How Can Agreement Errors Be Measured?

The most direct way to measure agreement errors in comprehension is by ask-ing participants to give a grammaticality judgment. Agreement errors result in judgment errors: An ungrammatical sentence will be incorrectly judged as ‘gram-matical’ when the agreement violation is overlooked and accordingly, an actually grammatical sentence will be judged as ‘ungrammatical’ in case of an apparent agreement violation. In addition to grammaticality judgments, agreement errors can also be observed in a more indirect way. Agreement violations lead to pro-cessing difficulties reflected in increased response times for various tasks and typ-ical brain responses. For agreement errors of type (ii), i.e., when the agreement violation is overlooked, the otherwise expected reflexes might fail to appear. Ac-cordingly, agreement errors of type (iii) are expected to elicit the violation-related pattern despite the absence of a real agreement violation.

Consider behavioral measures first. Readers/hearers usually notice an agree-ment violation, especially when asked to provide a grammaticality judgagree-ment (e.g., Blackwell et al., 1996; Kail and Bassano, 1997). Agreement violations are de-tected not only reliably but also fairly quickly. In an experiment using an incre-mental judgment procedure, participants rejected sentences containing an agree-ment violation earlier than sentences with another error type (Blackwell et al., 1996). Attraction errors as introduced in chapter 3, have been demonstrated to affect the accuracy to detect an agreement violation as well as response times. In contexts which elicited an agreement error in production, participants produced more judgment errors (Kaan, 2002; Nicol et al., 1997) and/or needed more time for their judgment (Kail and Bassano, 1997; Staub, 2009; Nicol et al., 1997). Agree-ment violations have also been investigated by means of experiAgree-mental methods not drawing the participants’ attention to grammaticality. Using a combination of

a comprehension and a production task, Blackwell et al. (1996) found that partic-ipants rejected sentences fragments containing an agreement violation in almost 100% of the cases. As soon as they encountered the critical word (the auxil-iary for subject–verb agreement, and, respectively, the noun for determiner–noun agreement) they either responded with ‘Can’t complete’ or changed the fragment in such a way that a grammatical sentence resulted. The performance was lower when the agreement violation occurred early in the sentence—slightly in the case of subject–verb agreement and substantially in the case of determiner-noun agree-ment. Note that a determiner-noun sequence involving a feature conflict still al-low for a grammatical continuation without changing the fragment. The sequence several sailor, for example, turns out to be grammatical when the next word is uniforms. And indeed, participants sometimes came up with such continuations.

Later in the sentence, such continuations are harder to find because the preceding context imposes syntactic as well as semantic restrictions. Furthermore, agree-ment violations have been shown to slow down readers’/listeners’ performance time in tasks like sentence matching (Freedman and Forster, 1985), lexical deci-sion (Jakubowicz and Faussart, 1998), verb naming (Sevald and Garnsey, 1995) or word-monitoring (Haarmann and Kolk, 1994; Tyler, 1992; Vos et al., 2001).

One might object that these effects are due to the additional task used in these experiments, but self-paced reading and eye-tracking studies revealed comparable effects—increased reading times and/or an increase in regressive saccades. In a self-paced reading experiment, Pearlmutter et al. (1999) found increased reading times for the spill-over region of the incorrect verb in ungrammatical sentences.

Furthermore, the study demonstrated that attraction elicits increased reading times in grammatical sentences and reduces the reading-time penalty in ungrammati-cal sentences. (Pearlmutter et al., 1999) found a comparable pattern in an eye-tracking experiment employing the same material. In addition the reading time effect, grammaticality also affected the probability of regressive saccades. On en-countering the agreement violation, some participants spent more time in the verb region while others performed regressive saccades. These reflexes of processing difficulty were reduced in attraction configurations (e.g., the key to the cabinets were rusty). At the same time, attraction hampered the processing of grammatical sentences which showed the same pattern as ungrammatical sentences although less pronounced. This finding suggests that participants struggled with an illusion-ary agreement violation. Pearlmutter (2000) could replicate the attraction effect in grammatical sentences whereas Wagers et al. (2009) found an attraction effect only in ungrammatical sentences. In will come back to this difference between grammatical and ungrammatical sentences in section 4.4.3.4 and in chapter 7.

Studies using event-related potentials (ERPs) have shown that syntactic vio-lations including agreement viovio-lations elicit specific brain responses that differ qualitatively from the brain response to semantic and pragmatic anomalies. Kutas

and Hillyard (1983) report that both agreement violations like *As a turtle grows its shell grow too and semantic incongruities such as He spread the warm bread with socks are reflected in a negativity peaking around 400 milliseconds after the onset of the critical item but with different scalp distributions. The negativity as-sociated with agreement violations has a left anterior distribution (hence termed Left Anterior Negativity or LAN for short) while the negativity in response to semantic anomalies has a central-parietal topography. The central-parietal neg-ativity related to semantic violations was first observed in (Kutas and Hillyard, 1980) and is commonly termed N400 in reference to its polarity and latency.² Subsequently, the finding of a LAN in correspondence to an agreement violation could be replicated in numerous studies and generalized for various languages (for English see Coulson et al., 1998b; Osterhout and Mobley, 1995; for Dutch see Ha-goort and Brown, 2000; Severens et al., 2008; Vos et al., 2001; for German see Burkhardt et al., 2007; Roehm et al., 2005; Rossi et al., 2005). In these studies, the LAN is typically accompanied by a late positivity.³ Some studies found just a late positivity but no LAN effect (e.g., Hagoort et al., 1993; Kaan et al., 2000;

Münte et al., 1997; Nevins et al., 2007; Osterhout and Holcomb, 1992; Osterhout et al., 1996; Osterhout and Nicol, 1999). Besides, both the LAN and the late positivity—commonly labeled P600 (Osterhout and Holcomb, 1992)⁴—are not restricted to agreement violations but occur with other (morpho-) syntactic viola-tions as well, e.g., subcategorization errors, auxiliary selection errors, verb form errors, word category errors, phrase structure violations, and also in grammatical but resource demanding constructions such as filler–gap dependencies, e.g., wh-constructions (cf. Felser et al., 2003; Fiebach et al., 2001, 2002; Kaan et al., 2000;

Kluender and Kutas, 1993; King and Kutas, 1995; Phillips et al., 2005).⁵ In addi-tion, LAN and P600 effects are reported for non-linguistic manipulations as well.

LAN effects, for instance, have been observed in response to cognitive sequenc-ing (e.g., Hoen and Dominey, 2000) while P600 effects are attested in response to harmonic incongruities in music (Patel et al., 1998) and rule-violations in arith-metic (Núnez-Peña and Honrubia-Serrano, 2003). The latter findings suggests that the two components are not specific responses to morphosyntactic illformedness and not even language specific but are related to the processing of rule-based

se-2Note that it is not the occurrence of the components which reflects a violation but rather the component’s parameters (latency, amplitude etc).

3The original study by Kutas and Hillyard ignored late time windows, but later reexamination of the data revealed a late positivity (cf. Kutas and Van Petten, 1994).

4Alternative terms are ‘syntactic positive shift’(SPS) (Hagoort et al., 1993), late positivity’

(Bornkessel and Schlesewsky, 2006).

5Again, the findings are not consistent. Some studies report a LAN or some other anterior negativity in the same time window (e.g., Kluender and Kutas, 1993; King and Kutas, 1995) while others report a P600 (e.g., Kaan et al., 2000) and yet others report a biphasic pattern with a LAN and a subsequent P600 (Phillips et al., 2005).

quences and perhaps working memory in a more general way. Strikingly, the cited examples for non-linguistic LAN/P600 effects are still related to syntax, though on a more abstract level.⁶On the other hand, semantic factors were also observed to affect the P600, but mainly in relation to thematic roles. Hence, as (Kuperberg, 2007) argues, the semantic effects may in fact result from interactions including semantic and syntactic factors. Taken together, the function of the two compo-nents is still an issue of debate (for recent reviews see Bornkessel-Schlesewsky and Schlesewsky, 2009; Kutas et al., 2006; Vos et al., 2001). Part of the solution to this problem might be to acknowledge that at least the P600 is not a unitary component but rather a family of components as now commonly assumed for the P300 with which the P600 may form a larger family. In any case, this discussion goes far beyond the scope of the thesis. For the moment, I want to review the suggested functions of the P600 in the light of agreement errors. The P600 has been discussed to reflect syntactic integration difficulty (Featherston et al., 2000;

Kaan et al., 2000), syntactic violation detection (Coulson et al., 1998a), subse-quent repair processes (Friederici et al., 1996; Münte et al., 1997; Osterhout et al., 1994) as well as conflict monitoring (Kolk et al., 2003; van Herten et al., 2006;

Vissers et al., 2006). Observing an increased P600 amplitude in the presence of an agreement violation is compatible with all these proposals. First of all, agreement checking is part of syntactic processing. Furthermore, a feature value mismatch between agreement controller and agreement target reasonably increases the inte-gration difficulty for the target word. And finally, it represents a syntactic violation the detection of which might initiate a reanalysis.

Attraction has been shown to affect the amplitude and latency of the P600 component (Kaan, 2002; Severens et al., 2008; Molinaro et al., 2009) as well as the emergence of a negativity in the 400ms time window (Severens et al., 2008).

Consider first the negativity which Severens et al. classify as an instance of a N400 because of its central-parietal distribution. In accordance with prior re-search, ungrammatical sentences elicited a negativity in response to the incorrect verb. Crucially, the N400 was visible only in sentences with a number match be-tween controller and distractor. This finding is compatible with the assumption that attraction hides an agreement violation. Severens and colleagues offer an al-ternative explanation. Following Kuperberg (2007) they assume two processing streams in comprehension: a syntactically shallow stream that is based on se-mantic memory and combinatory stream that uses multiple constraints including syntactic constraints. The N400 is assumed to reflect to first, semantically based stream. Under the assumption that this shallow processing streams builds up ex-pectations regarding the verb number based on the number specification of the first

6For a comprehensive discussion of similarities and differences between music and language see Patel (2008).

NP, an incorrect verb creates a conflict which in turn elicits an N400. Severens et al. suggest that the absence of the N400 in sentences with a mismatching dis-tractor results from an attempt to repair the conflict which is responsible for the subsequent P600. In sentences with two matching NPs, the shallow processing stream rejects an ungrammatical sentence since there is no NP available sharing the verb’s number specification. In ungrammatical sentences with two NPs dif-fering in their number specifications, the combinatory processing stream initiates a reanalysis. Note that Severens at al. seem to assume that the detection of the number conflict has no visible effect when reanalysis is promising. Even under the assumption that the two processing streams work in parallel, this is surprising given the difference in timing. The argumentation for reanalysis is, however, in line with studies reporting a P600 without a preceding violation-related negativ-ity (Hagoort et al., 1993; Kaan et al., 2000; Münte et al., 1997; Osterhout and Holcomb, 1992; Osterhout et al., 1996; Osterhout and Nicol, 1999).

The issue of reanalysis brings us to the second component associated with syntactic anomalies, the P600. The corresponding results are mixed. The evalu-ation of the source of the discrepancies is difficult since the studies involve dif-ferent languages—Dutch (Kaan, 2002; Severens et al., 2008) and Spanish (Moli-naro et al., 2009)—, different constructions—modifier attraction (Moli(Moli-naro et al., 2009; Severens et al., 2008) and object attraction (Kaan, 2002)—, and different tasks—reading (Molinaro et al., 2009; Severens et al., 2008) and grammaticality judgments (Kaan, 2002). Furthermore, the comparison tricky because of differ-ences in the choice of the baseline. Molinaro et al. (2009), for instance, compare brain responses in trials with a number match between controller and distractor and trials in which the two NPs differ in their number specifications. A P600 was elicited in sentences containing a singular subject and a plural distractor com-pared to all other number combinations. Since the study included only grammat-ical sentences, it can hardly be compared to studies contrasting grammatgrammat-ical and ungrammatical sentences (Kaan, 2002; Severens et al., 2008). While Severens et al. (2008) report a P600 in the mismatch conditions—ungrammatical sentences elicited a more positive waveform than grammatical sentences—which is absent in the match conditions, Kaan (2002) found a P600 in both sentence types but with a smaller amplitude in ungrammatical sentences containing a singular subject and a plural distractor. In fact, Kaan (2002) reports a larger P600 in ungrammatical sentences with a singular subject and a singular distractor compared to all other number combinations. The absence of an attraction effect in plural sentences can be interpreted in terms of a singular–plural asymmetry as observed in production studies as well as in other comprehension studies (e.g., Molinaro et al., 2009).

Supporting evidence comes from the judgment data obtained in addition to the electrophysiological recording. Judgment errors were most common in sentences with a singular subject and a plural distractor. The accuracy for reverse number

combination was slightly above the corresponding match condition. A more chal-lenging result is the finding that grammatical sentences show no signs of attraction in the P600 component. Although the two mismatch conditions show slightly in-creased mean amplitudes in the P600 time window, they are far away from the ungrammatical conditions. Under the assumption that P600 effects reflect pro-cesses of reanalysis, attraction in ungrammatical sentences might be simply too rare to cause visible effects. Furthermore, Kaan only reports ERPs for sentences that were judged correctly. Thus, instances of attraction in grammatical sentences that lead to rejection rather than to initiate successfully reanalysis are not included in the analysis.

In addition to differences in amplitude, Kaan (2002) also reports a difference in latency. The onset of the P600 was delayed in any ungrammatical condition involving a plural NP—either the subject or the distractor or both. Kaan offers an explanation that covers both the singular–plural asymmetry and the difference between grammatical and ungrammatical sentences. Her account disregards at-traction and refers to semantic complexity instead. Additional semantic and dis-course processes related to plural NPs may consume processing resources which otherwise could be used for reanalysis in ungrammatical sentences. As a result, reanalysis can only start later and accordingly the onset of the P600 is delayed.

Finally, Kaan (2002) found an early positivity around 250 ms for grammat-ical sentences compared to ungrammatgrammat-ical ones when the subject was singular and the distractor was plural. In line with studies showing a response to syntac-tic anomalies in the same time window or somewhat later (Neville et al., 1991;

Mecklinger et al., 1995), Kaan argues that the early positivity reflects reanalysis following attraction and triggered by the verb. Occasionally the subject’s number specification may initially be retrieved as plural (attraction) but on encountering the singular verb corrected to singular (reanalysis).