Results JudgmentsJudgments

Two sentences were excluded from analysis because of coding errors. The per-centages of correct judgments for the remaining items are given in Table 5.2. For further analyses, responses were scored as 1 for correct judgments and 0 for in-correct ones and fed into a mixed logit model that included participants and items

as random effects and Controller, Distractor and Grammaticality as fixed effect.

Table 5.3 summarizes the relevant measures for fixed effects.

Table 5.2: Percentages of Correct Judgments in Experiment 2

Singular Controller Plural Controller

Match Mismatch Match Mismatch Mean

grammatical 92 (2.2) 80 (3.3) 89 (2.6) 67 (3.9) 82

ungrammatical 89 (2.6) 74 (3.6) 84 (3.0) 55 (4.1) 76

Mean 90 77 86 61

Note. Values enclosed in parentheses represent standard errors.

Table 5.3: Summary of Fixed Effects in the Mixed Logit Model for Judgments in Experiment 2 (Log-Likelihood=−544)

Estimate SE z p

(Intercept) 1.686 .189 8.93 < .001 **

Grammaticality −.452 .168 −2.69 .007 **

Controller −.640 .168 −3.80 < .001 **

Distractor −1.410 .170 −8.31 < .001 **

Grammaticality×Controller −.167 .339 −0.49 .621

Grammaticality×Distractor −.023 .336 −0.07 .945

Controller×Distractor −.446 .336 −1.33 .184

Grammaticality×Controller×Distractor −.032 .672 −0.05 .963

**p<.01

The model indicates significant main effects of Grammaticality, Controller and Distractor: Judgments errors were more common in ungrammatical sentences than in grammatical sentences; plural sentences were more error prone than singu-lar sentences; the majority of judgment errors occurred in sentences with a number mismatch between controller and distractor. The numbers in Table 5.2 suggest some interactions—the effect of Distractor, for instance, seems to be bigger in ungrammatical sentences compared to grammatical ones and in plural sentences compared to singular sentences (see also Figure 5.1)—significant interactions are, however, absent in the model. This means that all three factors were effective independently of each other.

Attraction Rate

grammatical ungrammatical Error−rate difference 05101520253035

singular subject plural subject

Attraction Penalty

grammatical ungrammatical Response−time difference (ms) −50050100150

singular subject plural subject

Figure 5.1: Attraction rates and attraction penalty in Experiment 2

Judgment Times

Response times were corrected for outliers the same way as described for Ex-periment 1. Table 5.4 presents the resulting trimmed response times for correct judgments. For further analysis, a linear mixed-effects model was used with par-ticipants and items included as random effects and Grammaticality, Controller and Distractor as fixed effects. Table 5.5 lists the fixed effects and interactions.

Table 5.4: Mean Response Times for Correct Judgments (in ms) in Experiment 2 Singular Controller Plural Controller

Match Mismatch Match Mismatch Mean

grammatical 597 (23) 635 (29) 601 (24) 721 (38) 632

ungrammatical 702 (24) 655 (26) 806 (32) 752 (37) 728

Mean 649 645 700 735

Note. Values enclosed in parentheses represent standard errors.

On average, response times for correct judgments are about 100 ms longer in ungrammatical sentences compared to grammatical sentences. This means that participants needed more time to correctly reject an ungrammatical sentence than to accept a grammatical sentence. Closer inspection, however, reveals that this in-crease for ungrammatical sentences is almost absent in the mismatch conditions.

The model (Table 5.5) confirms this observation. Grammaticality is significant as a main effect but also in interaction with the factor Distractor. This

interac-Table 5.5: Summary of Fixed Effects in the Linear Mixed-Effects Model for Response Times in Experiment 2 (Log-Likelihood=−6672)

Estimate SE t pMCMC

(Intercept) 696.0 31.8 21.92 < .001 **

Grammaticality 88.0 17.8 4.94 < .001 **

Controller 70.7 17.7 3.99 < .001 **

Distractor 30.0 17.8 1.69 .098 +

Grammaticality×Controller 41.6 35.4 1.18 .238

Grammaticality×Distractor −125.4 35.4 −3.54 < .001 **

Controller×Distractor 49.9 35.3 1.41 .165

Grammaticality×Controller×Distractor −147.3 70.9 −2.08 .044 * Note. Only response times for correct responses entered analyses.

+ p<.1, * p<.05, **p<.01

tion also reflects the inconsistent effect of the factor Distractor which is further modulated by the factor Controller (see also Figure 5.1). The interaction involv-ing all three factors is significant. In grammatical sentences, a number mismatch between controller and distractor had no effect in sentences with a singular con-troller but resulted in longer response times in sentences with a plural concon-troller.

In ungrammatical sentences such a penalty is absent; instead, we see even a slight advantage for mismatch conditions, which is the same in singular sentences and plural sentences. As a main effect, Distractor is only marginally significant.

There was also a significant main effect of Controller. Mean response times for sentences with a plural controller are longer than mean response times for sen-tences with a singular controller (714 ms vs. 647 ms). This increase is present in all pairs except grammatical sentences in the match condition (597 ms vs.

601 ms).

5.3.1.3 Discussion

Experiment 2 has two main results: Sentences in which controller and distractor differed in number were more often judged incorrectly. Furthermore, plural sen-tences caused more processing difficulties than singular sensen-tences. The first effect attests attraction in the given relative-clause construction. Consistently, partici-pants produced more judgment errors whenever controller and distractor differed in number. Since most prior studies found only little attraction when the distrac-tor was part of a relative clause (Bock and Cutting, 1992; Bock and Miller, 1991;

Nicol, 1995; Nicol et al., 1997; Kail and Bassano, 1997 but see Solomon and Pearlmutter, 2004), this is an interesting result by itself. It suggests that relative clauses are not completely encapsulated from the embedding clause as one might

expect, e.g., based on binding data (cf. Lebeaux, 1991). Apparently, a distractor inside a relative clause can affect agreement processes in the embedding clause given the right properties. Note that the relative-clause construction in Experiment 2 differs from the relative clauses employed in the studies cited above. The syn-tactic connection between controller and distractor is much closer in Experiment 2 than in the previous studies. In contrast to previous studies, controller and distrac-tor are linked via the possessive relative pronoun located in the specifier position of the distractor NP. Moreover, the distractor NP occurs in a syntactically more prominent position than in the previous experiments where it was the object of the relative clause. In Experiment 2, in contrast, the distractor NP is the subject of the relative clause and it is located in the highest possible position, in SpecCP. Given that semantic integration has been shown to affects the incidence of interference (Solomon and Pearlmutter, 2004), the close semantic relation between controller and distractor resulting from the possessive structure may also contribute to the effect. Taken together, I conclude that a distractor inside a relative clause has a good chance to interfere in the agreement processes of the embedding clause if it is tightly linked syntactically and semantically to the agreement controller.

The attraction effect visible in the judgment data is not mirrored in response times. Apparently, attraction lead to judgment errors but did not consistently slow down correct judgments. Prima facie this suggests that the parser made no (or at least no visible) attempt to overcome the attraction error. This conclusion is reasonable for ungrammatical sentences: Attraction hides an agreement violation;

as a result, the sentence is perceived as grammatical and the parser has no mo-tive to revise its error. The slight speed-up in judgments times for ungrammatical sentences with number mismatch between controller and distractor supports this suspicion. The situation is different for grammatical sentences in which attraction results in an apparent agreement violation. Here the parser has a reason to ini-tiate a reanalysis. And in fact, response times in grammatical sentences show a penalty in the mismatch conditions. This penalty is rather small in grammatical sentences with a singular controller but substantial in grammatical sentences with a plural controller. Note that this asymmetry is parallel to the one found in error rates (though for error rates, the critical interaction failed significance). An attrac-tion penalty in terms of response times occurs in those sentences which have the highest error rates. This observation supports the suspicion that attraction-induced reanalysis is rare. Attraction is not a regular phenomenon but an exceptional er-ror; it occurs only from time to time. Furthermore, the existing evidence on the role of subject–verb agreement on the resolution of subject–object ambiguities im-plies that reanalysis would be hard to detect in the context of attraction. Locally ambiguous OS sentences disambiguated by number agreement cause a particu-larly strong garden path and are most of the time rejected in experiments making use of speeded-grammaticality judgments (cf. Bader and Meng, 1999; Meng and

Bader, 2000a). The high rejection rates indicate that the agreement violation re-sulting from an SO analysis does not always trigger reanalysis. Taken together, the instances of reanalysis following an attraction error might simply be too rare to leave a reliable mark in response times, at least when using a method like speeded-grammaticality judgments.

Turning back to judgment errors, it is striking that attraction occurred with both singular controllers and plural controllers. As Figure 5.1 illustrates, the at-traction effect was even stronger for plural subjects at least in ungrammatical sen-tences. Nevertheless, the mixed logit model used to evaluate the effects failed to show an interaction between any of the involved factors. Even if one is willing to accept an asymmetry (despite the lack of significance), one has to note that the asymmetry observed in Experiment 2 differs from the classical asymmetry in two respects. First, the asymmetry is non-strict. It is not the case that attraction oc-curs in one condition but is completely absent in the other condition. Instead, the observed asymmetry is an asymmetry with respect to the strength of attraction.

Second, the asymmetry is the reverse of the asymmetry in the classical instances of attraction: in Experiment 2, plural sentences exhibit higher attraction rates than singular sentences.

The incidence of singular attraction—whether resulting in no singular–plural asymmetry or a non-strict asymmetry—contrasts with previous research. Al-though the emergence or the strength of a singular–plural asymmetry varies across both languages and constructions, only few studies report a reversed asymme-try. In fact, a reversed asymmetry has been demonstrated for two cases (both in German): the production of sentences with an object distractor (Hemforth and Konieczny, 2003; Konieczny et al., 2004) and the comprehension of sentences with a modifier distractor (Hölscher and Hemforth, 2000). Object attraction may have a different source(cf. in Chapter 6); the comprehension study is also no convincing counterexample. The authors note themselves it suffers some method-ological problems (for details see Chapter 4). Note that Experiment 1 found the same singular–plural asymmetry that has been firmly established for German sen-tence production involving a distractor in modifier function. Under this perspec-tive, the attraction pattern in Experiment 2 is unexpected. Before drawing any conclusions based on this specific asymmetry consider the possibility that attrac-tion occurred twice in Experiment 2.

In addition to an attraction effect, Experiment 2 shows a plural effect—a plu-ral penalty. Processing difficulties for pluplu-ral sentences were visible both in error rates and in response times for correct judgments, especially in ungrammatical sentences. In an activation-based model this finding is explicable by decay if only plural is represented explicitly. A plural specification—a privative feature as in (cf. Eberhard, 1997) or a positive value of a continuously valued number feature as in Eberhard et al. (2005)—means a specific activation. Activation decays over

time; when the activation level falls below a threshold value the plural feature is no longer accessible. The loss of the plural specification results in an appar-ent agreemappar-ent violation since the now singular NP conflicts with a plural verb.

Accordingly, an actual agreement violation remains unnoticed in ungrammatical sentences containing a singular verb. As a result, participants produce a judgment error because they perceive a grammatical sentence as ungrammatical and vice versa.

An alternative explanation would be to blame the semantic complexity for the increased processing costs. Plural referents are semantically more complex than singular referents (Kamp and Reyle, 1993). At the current point, it is hard to de-cide between these two competing explanations for the plural effect. A first hint comes from the comparison of grammatical and ungrammatical sentences. The plural effect was virtually restricted to ungrammatical sentences. This difference between grammatical and ungrammatical sentences can easily be explained under the decay hypothesis: In grammatical sentences, the loss of the plural specification leads to an apparent agreement violation which might be occasionally overlooked or repaired via reactivation of the plural NP. In ungrammatical sentences, on the other hand, the loss of the plural specification leads almost inevitably to a judg-ment error since it hides a real agreejudg-ment violation. It may also be conceivable that the semantic complexity carries more weight in ungrammatical sentences, se-mantic complexity may boost the already existing processing difficulties. A more reliable test for differentiating between the two hypotheses concerns the distance between the encounter of the plural NP and the moment it is needed to be avail-able for agreement checking. Under the complexity hypothesis the plural effect should be independent of the distance between subject and verb whereas the decay hypothesis predicts that the plural effect is enhanced the more material intervenes.

Note that at this point I do not specify what ‘material’ means. The crucial point for the decay hypothesis is the decay of activation until the critical verb is reached whether decay is just a matter of time in which case material would be understood in string terms or due to the processing that has to be done in the interim in which case material would be understood in terms of some complexity measure. The finding that Experiment 1 where subject and verb were not separated by a relative clause exhibited only spurious signs of a plural effect argues for the decay hypoth-esis. I will come back to this issue in the discussion of Experiment 5 where subject and verb are also much closer to each other than in the present Experiment.