Proximity and Same Case Marking Do Not Increase Attraction Effect in Comprehension: Evidence From Eye-Tracking Experiments in Korean

Kwon, Nayoung; Sturt, Patrick

doi:10.3389/fpsyg.2019.01320

ORIGINAL RESEARCH article

Front. Psychol., 06 June 2019

Sec. Psychology of Language

Volume 10 - 2019 | https://doi.org/10.3389/fpsyg.2019.01320

This article is part of the Research Topic Featural Relations in the Brain: Theoretical and Experimental Perspectives on Grammatical Agreement View all 16 articles

Proximity and Same Case Marking Do Not Increase Attraction Effect in Comprehension: Evidence From Eye-Tracking Experiments in Korean

$\nNayoung Kwon$ Nayoung Kwon¹^*

Patrick Sturt²

¹Department of English, Konkuk University, Seoul, South Korea
²Department of Psychology, University of Edinburgh, Edinburgh, United Kingdom

Previous studies have suggested that during the on-line sentence processing, relevant memory representations are directly accessed based on cues at retrieval (McElree et al., 2003). Under this hypothesis, retrieval cues activate any memory representation with matching features, leading to the so-called attraction effect. This predicts that attraction effects would be modulated by memory representation of a distractor. Here, we investigated this possibility, focusing on two factors (i.e., proximity to the retrieval point and the number of matching features) that would affect representation of a distractor in three Korean eye-tracking experiments. We predicted that if memory representation of a distractor decays over time, a distractor close to a retrieval point would lead to stronger attraction effects. We also predicted that a distractor would be more likely to lead to interference when it shares a higher number of matching features with the retrieval cues of a dependency, relative to the target of the dependency, due to multiple direct accesses based on multiple matching cues. However, the results did not show evidence that proximity of a distractor to the retrieval point enhanced attraction effects. Likewise, there was no evidence that a greater number of matching cues of a distractor alone would trigger more mis-retrieval, in contrast to a previous finding that a greater number of mismatching cues of a licit antecedent in addition to a greater number of matching cues of a distractor did so (Parker and Phillips, 2017). On the other hand, the results suggested that a distractor marked with nominative case was more likely to be mis-retrieved as the subject of a verb, compared to a distractor marked with a dative case, suggesting that the subject grammatical role is a critical cue for a subject-verb agreement. These results are best compatible with the hypothesis that retrieval cues are weighted, possibly depending on the nature of the dependency that is currently processed.

Introduction

Successful processing of a long-distance dependency requires retrieval of linguistic items in working memory. For example, in (1) the head NP (i.e., the book) of a relative clause should be retrieved at the embedded verb (i.e., admired), where it can be associated with a thematic role within the relative clause, and in (2) the head NP (i.e., the key) of a complex NP should be retrieved at the main verb (i.e., was) to form a subject-verb agreement dependency.

(1) This was the book that the editor admired.

(2) The key to the cabinet was rusty from many years of disuse.

Results of previous studies have been argued to support a content-addressable direct access model of retrieval (McElree, 2000; McElree et al., 2003; c.f., Kintsch, 1970), according to which items stored in working memory are activated in parallel, based on the matching of retrieval cues (see Lewis and Vasishth, 2005) for an implementation of such a content-addressable cue-based retrieval model of sentence processing).

Importantly to the goal of the study, this cue-based parallel retrieval mechanism predicts that the processing of a linguistic dependency can be affected by elements in memory that are not licit parts of that dependency. This arises because the model predicts that any item in memory that matches in features with a retrieval cue will be activated to a certain extent, even when the feature match is only partial. Such effects have been discussed in terms of both inhibitory and facilitatory mechanisms. For inhibitory effects, it has been shown that when NPs (noun phrases) in memory are of the same type, this can result in increased processing costs, and this effect has been argued to be due to partial activation of the illicit NP (similarity-based interference: Gordon et al., 2001). Thus, as Lee et al. (2006) discussed, rereading times of NP1 and NP2 in sentences of the form of (3) were longer after reading the verb region (i.e., retrieval point) when they were both descriptive nouns than when they were of different types (e.g., descriptive NP1 and a pronominal NP2), probably due to enhanced difficulty in establishing legitimate syntactic or semantic relations between NPs and the verbs. These results suggest that retrieval cues activated all memory representations of linguistic items with matching features in parallel and that the mis-retrieved NP interfered with the processing of a subject-verb dependency.

(3) Two NPs with the nominative caser marker

[MAINCLAUSE subject NP₁ [EMBEDDEDCLAUSE subject NP₂

EMB^verb] MAINverb].

Turning to facilitatory effects, the so-called attraction effect is a case where the activation of an illicit element has been argued to lead to mis-retrieval of that element instead of the target on a proportion of trials, leading to an overall facilitation of processing (Lewis and Vasishth, 2005; Wagers et al., 2009). The attraction effect is a “grammatical illusion” where processing difficulty due to ungrammaticality is reduced when a feature-matching distractor is mis-retrieved. For example, sentences in (c) and (d) in (4) are both ungrammatical as the licit subject (the key) mismatches the verb (were) in its number feature, but they differ from each other in that sentence (d) has a distractor NP with a plural number feature (e.g., the cabinets…were) while sentence (c) does not (e.g., the cabinet…were).

(4) Experimental sentences in Pearlmutter et al. (1999)

If processing difficulty of the subject-verb agreement in sentences like (4) is only affected by the licit subject (i.e., the key), then an equal level of processing difficulty is predicted for sentences (c) and (d) in comparison to their grammatical counterpart sentences (a) and (b), regardless of different number features of a distractor in (c) and (d). However, experimental results have shown reduced processing difficulty sentences like (d) in comparison to sentences like (c), suggesting that the processing difficulty of the subject-verb agreement is also affected by a distractor item which does not participate in the dependency (see also Nicol et al., 1997; Pearlmutter et al., 1999; Thornton and MacDonald, 2003; Drenhaus et al., 2005; Vasishth et al., 2008; Xiang et al., 2009; Dillon et al., 2013; Tanner et al., 2014; Lago et al., 2015; Parker and Phillips, 2017; for related effects in production, see Bock and Miller, 1991; Bock and Cutting, 1992; Bock and Eberhard, 1993; Vigliocco and Nicol, 1998; Hartsuiker et al., 2001; Haskell and MacDonald, 2003; Thornton and MacDonald, 2003; for related inhibitory effects, see Lewis, 1996; Gordon et al., 2001, 2004, 2006).

Similar effects have been found in Korean as well. Although Korean does not have rich verbal agreement, and so verbs in Korean do not agree with their subjects in person or number in most cases (Sohn, 1999), the subject honorific marker –si is an exception to this. –Si attaches to the stem of a verb, marking the speaker's respect for its agreeing local subject as in (5). Accordingly, –si can only occur with a honorifiable NP such as grandpa, uncle or teacher but not with nouns such as kid or burglar, which would be regarded as denoting people who are low in their perceived social status as shown in (6). In certain aspects, the –si subject-verb agreement in Korean is different from the number or person subject-verb agreement in English, given that –si agreement is pragmatically motivated, and given that the use of –si is optional such that its omission does not render the sentence unacceptable as shown in (5). Yet, in a previous study, Kwon and Sturt (2016b) showed that retrieval processes underlying the processing of –si agreement in Korean are similar to those of the number or person subject-verb agreement in English. Thus, in a sentence where a second NP forms a dependency with an embedded verb as in (7), a distractor with a matching honorific feature (e.g., chair) has been shown to reduce the processing difficulty due to honorific feature mismatches between a verb and its licit subject (e.g., embedded verb with –si and Inho, personal names in Korean, without honorific feature; Kwon and Sturt, 2016b).

yes

In many cases, attraction effects have been found in ungrammatical sentences at a relatively late processing stage after the ungrammaticality of sentences is detected (Pearlmutter et al., 1999; Wagers et al., 2009; Dillon et al., 2013; Lago et al., 2015; cf. Van Dyke, 2007). Given this, and given the fact that these studies did not find reliable evidence of attraction for grammatical sentences, it has been proposed that the attraction effects arise as an error-driven process, whereby distractors are retrieved as a repair strategy (cf. Wagers et al., 2009). However, the effect of a distractor found in Kwon and Sturt (2016b) is only partially compatible with this proposal. While the grammaticality effect (i.e., cost of ungrammatical relative to grammatical conditions) indeed preceded the effect of a distractor in the eye-movement record, the effect of the distractor was not limited to ungrammatical sentences. Instead, the processing of grammatical sentences was also affected by a distractor, leading to longer reading times when the distractor (e.g., Mary without honorific feature) mismatched the embedded verb in honorific features. These results were taken to suggest that the attraction effect is not an error-driven mechanism. Instead, Kwon and Sturt proposed that the effect of a distractor is likely to result from general working memory principles of activating of potential items in memory (see also Vasishth et al., 2008). Note, however, that the effect observed by Kwon and Sturt showed a facilitation for grammatical sentences where the distractor matched the honorific features of the verb, while the model proposed by Lewis and Vasishth (2005) would have predicted inhibition in this condition due to similarity-based interference. We return to this issue in the discussion.

Despite some differences in the interpretation, these results suggest the possibility that the processing of the retrieval process is modulated by memory representations of a distractor as well as that of a licit target item. For example, while a larger pool of items in memory does not affect the retrieval speed of a target item (McElree, 2000), assuming that retrieval of a linguistic item in memory is preceded by the reactivation of its memory representation which decays over time, it is possible that a distractor is more easily activated and thus interferes more with the processing of a dependency when it is closer to a retrieval point than when it is further away (Lewis and Vasishth, 2005)¹. In fact, manipulating the linear distance between the subject and the verb as well as the number feature of the subject and its intervening object, Kaan (2002) showed that participants better remembered the number features of the licit subject when the linear distance of the subject and the verb was shorter, suggesting the relevance of linear distance in the context of retrieval processes. Likewise, if retrieval is based on a parallel feature matching process, it is also possible that a distractor is more prone to affect retrieval of a target item when it has a larger number of matching features than when it has fewer. Supporting evidence for this latter observation comes from a study on the processing of reflexives, reported by Parker and Phillips (2017). Previous studies had shown that unlike number/person subject-verb agreement or negative polarity items, the processing of reflexives is not easily affected by a distractor (Sturt, 2003; Xiang et al., 2009; Dillon et al., 2013). For example, in Sturt's (2003) study of the processing of reflexives in sentences like (8), readers slowed down in the early parsing stages only when the licit subject (He/She) mismatched the reflexive in gender features. On the other hand, the gender feature of the intervening distractor (the surgeon) did not affect the processing until later processing stages, during the final interpretation (Sturt, 2003). However, Parker and Phillips (2017) showed that, when a distractor is associated with a larger number of matching (and probably semantically more salient) features [i.e., animacy as well as gender features in (9)], the processing of an emphatic reflexive is affected by a distractor in a similar manner to that of number/person agreement. Thus, the processing difficulty due to animacy/gender mismatches in a dependency involving a reflexive (e.g., the discovery-himself) was significantly reduced when there was a feature matching distractor (e.g., the researcher). These results suggest a possibility that retrieval is sensitive to memory representations of a distractor.

(8) He/She remembered that the surgeon had pricked himself/herself with a used syringe.

(9) The doctor/discovery that the researcher/report described meticulously was certified after debunking the urban myth himself in the new scientific journal.

Thus, in the experiments described below, we aimed to further investigate the effect of memory representations of a distractor on retrieval. We examined the effect of proximity by manipulating the linear distance of a distractor to a retrieval point, and the effect of the degree of similarity by manipulating the number of matching features between retrieval cues and a distractor, across the three experiments. To this aim, we ran eye-tracking studies of the processing of a subject-verb –si agreement using subject control, object control, and center-embedded clause constructions in Korean. The target stimuli of Experiments 1 through 3 are schematically presented in sentences (10), (11), and (12) respectively. As illustrated in these sentences, the structurally legitimate subject NP for the embedded verb differs depending on the sentence type. For example, in the subject control construction in Experiment 1, NP1 is the licit subject for the embedded verb, while NP2 is the licit subject in the object control construction in Experiment 2 and in the center-embedded construction in Experiment 3. Thus, in Experiment 1, NP2 is a distractor (i.e., NP in the dotted square), while in Experiments 2 and 3, NP1 is. Accordingly, the distractor linearly intervenes in the subject-verb dependency in Experiment 1 but it does not in Experiment 2 or Experiment 3. Thus, the comparison of the results of Experiment 1 and 2 could reveal the effect of proximity of a distractor to its retrieval point. If a distractor whose memory representations are more highly activated is more likely to lead to interference or attraction, then it can be predicted that an effect of a distractor will be stronger in Experiment 1 than in Experiment 2.

yes

On the other hand, the comparison of Experiments 2 and 3 could reveal the effect of the number of matching features of a distractor. Previous production studies have shown that similarity in case marking of the licit subject and a distractor led to increased attraction effects (Hartsuiker et al., 2001). Likewise, assuming that case information is a retrieval cue in comprehension, the morphology on the verb in Experiment 2 encodes a retrieval cue of dative case for its (overt) subject argument, while in Experiment 3, the retrieval cue is for nominative case. Thus, the distractor in Experiment 3 has more retrieval cue features than the distractor in Experiments 1 or 2. If a greater degree of feature overlap between distractor and retrieval cues leads to greater interference or attraction, then it can be predicted that the effect of the distractor will be stronger in Experiment 3, where the distractor matches the case retrieval cue, than in Experiment 2, where it does not. Thus, comparison of the general patterns of the results of these experiments would further our understanding of how memory representations of a distractor affect the retrieval processes, leading to a fuller understanding of the mechanisms underlying the retrieval processes in general. Below, we first report these three eye-tracking experiments, and then go on to present cross-experiment comparisons to address these questions.

Experiment 1

Experiment 1 investigated the processing of subject-verb honorific agreement in the subject control construction with the suffix –keyss “will” or “plan to” in Korean. In this particular construction, as used in Experiment 1, the embedded verb, marked with –keyss, indicates that the matrix subject (NP1) is the controller for PRO, as shown in (13). The construction is roughly translated as the subject “planning to” execute the action predicated in the embedded verb. Thus, the dependency is formed between NP1 and the embedded verb (via PRO), as illustrated in (10) above. On the other hand, the dative marked indirect object cannot serve as a controller although it is linearly closer to the embedded verb, intervening with the subject-verb honorific agreement.

(13) Subject control construction used in Experiment 1

[NP1_i-NOM NP2_k-DAT

[PRO_i leave-HON-SUBJECT.CTRL-DECL-comp] said]

‘NP_i told NP2_k that PRO $_{i /^{*} k}$ will leave.'

Personal names in Korean do not have honorific features. Thus, following Kwon and Sturt (2016b), to investigate the memory retrieval processes underlying the processing of a dependency, we manipulated the honorific features of NP1 and NP2 of the experimental sentences, by using either personal names (i.e., NH: non-honorifiable) or descriptive NPs (i.e., H: honorifiable)². On the other hand, the embedded verb is always marked with an honorific marker –si (see R5 in Table 1 below). Accordingly, there are two congruous conditions (i.e., H_NP1-NH_NP2 and H_NP1-H_NP2) and two incongruous conditions (i.e., NH_NP1-H_NP2 and NH_NP1-NH_NP2), as shown in Table 1.

TABLE 1

Table 1. Example experiment item in Experiment 1: Subject control construction.

Since the honorific suffix –si should agree with the verb's subject in Korean, we predict that mismatched honorific features of the incongruous conditions with the NH subject (NH-H and NH-NH conditions) will incur processing difficulty (Kwon and Sturt, 2016b). Thus, the incongruous conditions will show longer reading times at the embedded verb marked with –si– when compared with their congruous counterpart sentences with the H subject (H-H and H-NH conditions). Crucially, although control information is accessed early during on-line sentence processing (Kwon and Sturt, 2014, 2016a), it has been also shown that the processing of subject-verb honorific agreement can be affected by a feature-matching yet structurally illicit distractor in a control construction (Sturt and Kwon, 2015). If so, we predict that the reading time penalty for the incongruous condition will be reduced in the NH-H condition, where the distractor matches the honorific features of the verb, compared to the NH-NH condition, where it does not, resulting in an interaction between the honorific features of the subject and the dative object at the embedded verb position. This interactive pattern would be consistent with previous studies investigating subject-verb number agreement in English (e.g., Wagers et al., 2009; Dillon et al., 2013).

Participants

Twenty eight native speakers of Korean (mean age: 23.46; range 19–27) participated in the study. At the time of the experiment, they were all either undergraduate or graduate students at Konkuk University, Korea and received KRW 10,000 per hour for their participation. They all had normal or corrected-to-normal vision. Written informed consent was obtained from all the participants.

Materials

Forty sets of experimental sentences like those in Table 1 were created. All the experimental sentences contained two NPs with a + or – value for the honorific feature. On the other hand, the embedded verb was always marked with –si, and this verb formed a dependency with one of these NPs.

Before the main experiment, we first conducted a norming study to control for the plausibility of the event described with an embedded verb with NP1 or NP2 as a potential subject. For example, for the sentences in Table 1 four norming sentences were created using NP1 or NP2 (H_NP1, NH_NP1, H_NP2, NH_NP2) and the embedded verb, where the verbs of the norming sentences were matched with these NPs in their honorific features, as shown in (14) and (15). Thirty-two native Korean speakers participated in the norming study, each receiving KRW 3,000. At the time of the study, they were undergraduate students at Konkuk University, Korea. The norming sentences were split into four lists based on a Latin-square design along with 40 filler sentences with similar complexity. They were pseudo-randomized such that no two sentences from the same condition appeared in a row. Participants were asked to rate the plausibility of the sentences on a scale of 1 (very unlikely) to 5 (very plausible). The rating results were then analyzed using Linear Mixed Effect Regression (LMER) analysis (Baayen, 2008; Baayen et al., 2008; Jaeger, 2008). Models were constructed with the maximal random effect structure and were only simplified when the model did not converge (Barr et al., 2013). The results showed that the plausibility of the four conditions did not significantly differ from each other (|t| < 0.14 for all comparisons: H-NP1 vs. NH-NP2, NH-NP1 vs. NH-NP2, H-NP2 vs. NH-NP2), with the mean ratings of 4.69 (se: 0.042), 4.71 (se: 0.04), 4.72 (se: 0.039), and 4.62 (se: 0.046) for the H-NP1 (e.g., teacher), NH-NP1 (e.g., chair), H-NP2 (e.g., Swujin), and NH-NP2 (e.g., Inho) conditions, respectively.

yes

Given the results of the norming study, the experimental stimuli were distributed over four lists based on a Latin square design, along with 80 filler sentences of similar length and complexity. No two experimental items from a same condition were presented in a row.

Procedure

There were three practice trials before the main experiment started. Participants' eye movements were recorded using an SR Research Eyelink 1,000 Plus eye-tracker at a rate of 1,000 Hz. Each recording session began with a calibration procedure, using a standard 9 point calibration routine before the experiment started, and recalibration was performed whenever necessary throughout the experiment. In each trial, a participant was asked to fixate on a black square on the left side of the screen, where the first character of the upcoming sentence would be presented. When a participant's fixations were successfully detected on the black square, the square was automatically replaced by the experimental stimuli. Participants were asked to read the sentences at their natural speed and answered a yes/no comprehension question for all the sentences. The comprehension questions probed general understanding of the sentences. For example, for the H-H condition sentence in Table 1, “Will the teacher wait around the event place?” was asked. The experiment took about 40 min.

Data Analysis

For data analysis, following standard eye-tracking data analyses (Rayner, 1998; Sturt, 2007) we first merged short fixations under 80 ms into longer fixations within the distance of the visual angle of 0.05. If there was no such fixation, we removed the short fixations. We also removed fixations longer than 1,200 ms. This procedure affected 1.5% of the trials. Three eye-movement measures are presented. First pass reading times are the sum of all fixations in a target region from the first entry into the region until leaving the region either to the left or right. Go-past times (also called regression path duration) are the sum of all fixations on a given region from the first entry into the region from the left until leaving it to the right. Total time is the sum of all fixations in a given region. For First-pass reading time and Go-past times, we excluded the trials in which a target region was not fixated on in initial reading. For Total Time, we excluded the trials in which the target region received no fixation at all. The proportion of missing data points, due to zeros or track losses were less than 1%.

Statistical analyses were conducted for Region 5 as defined in Table 1, as the region is critical for retrieval processes. The region included the embedded verb and its following adverbial word (mean length = 2.5 syllables). To lower the rate of false positives due to multiple comparisons (von der Malsburg and Angele, 2017), we report statistical analyses for only one region, while reporting means for all regions. We first log-transformed reading times, and the resulting reading time data were analyzed based on Linear Mixed Effect Regression (LMER) analysis (Baayen, 2008; Baayen et al., 2008; Jaeger, 2008), using the lme4 R package (Bates et al., 2015; version 1.1–8). The comprehension accuracy rates were analyzed using a generalized LME model with a binomial distribution. The regression models incorporated two fixed-effect factors (the honorific features of the mrain and the embedded subject: H vs. NH), their interaction and crossed random effects for participants and items. The fixed-effect factors were coded numerically using sum coding, with the two levels of each factor coded as 1 and −1. Models had the maximal random effect structure whenever possible, including both intercepts and slopes, and were only progressively simplified when the model did not converge (Barr et al., 2013). In case of non-convergence, we simplified the model by backwards elimination, following the hierarchy principle, such that the interaction slope parameters were removed, and convergence checked, before attempting to remove either of the random main effect parameters. Also, at each stage of model simplification, convergence was checked both including and excluding random correlation parameters. Random slope parameters corresponding to fixed-effects are reported in in the “slope” column of Table 4 if they were included in the model for participants or items. Table 4 also shows coefficients, standard errors and t-values (z-values for the logit model) for each fixed effect and interaction from the analyses. P-values were obtained using LmerTest (Kuznetsova et al., 2017), and were corrected for multiple comparisons (9 comparisons: three eye-tracking measures × two main effects and one interaction) using Holm's correction (Holm, 1979; Abdi, 2010). For the analysis of comprehension accuracy based on a binomial logit model, p-values were calculated from the Z score, and were also corrected for multiple comparisons (three comparisons: two main effects and one interaction) using Holm's correction. Finally, planned (paired) contrasts were made based on the Tukey test (using the glht function of multcomp package: Hothorn et al., 2008; version 1.4–1) in R (R Core Team, 2018).

Results and Discussion

Comprehension accuracy and mean reading times for each condition are given in Tables 2, 3, respectively. Statistical analysis results for reading time measures are given in Table 4.

TABLE 2

Table 2. Mean comprehension accuracy rates in Experiment 1.

TABLE 3

Table 3. Means (and standard errors), aggregated by participants, for first pass, go-past, and total times in Experiment 1.

TABLE 4

Table 4. Generalized Linear Mixed Effects results for reading times in Experiment 1.

Region 5 (the critical embedded verb and the spill over region; ‘wait-HON again')

At R5, there was a main effect of NP1 with the (ungrammatical) NH_NP1 conditions taking longer to read than the (grammatical) H_NP1 conditions. The effect was marginal in First pass times and significant in Go-past and Total times. In addition, there was a marginal main effect of NP2 in Total times with longer reading times for the H_NP2 condition than for the NH_NP2 condition.

The grammaticality effect (i.e., the main effect of NP1) suggests that the subject control information was accessed from an early processing stage, allowing, and constraining dependency formation between the verb and its licit subject throughout the various processing stages. On the other hand, the marginal effect of NP2 in Total times suggests a tendency for a distractor with matching features to be activated regardless of grammaticality of the target sentences. For both grammatical and ungrammatical sentences, a distractor matching the verb in the honorific feature led to a slow-down in Total Time, relative to conditions where the distractor mismatched, presumably reflecting a later stage of processing.

The grammaticality effect shows that readers were sensitive to the relevant control information in forming the dependency. On the other hand, the (marginal) effect of the distractor was different from the so-called “attraction effect,” where facilitation is limited to the ungrammatical conditions (Wagers et al., 2009). Instead, the direction of the effect was such that the distractor interfered with and slowed down reading, when it matched the honorific features of the verb, regardless of the grammaticality of the sentence. We note that this effect is different from the general pattern observed by Kwon and Sturt (2016b), where a matching distractor tended to facilitate processing. We return to this point in the general discussion.

In summary, in Experiment 1 the grammaticality effect preceded any effect of the distractor. That is, while the grammaticality effect was found both in Go-past and Total times on the critical verb, the effect of the distractor was found only in Total Time. In addition, the marginal main effect of the distractor suggests that the distractor may affect the processing of the subject control construction regardless of grammaticality of the target sentences.

Experiment 2

Experiment 2 investigated the processing of subject-verb honorific agreement in the object control construction with –la in Korean. The embedded verb marked with –la signals that the indirect object (NP2) marked with a dative case marker is the licit controller for PRO, as shown in (16). Thus, the dependency is formed between the dative marked NP2 and the embedded verb (via PRO), as illustrated in (11) and (16). On the other hand, the nominative marked main clause subject (NP1) cannot serve as a controller.

(16) Object control construction used in Experiment 2

[NP1_i-NOM NP2_k-DAT

[PRO_k leave-HON-OBJECT.CTRL-comp] said]

‘NP_i told NP2_k to PRO $_{^{*} i / k}$ leave.'

As in Experiment 1, we predicted that mismatched honorific features of the incongruous conditions with the NH subject (NH dative object conditions: H_NP1-NH_NP2 & NH_NP1-NH_NP2) would incur processing difficulty (Kwon and Sturt, 2016b), leading to longer reading times at the critical region in these conditions than in the congruous conditions (H-H & NH-H conditions). In addition to the grammaticality effect, we also predicted an effect of the distractor. In particular, if interference from a distractor is affected by memory representations of a distractor, it is likely that a memory representation that is more highly activated at the point of retrieval will be more likely to lead to interference. If so, we predict that the effect of a distractor will be weaker in Experiment 2 than in Experiment 1, given that the distractor position is further away from the embedded verb and so its memory representation could be more decayed at the retrieval point in Experiment 2.