 2018), and stronger effects of retroactive (backward) interference compared to proactive (forward) interference (Van Dyke & McElree, 2011). These effects cannot be attributed to retrieval interference. Degradation of memory representations due to encoding interference should also make retrieval of distant chunks difficult.

In sum, while processing models assume different sources of processing difficulty, they agree that memory representations become more difficult to retrieve as they become distant. If representations of arguments are subject to such degradation and they have to be retrieved at the verb, then these models predict argument-verb locality effects.

1.2 Experimental results and explanations

Table 1 lists previous experimental studies that investigated the argument-verb locality effect and their results. The results have been mixed, but they are not completely random. Below we review key experimental results and existing explanations for them.

Grodner and Gibson (2005) is a classic case for locality effects. They conducted a self-paced reading (SPR) experiment using sentences like (2), manipulating the distance between the subject the nurse and the verb supervised, and whether the clause is embedded as a relative clause (RC) or not. Based on DLT, they predict a main effect of distance and an interaction of distance and embedding. The latter is predicted because processing a verb in an RC involves additional retrieval of the relative pronoun who.

As expected, they found that the verb was read more slowly when the subject was more distant, and this effect was stronger in an RC. Bartek et al. (2011) confirmed their predictions in eyetracking experiments using the same structures. Notably, Bartek et al. found simple effects of distance even in the matrix clause condition (2a), where no additional retrieval of a relative pronoun is involved. This suggests that argument-verb dependencies themselves are subject to the locality effect.

However, several studies in verb-final languages have failed to find such a locality effect (Konieczny, 2000; Levy & Keller, 2013; Nakatani & Gibson, 2008, 2010; Vasishth & Lewis, 2006). Various factors are claimed to be responsible for the lack of a locality effect.

One such factor is expectation. Consider the following German sentences, used in Konieczny (2000)’s SPR study. The distance between the object NP (die Rose ‘the rose’) and the verb participle ((hin)gelegt ‘laid (down)’) was manipulated by the position and length of an RC that modifies the object NP, and the presence and length of a PP that modifies the verb, as shown in (3).

Konieczny (2000) found that the verb participle was read faster when more words intervene between the object and the verb, contrary to the prediction. Such speed-up with an increased argument-verb distance is called an anti-locality effect. Levy (2008) attributed the anti-locality effect in such verb-final configurations to the increased expectation for the verb (also see Konieczny, 2000; Levy & Keller, 2013).

Another factor that may be responsible for the anti-locality effect is reactivation (Vasishth & Lewis, 2006). Attaching a modifier to an argument (e.g., attaching an RC to the object die Rose in (3)) involves retrieval of the argument. In the activation model, such retrieval boosts the activation of the argument and makes its later retrieval at the verb easier. Similarly, attaching an adverbial to the predicted verbal node boosts the activation of the verbal node, making its retrieval at the verb easier.

A problem with explanations based on expectation or reactivation is why we do not find a similar anti-locality effect in English sentences like (2), where the subject-modifying RC should reactivate the subject and increase the expectation for the appearance of the verb.⁴ Still, these explanations for anti-locality effects tell us that expectation and reactivation have to be controlled for in order to investigate the underlying locality effect.

Other factors are also claimed in the literature to affect the presence or absence of a locality effect. Nakatani and Gibson (2010) (also see Nakatani & Gibson, 2008) report that no argument-verb locality effect was observed in a Japanese SPR experiment that controlled for expectation and reactivation factors. They used center-embedded sentences like (4) and manipulated the argument-verb distance by scrambling the complement clause to the front. For example, the matrix subject syoki-ga ‘secretary-NOM’ is more distant from the verb hookokusita ‘reported’ in (4a) than in (4b).

Nakatani and Gibson (2010) found no significant correlation between DLT’s integration cost (i.e., argument-verb distance) and the reading time at the verb. It is difficult to attribute this observation to the aforementioned confounding factors of expectation and reactivation, since all conditions use the same subjects and objects, and they always appear before the verb. It should be noted, however, that the use of scrambling may have masked the underlying locality effect by increasing the processing cost in (4b), as the authors admit.

Nakatani and Gibson (2010) offer two possible accounts of why the locality effect is absent in some experiments, including theirs. One possibility is that the processing strategy is parametrized depending on the language, and head-final languages are “heavily expectation-oriented” (p. 108) so that there is no cost for distant dependencies. Another possibility is that locality only affects processing of certain types of dependencies, such as filler-gap dependencies, and not argument-verb dependencies. This account is corroborated by the observation that the lack of a locality effect is concentrated on plain thematic dependencies, i.e., argument-verb dependencies without additional complexities such as relativization, a wh-dependency, or negative concord (Nakatani, 2021a; Ono & Nakatani, 2014).

Levy and Keller (2013)’s eyetracking data in German also suggests that the strength of a locality effect depends on the configuration. The sentences they used are schematically shown in (5). In Experiment 1, the distance between the subject and the verb was manipulated by insertion of an AdvP and/or a dative NP, which are shown in parentheses in (5a). In Experiment 2, the entire clause was embedded as an RC (5b). (6a,b) shows actual sentences corresponding to (5a,b), with both the AdvP and dative NP present in the critical clause.

They found an anti-locality effect due to the dative NP in Experiment 1. However, this effect disappeared in Experiment 2 when both the AdvP and the dative NP were present. Levy and Keller’s account for this result was as follows. Generally, both expectation and locality affect reading times. The relative weighting of these factors is language-dependent, and speakers of verb-final languages are less prone to locality, since they are more practiced at non-local dependencies. Therefore, in German, a locality effect surfaces only when it is strong enough to overwhelm the expectation-based effects. Unlike Nakatani and Gibson (2010), this account does not take the type of dependency as the sole decisive factor. Rather, reading times are assumed to result from an interplay between expectation and locality effects (also see Vasishth & Drenhaus, 2011, for a similar proposal).

In short, Nakatani and Gibson (2010) and Levy and Keller (2013) both suggest that the type of the dependency and the word order (verb-final or verb-medial) are relevant for the presence or absence of a locality effect. Both suggest that plain thematic dependencies in verb-final languages do not exhibit a locality effect. However, there is evidence that plain thematic dependencies can exhibit a locality effect even in verb-final languages. Safavi et al. (2016) conducted SPR and eyetracking experiments in Persian, a verb-final language, and showed that plain thematic dependencies can yield a locality effect. In this study (Experiments 2 and 4), they manipulated the distance between the matrix arguments and the verb by the length of the NP in the intervening PP, as shown in (7). This design is intriguing, because the distance is manipulated without bringing confounding factors in. Neither the predictability of the verb nor the activation of the matrix arguments are affected by the additional NPs inside the PP.⁵ The linear order of the arguments is constant across conditions (i.e., no scrambling is used). They also manipulated the predictability of the verb from the object using object-verb complex predicates, such as a:rezouyee kard ‘wish do (make a wish)’.

They found a clear locality effect at the verb. This is unexpected if plain thematic dependencies are inherently locality-insensitive, as Nakatani and Gibson (2010) suggested. There was also an effect of predictability, but it did not interact with the locality effect. This indicates that locality and expectation are two independent factors that both affect reading times, in line with Levy and Keller (2013)’s view (also see Ono & Sugi, 2018). The Persian data is still problematic for Levy and Keller (2013)’s account since it shows a locality effect in an arguably very simple argument-verb dependency in a verb-final language. A possible explanation for this data will be discussed in Section 2.

Summarizing the review so far, the argument-verb locality effect is predicted from two assumptions: representations in human working memory degrade gradually, and arguments have to be retrieved at the verb. However, experimental results have been mixed. In particular, several studies on verb-final structures have failed to observe the effect. Various explanations for the mixed results can be grouped into two views. The underlying locality view points to confounding factors that may have masked the underlying locality effect, such as expectation, reactivation by modification, and scrambling. In the current study, these factors are controlled for. We add the lack of statistical power to detect the locality effect as another possibility along this line. The inherent insensitivity view states that certain languages and certain types of dependencies are inherently insensitive to locality. Namely, thematic dependencies in verb-final languages are argued to be insensitive. We have seen that there is a counterexample to this view, however. In the next section we argue for another view, the pre-verb reactivation hypothesis, which is able to account for the counterexample and the general lack of locality effects in verb-final thematic dependencies at the same time.

2.1 Pre-verb reactivation based on a dependency chain

As mentioned in Section 1, Japanese fails to show a locality effect in plain thematic dependencies, but does show an effect when the dependency involves a wh element or a negative polarity item (NPI). Nakatani (2021a) proposes an account for this contrast based on the pre-verb reactivation of the arguments and adjuncts (also see Nakatani, 2021b). He assumes that elements that (are predicted to) depend on the same head form an object called dependency chain, and the chain is reactivated every time another member is added to it. Such reactivation counteracts the degradation of the representations of the thematic elements, which could lead to a locality effect. Non-thematic dependencies, on the other hand, do not receive such intermediate reactivation, and are therefore subject to a locality effect. Consider the Japanese SOV sentence (8a), for example. The matrix elements gradually form a chain ⟨John, in a park, apple, ate⟩ as the sentence unfolds. Since the chain that contains John is reactivated every time one of these elements is added, there is no processing difficulty due to the surface distance between John and the verb tabe-ta ‘ate.’ If the sentence involves an NPI sika ‘only,’ as in (8b), there is an additional chain ⟨only, NEG⟩, hence processing difficulty due to the distance between sika and NEG. Nakatani also applies this account to locality effects observed in RCs, which have a distinct dependency chain between the relative pronoun and the trace.

While Nakatani (2021a) argues that the dependency chain hypothesis explains why thematic dependencies are insensitive to locality, it also predicts locality effects between adjacent clausemates, in our view. This prediction is consistent with the aforementioned Persian data (7). In that experiment, there was a locality effect between the prepositional head and the verb. This is predicted by the dependency chain hypothesis as follows. The chain that would be formed in this sentence is ⟨Ali, chocolate, for, bought⟩. The chain is retrieved when, and only when, each of these elements is encountered. Therefore, the distance between for and bought causes a locality effect at the latter.

Thus, pre-verb retrieval of arguments based on dependency chains is able to account for the complicated pattern of argument-verb locality effects observed in previous studies. However, theoretical questions remain. What is a dependency chain, and why does adding an element to a chain result in cancelling the decay of all the elements that belong to that chain? That is not expected under the standard notion of dependency. For example, a subject and an object are not usually considered to form a dependency, hence an object is not expected to trigger the retrieval of the preceding subject. A possible answer to these questions will be discussed in 4.2. We will use the term pre-verb reactivation hypothesis in a sense that is neutral with respect to the underlying syntactic representation. It refers to the psycholinguistic hypothesis that arguments and adjuncts can be reactivated when a new clausemate is encountered.

2.2 Locality effects between adjacent clausemates

The pre-verb reactivation hypothesis makes a unique empirical prediction: there should be locality effects between adjacent clausemates. For example, sentences (9a,b) differ in the distance between the subject John-ga and the head of the locative adverbial kooen-de ‘in a park.’ The current hypothesis predicts that the reading time at kooen-de is longer in the long distance condition than in the short distance condition.

Let us formalize this prediction using the activation model of R. L. Lewis and Vasishth (2005). As mentioned earlier, this model assumes that each chunk (maximal projection) has a fluctuating activation that reflects its encoding and retrieval history. The base level activation of item i is given by the equation

$B_i=\ln \left({\displaystyle \sum _{j=1}^n{t}_j^{-d}}\right)$

where t_j is the time since jth encoding or retrieval of the item, and d is the fixed decay rate. We adopt d = 0.5, following R. L. Lewis and Vasishth (2005). We further assume that every region takes a constant time (500 ms) to read. We admit that this is a huge simplification. But calculating exact reading times from the model requires specification of numerous factors, including the exact structure for the entire sentence, all the cues used in all the retrievals, a linking hypothesis between the retrieval time and reading time, and so on. That is beyond the scope of this paper. Another problem is that the standard activation model does not take encoding interference into account, despite its empirical plausibility (see Section 1). The crucial prediction is not affected by such details, however. What is important is that the two conditions substantially differ in the time that passes between the encoding of the subject and its first retrieval at the locative adverbial. We therefore confine the use of the activation model to the simulation of activation fluctuation that varies by the distance manipulation, and combine the simulation results and other considerations to obtain predictions. Figure 1 shows how the activation of the subject noun John-ga in (9) would fluctuate under the current assumptions. As presented in the figure, the locative adverbial is predicted to show a locality effect due to the distance from the subject. The subject is also retrieved at the object and at the verb, but these retrievals should result in little locality effect because the activation of the subject has become largely equal between the Short and Long Distance conditions once it is activated at the locative adverbial.

An important assumption here is that the parser analyzes the locative adverbial, the accusative-marked noun, and the verb as clausemates of the subject. Previous experimental results support the view that the parser adopts such a single clause analysis unless the input suggests the contrary (Kamide et al., 2003; Miyamoto, 2002).

In summary, we argued that (i) pre-verb reactivation of arguments can account for the complicated facts regarding the presence and absence of the argument-verb locality effect, and (ii) the pre-verb reactivation hypothesis predicts locality effects between adjacent clausemates. The next section reports two experiments that test this prediction. To anticipate the results, the experiments show locality effects between the subject and a locative adverbial (like kooen-de ‘in a park’ in the example above), lending support to the pre-verb reactivation hypothesis.

Figure 1

Simulated fluctuation of the activation of the subject in the sentences in (9). Solid line: short distance condition; Dashed line: long distance condition. The asterisk indicates where the locality effect is predicted.

3.1 Experiment 1

3.1.2 Predictions

The three views we are considering make different predictions regarding the reading times, as summarized in Table 2. The underlying locality view predicts a classic argument-verb locality effect. That is, the Long Distance conditions should be slower at the matrix verb, because retrieval of the subject is more difficult when it is more distant. Panel A of Figure 2 illustrates the simulated fluctuation of the subject’s activation under this view. The inherent insensitivity view simply predicts no locality effect. To state such a prediction within the activation model, one would have to argue either that the decay rate of the relevant element is exceptionally low, or that the effect of the activation level of the relevant element on its retrieval latency is exceptionally small. Panel B of Figure 2 illustrates the former possibility, with the decay rate d = 0.25 instead of d = 0.5 (the latter possibility is not represented since the activation profile would be the same as Panel A). Finally, the pre-verb reactivation hypothesis predicts a locality effect between the subject and the locative adverbial (Panel C).

Table 2

Predictions of different views and results of the experiments. D indicates a main effect of Distance; (D) means that the main effect of Distance is not predicted explicitly in the literature, but should be predicted if adjunction of the adverbial is taken into account; D:I indicates Distance-Interference interaction; – means neither effect is predicted or observed.

	Adverbial	Verb
Underlying locality	(D)	D, D:I
Inherent insensitivity	–	–
Pre-verb reactivation	D, D:I	–
Result – Experiment 1	D, D:I	–
Result – Experiment 2	D, D:I	–

Figure 2

Simulated fluctuation of the activation of the subject in the sentences in Experiment 1. Panel A: underlying locality view; Panel B: inherent insensitivity view (one instantiation; see the main text); Panel C: pre-verb reactivation hypothesis. Solid line: Short Distance condition; Dashed line: Long Distance condition. The asterisks indicate where the locality effect is predicted.

The locality effect (the main effect of Distance) between the subject and the locative adverbial can also be due to the retrieval of the attachment site of the adverbial, however. When the adverbial phrase is completed by its head (mae-de in the above example), it has to be attached to a node in the matrix clause structure. If the attachment site is predicted when the subject is introduced, as would be assumed in the left-corner parsing strategy (Abney & Johnson, 1991), there would be a locality effect due to the retrieval of the attachment site. Therefore, a main effect of distance between the subject and the locative adverbial does not provide convincing evidence for the retrieval of the subject noun at the adverbial. Due to this consideration, the Distance factor was crossed with Interference. The idea is that higher similarity between the matrix subject and the RC noun would increase the retrieval difficulty of the matrix subject via encoding interference (see 1.1 for encoding interference). The Interference effect should be stronger in the Long Distance condition, because retroactive interference is stronger than proactive interference (Van Dyke & McElree, 2011). Thus, the pre-verb reactivation hypothesis predicts an interaction between Distance and Interference at the locative adverbial, such that the Long Distance, High Interference condition is the slowest. The underlying locality view predicts the same interaction at the verb.

Another possible source of the locality effect at the locative adverbial is the spillover from the RC. In SPR experiments, processing difficulty due to the material in a region can be reflected not only in the reading time of that region but also those of subsequent regions (spillover; Bartek et al., 2011; Mitchell, 1984). To separate out such spillover effects from effects inherent to the material of the current region, we included the reading time of the previous region in the statistical model (Vasishth, 2006) (see 3.1.4 for details). But this approach is only able to factor out the spillover as long as the effect caused by the material in the previous region surfaces both in the previous and the current regions. Given that the button can be pressed in principle once the visual encoding is done, effects of syntactic processing may start to show up only in subsequent regions. This is problematic in interpreting the data in two ways. Take the matrix locative adverbial at Region 6, for example. First, even if some effect is observed at Region 6, it is still possible that this is a delayed effect from the relativization process initiated by the RC head at Region 5 (in Japanese, the presence of an RC becomes evident only when the RC head appears). Second, the effect of the locative adverbial may only appear with delay, in subsequent regions. But it is also possible that the effect at Region 6 is a genuine effect of the locative adverbial. Due to these considerations, we look at the reading times of not only the locative adverbial itself (Region 6) but also the preceding Region 5 and the following Region 7. Effects in Region 7 are more likely to be due to the processing of the locative adverbial than effects in Region 6.

3.1.5 Results

Comprehension accuracy. The overall comprehension accuracy was 85.5% after the initial screening of participants and items, indicating that participants paid attention to the stimuli. Table 3 shows comprehension accuracy for target sentences by condition after the screening. GLME modeling showed no significant effects.

Table 3

Comprehension accuracy by condition, Experiment 1.

Condition	Mean (%)
Short Distance, Low Interference	87.2
Short Distance, High Interference	85.3
Long Distance, Low Interference	86.8
Long Distance, High Interference	83.0

Reading time. Figure 3 shows log-transformed reading times by region and condition after the outliers are removed. Table 4 shows model estimates for the regions of interest.

Figure 3

Reading times by region and condition, Experiment 1. Error bars indicate standard errors.

Table 4

Model estimates for selected regions of Experiment 1. Significance codes: *** for p < 0.001, ** for p < 0.01, * for p < 0.05, . for p < 0.1.

Region 5	Estimate	SE	df	t	p
(Intercept)	6.3933	0.0546	54.92	117.118	0.0000	***
Distance	–0.0370	0.0187	46.22	–1.978	0.0539	.
Interference	–0.0163	0.0114	1362.03	–1.438	0.1506
Spillover	0.0941	0.0395	41.57	2.383	0.0218	*
Distance:Interference	–0.0145	0.0114	1370.43	–1.272	0.2035
Region 6	Estimate	SE	df	t	p
(Intercept)	6.2681	0.0459	54.70	136.669	0.0000	***
Distance	0.1195	0.0171	46.63	6.969	0.0000	***
Interference	–0.0173	0.0106	1353.10	–1.623	0.1049
Spillover	0.1102	0.0372	44.69	2.966	0.0048	**
Distance:Interference	0.0214	0.0107	1367.73	2.000	0.0458	*
Region 7	Estimate	SE	df	t	p
(Intercept)	6.2345	0.0466	58.96	133.8569	0.0000	***
Distance	0.1071	0.0105	1410.00	10.1616	0.0000	***
Interference	0.0002	0.0102	1398.59	0.0180	0.9857
Spillover	0.1156	0.0301	43.33	3.8391	0.0004	***
Distance:Interference	–0.0023	0.0102	1399.32	–0.2287	0.8191
Region 8	Estimate	SE	df	t	p
(Intercept)	6.1041	0.0299	47.29	203.9322	0.0000	***
Distance	0.0072	0.0083	1426.41	0.8715	0.3836
Interference	0.0026	0.0081	1411.27	0.3192	0.7496
Spillover	0.1697	0.0298	46.05	5.7018	0.0000	***
Distance:Interference	0.0081	0.0081	1410.60	1.0109	0.3122
Region 9	Estimate	SE	df	t	p
(Intercept)	6.0051	0.0293	49.74	204.7173	0.0000	***
Distance	0.0108	0.0068	1373.22	1.6045	0.1088
Interference	0.0007	0.0079	46.52	0.0851	0.9325
Spillover	0.2120	0.0322	51.44	6.5861	0.0000	***
Distance:Interference	0.0069	0.0067	1371.05	1.0225	0.3067
Region 10	Estimate	SE	df	t	p
(Intercept)	6.0057	0.0277	52.89	216.9538	0.0000	***
Distance	0.0066	0.0077	29.59	0.8576	0.3980
Interference	–0.0143	0.0081	28.61	–1.7623	0.0887	.
Spillover	0.3260	0.0400	49.58	8.1442	0.0000	***
Distance:Interference	0.0054	0.0062	1374.50	0.8810	0.3785
Region 11	Estimate	SE	df	t	p
(Intercept)	6.0100	0.0282	47.56	212.9729	0.0000	***
Distance	0.0009	0.0065	1439.41	0.1425	0.8867
Interference	0.0021	0.0065	1434.40	0.3221	0.7474
Spillover	0.3713	0.0372	56.85	9.9802	0.0000	***
Distance:Interference	–0.0009	0.0064	1430.48	–0.1339	0.8935

Region 5 (the genitive noun modifying the locative adverbial; the head noun of the RC in the Long Distance conditions) showed a marginal effect of Distance, such that the Long Distance conditions were faster than the Short Distance conditions. Both Region 6 (the locative adverbial) and Region 7 showed a significant main effect of Distance, such that the Long Distance conditions were slower than the Short Distance conditions. Region 6 also showed a significant interaction of Distance and Interference. Nested analysis indicates that this is because the effect of Distance is larger in the High Interference condition. Region 10 (the matrix verb) showed a marginal effect of Interference, such that the High Interference conditions were faster than the Low Interference conditions.

3.2 Experiment 2

3.2.5 Results

Comprehension accuracy. The overall comprehension accuracy was 84.5% after the initial screening of participants and items, indicating that participants paid attention to the stimuli. Table 5 shows comprehension accuracy for target sentences by condition after the screening. GLME modeling showed no significant effects.

Table 5

Comprehension accuracy by condition, Experiment 2.

Condition	Mean (%)
Short Distance, Low Interference	85.1
Short Distance, High Interference	86.3
Long Distance, Low Interference	84.1
Long Distance, High Interference	82.6

Reading time. Figure 4 shows log-transformed reading times by region and condition after the outliers are removed. Table 6 shows model estimates for regions of interest.

Figure 4

Reading times by region and condition, Experiment 2. Error bars indicate standard errors.

Table 6

Model estimates for selected regions of Experiment 2. Significance codes: *** for p < 0.001, ** for p < 0.01, * for p < 0.05, . for p < 0.1.

Region 5	Estimate	SE	df	t	p
(Intercept)	6.4797	0.0553	54.04	117.0711	0.0000	***
Distance	–0.0348	0.0131	1345.36	–2.6561	0.0080	**
Animacy	0.0205	0.0128	1328.31	1.5994	0.1100
Spillover	0.1310	0.0446	43.35	2.9387	0.0053	**
Distance:Animacy	0.0016	0.0128	1323.85	0.1238	0.9015
Region 6	Estimate	SE	df	t	p
(Intercept)	6.3590	0.0495	44.62	128.4793	0.0000	***
Distance	0.0983	0.0182	44.34	5.3939	0.0000	***
Animacy	0.0174	0.0113	1290.09	1.5469	0.1221
Spillover	0.0727	0.0423	43.55	1.7206	0.0924	.
Distance:Animacy	0.0091	0.0112	1276.36	0.8096	0.4183
Region 7	Estimate	SE	df	t	p
(Intercept)	6.3427	0.0455	52.11	139.3374	0.0000	***
Distance	0.0914	0.0166	44.58	5.5038	0.0000	***
Animacy	–0.0011	0.0112	1297.20	–0.0953	0.9241
Spillover	0.1285	0.0417	40.49	3.0802	0.0037	**
Distance:Animacy	0.0232	0.0112	1291.42	2.0779	0.0379	*
Region 8	Estimate	SE	df	t	p
(Intercept)	6.1585	0.0335	45.52	183.8979	0.0000	***
Distance	–0.0084	0.0081	1350.57	–1.0342	0.3012
Animacy	–0.0009	0.0080	1346.04	–0.1177	0.9063
Spillover	0.1319	0.0311	39.88	4.2360	0.0001	***
Distance:Animacy	0.0007	0.0080	1343.02	0.0904	0.9280
Region 9	Estimate	SE	df	t	p
(Intercept)	6.0983	0.0247	54.57	247.0476	0.0000	***
Distance	–0.0010	0.0096	28.80	–0.1004	0.9208
Animacy	–0.0006	0.0086	42.77	–0.0741	0.9413
Spillover	0.2693	0.0431	51.07	6.2510	0.0000	***
Distance:Animacy	0.0008	0.0074	1290.34	0.1013	0.9194
Region 10	Estimate	SE	df	t	p
(Intercept)	6.0693	0.0262	48.21	231.3166	0.0000	***
Distance	–0.0038	0.0068	1351.45	–0.5639	0.5729
Animacy	–0.0006	0.0067	1348.90	–0.0942	0.9250
Spillover	0.3040	0.0393	41.33	7.7305	0.0000	***
Distance:Animacy	0.0042	0.0067	1342.87	0.6207	0.5349
Region 11	Estimate	SE	df	t	p
(Intercept)	6.0742	0.0274	46.97	221.6915	0.0000	***
Distance	0.0053	0.0061	1343.70	0.8758	0.3813
Animacy	–0.0027	0.0060	1333.68	–0.4442	0.6570
Spillover	0.4029	0.0429	57.77	9.3859	0.0000	***
Distance:Animacy	–0.0032	0.0060	1330.30	–0.5239	0.6004

Region 5 (the genitive noun modifying the locative adverbial; the head noun of the RC in the Long Distance conditions) showed a significant effect of Distance, such that the Long Distance conditions were faster than the Short Distance conditions. Both Region 6 (the locative adverbial) and Region 7 showed a main effect of Distance, such that the Long Distance conditions were slower than the Short Distance conditions. Region 7 also showed a significant interaction between Distance and Interference. The nested model indicates that this is because the effect of Distance is larger in the High Interference condition.

3.3 Between-experiment comparison

Half of the target sentences used in Experiments 1 and 2 were the same. Namely, the Low Interference conditions in Experiment 2 are the same as the High Interference conditions in Experiment 1. We noticed that the same sentences were read slightly differently in these experiments. Specifically, Experiment 2 was generally slower, especially around Region 6, where we observed the subject-adverbial locality effect. This is unlikely to be due to any systematic difference in the participant groups. They were recruited from the same communities and assigned one experiment randomly. The mean age of the participants were close (Experiment 1: 22.3 (SD: 5.1) / Experiment 2: 21.7 (SD: 2.4)). The experimental method was exactly the same in both experiments. The only difference was the other half of the target sentences.

One possible account for this difference is that the pre-verb retrieval of the subject is a reflection of the parser’s strategic effort to have a better understanding of the event, rather than a syntactically required operation. If so, whether a pre-verb retrieval takes place may be affected by various linguistic and extralinguistic factors, including task difficulty. In the current case, specifically, the matrix subject nouns were subject to more interference in Experiment 2 than in Experiment 1. This may have changed the participants’ perception of the difficulty of maintaining and remembering the subject, and encouraged them to perform strategic retrieval of the subject more often. This would result in an interaction between Distance and Experiment at the locative adverbial, and main effects of Experiment at the object and the verb. To investigate this possibility, we additionally conducted a between-experiment analysis of the reading time data.

The relevant target sentences are repeated in (14) below. The difference between the two conditions is the distance between the subject and the other matrix elements.

To compare the by-region reading times in the two experiments, LME models were constructed. The full model included the fixed intercept and slopes of Distance, Experiment, their interaction, and Spillover, as well as corresponding random intercepts and slopes for participants and items. Experiment 1 was coded as –1 and Experiment 2 as 1. Coding of other factors and model fitting were done in the same way as in Experiments 1 and 2.

Figure 5 shows the reading times by region, condition, and experiment. Table 7 shows model estimates for the regions of interest. The modeling results indicate that Experiment 2 was generally read more slowly. Such a trend persisted from Region 5 to 11, was significant at Regions 9 and 10, and was marginally significant in Region 7. In addition, Region 6 showed a significant interaction of Distance and Experiment. A nested model indicates that there were simple effects of Distance in both experiments, such that the Long Distance condition was slower, but the effect was larger in Experiment 1.

Figure 5

Reading times by region, condition, and experiment. Error bars indicate standard errors.

Although this analysis is exploratory, the results are striking in that they demonstrate that participants changed the way they read the sentences in response to a slight change in other experimental sentences. Specifically, stronger interference between the matrix subject and the RC subject led to a slowdown in a number of regions. This observation suggests that participants strategically commit to more retrieval of the subject when faced with more interference.

The Distance-Experiment interaction at Region 6 is particularly important, since it implies the strategic nature of the pre-verb reactivation of the subject at the adverbial, but it should be interpreted with caution. First, the interaction was found in Region 6 but not in Region 7. An effect that appears in Region 6 alone may be due to the processing of the preceding RC, for reasons discussed earlier with regard to Experiment 1. More specifically, it may be a delayed effect of the retrieval of the RC subject triggered by the RC verb. Even so, the interaction with Experiment still constitutes evidence that the retrieval of the argument is strategic, since it means that the same sentences are processed differently between the two experiments. Another caution concerns the way Distance and Experiment interact. If strategic retrieval was simply more frequent in Experiment 2, we should observe it as a larger locality effect, i.e., longer reading times in the Long Distance condition of Experiment 2. However, the actual interaction was driven by the Short Distance condition in Experiment 2 being read more slowly than the same condition in Experiment 1. This interaction can be attributed to a ceiling effect. Suppose that strategic retrieval is affected not only by other sentences in the experiment but also by the current sentence being processed. This can make the Long Distance conditions in both experiments “hit the ceiling,” i.e., the parser always commits to strategic retrieval to save the subject from degradation. Then, a difference from other experimental sentences should exist only in the Short Distance condition. In sum, the results of the between-experiment comparison are consistent with the idea that the retrieval of the subject is strategic. In 4.2 we will discuss why this is the case.

4.1 Summary of the results

Overall, the current results were consistent with the pre-verb reactivation hypothesis. To recapitulate, the pre-verb reactivation hypothesis states that arguments are often reactivated before the verb, and that counteracts their degradation. We argued that this hypothesis accounts for the mixed results regarding the argument-verb locality effect in previous studies. In two experiments, we manipulated the distance between the matrix subject and the following locative adverbial, and the similarity of an intervening noun to the matrix subject. The results of both experiments are consistent with the hypothesis that the subject is retrieved at the adverbial. Both Experiments 1 and 2 showed a main effect of distance and an interaction between distance and similarity.

In contrast, the current results do not provide evidence for the underlying locality view, which states that the argument-verb locality effect is masked by confounding factors. The current experiment controlled for such factors, but revealed no argument-verb locality effect. We also ensured that the experiments had sufficient statistical power to detect locality effects of a reasonable size. Admittedly, it is still possible that the size of the argument-verb locality effect in Japanese is very small, and can be observed with a larger sample size and/or in a lab experiment with less noise in the data. We leave this possibility for future research.

The results are also problematic for the inherent insensitivity view, which assumes that verb-final languages and/or thematic dependencies are generally insensitive to locality effects. We did observe a locality effect between the subject and the subsequent locative adverbial in a verb-final language.

4.2 An account of pre-verb reactivation based on constructivism

We now return to the question that we left unanswered in 2.1: what is the underlying structural representation that motivates the pre-verb reactivation? Isono and Hirose (2022) briefly suggest that the constructivist analysis of argument structure, which has been developed in generative syntax, can be an answer. As a possible theoretical implementation of pre-verb reactivation, we elaborate on this idea below.

According to the constructivist view, often couched in terms of Distributed Morphology (Halle & Marantz, 1993), some or all thematic arguments are introduced by functional heads rather than the lexical argument structure of the verb (Borer, 2005a, 2005b; Folli & Harley, 2005, 2007; Hale & Keyser, 1993; Harley, 2011; Kratzer, 1996; Lohndal, 2012; Marantz, 1997, 2013, among many others). This contrasts with what is called the projectionist view, which assumes that all arguments are licensed in the verb phrase that projects the lexical argument structure of the verb (see Borer, 2005a; Harley, 2011; Marantz, 2013, for discussions on constructivism vs. projectionism).

The precise structural analysis of argument structure is still hotly debated within constructivism. For current purposes, it is sufficient to adopt the assumption that there is a functional head (a particular type of ‘Voice’ head) that introduces an agent argument in a position that is external to the lexical verb (Borer, 2005b; Hale & Keyser, 1993; Kratzer, 1996; Marantz, 1997, among many others). Under this view, the structure for the transitive sentence (8a) would be as in (15) below. The subject John is base-generated as the specifier of the VoiceP, where it receives the agent role, and then raises to the specifier of the TP. We remain agnostic about the structure below Voice, represented as vP.

The separation of the external argument is supported by the observation (due to A. Marantz) that external arguments rarely trigger a special interpretation of the verb, while internal arguments often do (Kratzer, 1996). Kratzer argues that this is because the interpretation of the external argument is determined fully compositionally, as formulated in (16), which is the Neo-Davidsonian logical form that corresponds to (15).¹¹ Here, e denotes the event described by the sentence, and the predicates past, agent, eat, and in_the_park all take e as an argument, meaning that they describe aspects of the event e. Notably, the agent John is described as the agent of the event, and only indirectly associated with the verb eat via e.

Although such constructivist analyses have developed in theoretical linguistics, which is concerned with competence and based on offline judgments, it is reasonable from the perspective of theoretical parsimony to assume that these structures are also relevant for online processing (Bresnan & Kaplan, 1982; Chomsky, 1965; S. Lewis & Phillips, 2015; Phillips, 2012; Steedman, 2000). If constructivist structures underlie online processing, we must reconsider the view that the verb’s lexical content is the sole pivot of argument structure that specifies the thematic roles of the arguments all at once, although such an idea seems to tacitly underlie existing processing models. At the same time, given previous experimental results that show similarity-based interference with subjects (Van Dyke, 2007; Van Dyke & McElree, 2006, 2011), we cannot deny that agent arguments can be retrieved at the verb. Precisely speaking, those studies do not explicitly control for the thematic role of the subject, but the critical verbs used for illustration in the papers (complain, sail, fix, compromise) seem compatible with the agent role. The trigger of the retrieval of the agent at the verb is probably the indirect association between the verb and the agent argument via the event variable. Then, it further follows that the adverbial should also be able to trigger a retrieval of the preceding agent, since the adverbial and the agent are indirectly associated with the event variable in the same fashion as the verb and the agent.

An advantage of the account of pre-verb reactivation based on constructivism is that it naturally predicts the strategic nature of the retrieval that we observed in the between-experiment analysis. Under this account, the pre-verb reactivation is triggered by the indirect association between clausemates via the event variable. This is not required to build a grammatical syntactic structure, unlike the thematic role assignment from the verb to the arguments under a projectionist structural assumption. Rather, it is an optional operation that can be performed to have a better understanding of the event described by the relevant predicates.

In short, the constructivist analysis of argument structure can provide theoretical motivation for the pre-verb reactivation of an argument (or an adjunct) by a subsequent clausemate, which is a plausible account of the previous and current experimental results.

A more general implication of this constructivist processing hypothesis is that a structural analysis that has been developed in the study of competence grammar can derive a reasonable hypothesis for real-time performance. This is a desirable direction, since having a single generative system is conceptually more plausible than having two or three, and there is evidence that other aspects of competence grammar, such as island constraints, are actively used in real-time processing (S. Lewis & Phillips, 2015). Ultimately, this approach will have to face the issue of how to formulate the grammar in a way that allows left-to-right structure building (cf. Bresnan & Kaplan, 1982; Phillips, 2003; Sag & Wasow, 2011; Stabler, 2013; Steedman, 2000). Currently, constructivism is primarily formulated in Distributed Morphology (Halle & Marantz, 1993) and Minimalism (Chomsky, 1995), which describe syntactic structures as built in a bottom-up manner, especially in formulating principles such as the phase and Late Insertion. Bottom-up structure building is not problematic as far as competence grammar is concerned, and confining one’s interest to competence grammar is a valid research strategy. Still, given the theoretical and empirical plausibility of integrating competence grammar and performance, we should ask how constructivist structures can be built in a left-to-right manner. We leave this issue for future research.

4.3 Remaining issues

There are several other remaining issues for future research. First, the timing of strategic retrieval of arguments needs further investigation. The current study found evidence for two factors that affect strategic retrieval: addition of a clausemate, and difficulty of other sentences in the same experiment. But this does not have to be an exhaustive list. The effect of other experimental sentences points to the more general possibility that the experimental task affects retrieval strategy. There may also be individual differences. Individuals with low working memory capacity may resort to more retrievals to compensate for the low capacity, or may avoid the computational cost with additional retrieval. Indeed, interactions of working memory capacity and memory effects in reading have been reported, but their results vary as to whether low capacity individuals are more prone to memory effects or less prone to them (Nicenboim et al., 2016; Tan et al., 2017). Taking into account the strategic aspect of retrieval in sentence processing may shed light on this issue.

A related issue is whether retrieval of arguments takes place at the verb. The constructivist approach suggests that verbs do not necessarily trigger the retrieval of arguments, since verbs are not the sole pivot of argument structure. If thematic roles can be assigned by functional heads such as Voice before the verb is revealed, retrieval of arguments for thematic role assignment at the verb would be required only when the predicted role mismatches the verb (putting aside strategic retrieval based on association with the event variable). Thus, it is possible that whether arguments are retrieved at a given verb depends on the argument structure of the verb.¹² The current study is unable to directly speak to this issue. Even if some verbs in the experiments trigger the retrieval of the subject, we would not be able to observe that as an effect of Distance (or any other factor), since the distance between the verb and the last reactivation of the subject, i.e., the object, is constant across conditions. Studies using the cross-modal priming and visual world paradigms have found differences in the activation profile of subjects in sentences with unergative and unaccusative verbs (Burkhardt et al., 2003; Friedmann et al., 2008; Koring et al., 2012). A comparable contrast between unergatives and unaccusatives is also reported in sentence production (Momma et al., 2018). Future study may investigate if an unergative-unaccusative contrast can also be observed in reading times, and whether transitive verbs also show contrasts between verb types, such as agent, causer, and experiencer verbs (Belletti & Rizzi, 1988; Folli & Harley, 2005).

The processing of functional heads must also be further investigated. In the current study, we did not work out how functional heads are generated and reactivated during incremental processing. This is because the hypothesis space is too large. Indeterminacy resides both in the syntactic structure, which is still hotly debated in theoretical linguistics, and in how the structure is built incrementally. There could be locality effects between non-adjacent clausemates due to retrieval of some functional heads that are predicted in advance, such as T and Voice. However, we did not observe such effects. One possible reason for the lack of such effects is that functional heads are conflated into a single chunk by head movement. In that case, the single chunk is retrieved every time an argument or an adjunct is added to the clause, so there would be locality effects only between adjacent clausemates. Future research may investigate particular aspects of the processing of functional heads by experiments specifically designed for that purpose. We believe this is where collaboration of psycholinguistics and theoretical linguistics would be particularly fruitful.