4.1 Problems with existing accounts

As can be seen, the two major psycholinguistic theories of attraction outlined in the preceding section have evolved to meet the interface facts of agreement by becoming very much more nuanced, and also by de facto becoming not too dissimilar. Marking and Morphing captures inescapable semantic interference by allowing semantics to sanction referents at the initial marking phase as either singular or plural at the phrase level, even before the phrase is assembled. Encapsulation is taken to prevail during the assembly phase itself. Maximal Input captures morphological effects by recognising that S–V agreement operations unfold primarily in the home domain of grammatical encoding, after conceptual inputs and before phonological spell-out. Semantic effects at this stage are explained by simply relaxing the level of encapsulation of each phase, allowing conceptual influences from the previous encoding cycle if beneficial and penalising them if detrimental (Berg Reference Berg1998). However, there are a number of problems with both accounts. In the first place, Marking and Morphing, being a formal account in origin at least, is easier to falsify since formal accounts have typically neater predictions. Two sets of data in particular are hard to reconcile with this model. The first set of data concerns the fact that attraction has been found to exist when object NPs (both clitics and full NPs) intervene between subject NPs and their agreeing verbs (Fayol et al. Reference Fayol, Largy and Lemaire1994; Hartsuiker et al. Reference Hartsuiker, Antón-Méndez and Zee2001; Hemforth & Konieczny Reference Hemforth, Konieczny, Alterman and Kirsh2003; Franck et al. Reference Franck, Lassi, Frauenfelder and Rizzi2006, Reference Franck, Soare, Frauenfelder and Rizzi2010). Since marking affects the entire referential phrase and the object NP is another referential phrase, the percolation path envisaged to account for such effects would actually be tantamount to a relaxation of the notion of fully encapsulated cycles (making Marking and Morphing even more similar to Maximal Input).Footnote ¹⁴ The second set of data concerns the fact that, the encapsulation part of Marking and Morphing rests on two series of findings: of semantic effects if these reside in the head noun or the entire NP, and of morphological effects if these reside in the local noun. However, Hupet, Fayol & Schelstraete (Reference Hupet, Fayol and Schelstraete1998) and Thornton & MacDonald (Reference Thornton and MacDonald2003) have found semantic effects by manipulating the plausibility of the verb relative to the two nouns in the complex NP. For instance, in an experiment where both nouns could be plausible passive subjects (e.g. the album by the classical composers... BE praised) or only the head noun could be so (the album by the classical composers... BE played), there were more agreement errors when both nouns were plausible subjects than when only the head noun was plausible (Thornton & MacDonald Reference Thornton and MacDonald2003).Footnote ¹⁵ There are two problems with this finding: firstly, it indicates that the local noun can exercise semantic influence (perhaps simultaneous with the morphing process for which Marking and Morphing has provided convincing evidence); secondly, and maybe more importantly, effects are obtained by manipulating a position (the verb) where, in Marking and Morphing terms, all kinds of agreement decisions (reconciliation) are supposed to have been left behind and the verb's agreeing form is supposed to be merely inhereted via control (Wagers et al. Reference Wagers, Lau and Phillips2009). In fact, this attested role of the verb is consonant with two other facts of grammar: (i) inheritance (or percolation or copying) is problematic in the case of absent controllers (see footnote 5 above); and (ii) often the mereology feature of a subject NP (like a collective) cannot be determined until the verb is chosen, as in, for example, the committee is old vs. the committee are old (in the second case, for instance, the committee may be newly composed of people who happen to be old; see Sauerland & Elbourne Reference Sauerland and Elbourne2002). In such cases, even truth conditions are affected.

4.2 Production and comprehension

Models of attraction developed with production in mind, as research in comprehension has only barely begun. There is in fact no theory of attraction in comprehension yet. However it is evident that in reversing the direction of encoding the interference dynamics may change accordingly: in production, message features may spill into form (this constituting a whole research agenda); in comprehension this is less likely (Berg Reference Berg1998; Acuña-Fariña Reference Acuña-Fariña2009). A positive aspect about the new, comprehension studies is the fact that they make it possible to study attraction in grammatical sentences (Wagers et al. Reference Wagers, Lau and Phillips2009; Acuña-Fariña, Carreiras & Meseguer Reference Acuña-Fariña, Carreiras and Meseguer2011). One of the first studies of this kind was conducted by Nicol, Forster & Veres (Reference Nicol, Forster and Veres1997) . By using a maze task (a timed, incremental sentence-construction task) and a whole-sentence grammaticality judgement task, they showed that a mismatching plural local noun interfered with the agreement process in the form of increased processing time on the verb. The authors suggested that a plural feature percolates from the local noun to the head noun in comprehension just as it is supposed to do in production. However, there are obvious limitations to the methodology employed. A tighter control of methodological issues was provided by Pearlmutter, Garnsey & Bock (Reference Pearlmutter, Garnsey and Bock1999), with two self-paced reading studies and one eye-tracking study in which they replicated the mismatch effect. Interestingly, eye-tracking revealed no effects for first pass reading times, but significantly more regressions from the verb region for mismatching local nouns. Additionally, Thornton & MacDonald (Reference Thornton and MacDonald2003) also replicated the plausibility effects of their production study (recall the album by the classical composer was praised/played mentioned above) in an accompanying self-paced reading experiment in which they obtained longer reaction times (RTs) at the verb when both nouns were plausible subjects than when only the head noun was plausible. The mechanism suggested for adapting these congruent data from comprehension to the much more established data from production is a ‘head overwriting’ operation whereby the feature on the legal head noun is ‘overwritten’ by the feature of the attractor (Pearlmutter et al. Reference Pearlmutter, Garnsey and Bock1999; Wagers et al. Reference Wagers, Lau and Phillips2009).

However, Wagers et al. (Reference Wagers, Lau and Phillips2009) have reasoned that there is in fact no attraction in comprehension and that the effects obtained in recent research are spurious: they are not syntactic but morphological, a side-effect of the fact that plurals are harder to process than singulars, on their own. In their view, the added difficulty of the plural ‘is particularly problematic for on-line experiments in which the attractor region is immediately adjacent to the verb, as the strongest effects of plural processing complexity would likely be felt here’ (Wagers et al. Reference Wagers, Lau and Phillips2009: 210). The theoretical significance of this methodological issue is important: in percolation/copying terms, attraction unfolds independently of the verb due to architectural requirements affecting the way referential NPs are projected. If attraction is due to the faulty representation of subject number, then not registering it in grammatical sentences would mean that it is the to-be-produced verb that causes it in the production studies. Wagers et al. point out that this is the case and that it suggests that attraction is due to a faulty access to the NP's features at the moment of reactivating them at the verb (not to automatic percolation in a step-by-step syntactic derivation of the NP part of structure). Partly aware of that, Acuña-Fariña et al. (Reference Acuña-Fariña, Carreiras and Meseguer2011) conducted an eye-tracking experiment in Spanish in which they manipulated both number and gender. By using eye-tracking (and eliminating certain limitations of Pearlmutter et al.'s 1999 study), they hoped to provide the best possible record of the reading data. By adding a gender manipulation, two other positive things were made possible: (i) since gender belongs with the stem of the noun in Spanish (it is ‘intrinsic’, see Eberhard et al. Reference Eberhard, Cutting and Bock2005: 23), it should not induce any extra level of difficulty relative to any base form (as is the case, hypothetically, of plurals vs. singulars in English); and (ii) since gender appears on adjectives but not on verbs, the co-indexation site occurs after the verb (a copula), with some distance between the attractor and that site, a fact that is likely to dissolve any local morphological effect (e.g. el nombre del niñ oera extrañ o ‘the name of the child was strange’). What they found was the following: (i) disruption, as in production, in cases of mismatch; (ii) no asymmetry, as effects were obtained for mismatching singulars as well (albeit less strong); (iii) the fastest possible reaction times for number (not only total reading time and regressions into the region but even first pass regressions, and first fixation duration), in contrast to the data from English, where only regression measures yielded differential results (Pearlmutter et al. Reference Pearlmutter, Garnsey and Bock1999); (iv) also surprisingly fast RTs for gender registered at the verb, that is, at a location prior to the actual co-indexation site (not only cumulative reading time but also first pass regressions out, a tendency which was preserved when the verb and the predicate were analysed together as a region); and (v) no semantic, distributivity effects, using the same materials with which Vigliocco et al. (Reference Vigliocco, Butterworth and Garrett1996a, Reference Vigliocco, Hartsuiker, Jarema and Kolkb) obtained such effects in production for the same language.

This pattern of results is difficult to accommodate by any existing model.Footnote ¹⁶ In the first place, it casts doubt on the markedness theory of attraction, a theory taken almost for granted nowadays (Bock & Eberhard Reference Bock and Eberhard1993; Vigliocco et al. Reference Vigliocco, Butterworth and Semenza1995; Vigliocco et al. Reference Vigliocco, Butterworth and Garrett1996a, Reference Vigliocco, Hartsuiker, Jarema and Kolkb; Eberhard Reference Eberhard1997; Antón-Méndez, Nicol & Garrett Reference Antón-Méndez, Nicol and Garrett2002; Eberhard et al. Reference Eberhard, Cutting and Bock2005; Wagers et al. Reference Wagers, Lau and Phillips2009; Haskell, Thornton & MacDonald Reference Haskell, Thornton and MacDonald2010).Footnote ¹⁷ Note, however, that the little known data from the Romance languages are consistent with the view that the role of markedness should be re-evaluated: Vigliocco et al. (Reference Vigliocco, Butterworth and Semenza1995) and Franck et al. (Reference Franck, Lassi, Frauenfelder and Rizzi2006) did not find strong asymmetry in Italian either, and Franck et al. (Reference Franck, Vigliocco and Nicol2002) could not find it in French. It is interesting that Vigliocco et al. (Reference Vigliocco, Butterworth and Garrett1996a) did find a robust effect in Spanish in their production experiment with the same materials. Why should a plural vs. singular contrast matter less in a comprehension experiment in Spanish than in all the experiments on production in English? One likely reason is that in a language fraught with morphological specifications of all kinds of features (person, gender and number being marked on determinatives, adjectives, participles, nouns and verbs), the singular vs. plural contrast is not the same kind of conspicuous event that it is in English.Footnote ¹⁸ A short phrase such as nosotros los ricos altos y guapos ‘us tall and handsome rich guys’ contains a variety of eleven morphological features in Spanish, but only two – of number – in English. Additionally, when it comes to gender, for most nouns there is no base form that is gender-less in Spanish, where a gender specification must be part of the lemma for each word (see Lorimor et al. Reference Lorimor, Bock, Zalkind, Sheyman and Beard2008 and Acuña-Fariña Reference Acuña-Fariña2009 for differences between gender and number in processing). Another likely reason for the absence of a plural markedness effect may be the fact that, unlike Vigliocco et al.'s (Reference Vigliocco, Butterworth and Garrett1996a) experiment with the same materials, this was a comprehension experiment, where form precedes meaning. We will return to this. Finally, a third potential reason is that, to make morphology matter even more, Spanish NPs house highly informative determiners: el (sg, masc), la (sg, fem), los (pl, masc), las (pl, fem), all meaning ‘the’. This paves the way for our next point.

4.3 Probabilistic agreement

A problem with existing linguistic theories of agreement is that they tend to be too rigid. However, the facts of agreement are typically probabilistic. In a number completion study in English and German, Berg (Reference Berg1998) obtained consistent differences between the two languages. Thus, for instance, ‘number transparent’ nouns (series, number, lot, etc.) tended to establish semantic agreement in English (e.g. a series of reports were issued) but formal agreement with the head in German (eine Reihe von Berichten wurde erstellt ‘a series of reports was issued’). However, results were tendencies, never absolute or near-absolute laws. There were plenty of cases where the statistical pattern met with exceptions. Acuña-Fariña (Reference Acuña-Fariña2009) did a small-scale pilot study involving a comparison of Spanish and English, and replicated its findings with Spanish siding with German, and Riveiro-Outeiral & Acuña-Fariña (Reference Riveiro-Outeiral and Acuña-Fariña2012) further corroborated them. Some of the outliers in these three studies had to do with differing levels of bleaching (loss of lexical content as a result of grammaticalisation). Thus, bleaching seems to account for the behaviour of nouns like series, number, or lot, but not for nouns like bouquet or bunch in expressions like a bouquet/bunch of (roses), where the NP1 position retains a strong lexical meaning. These latter cases are usually conceptualised as units (a bouquet of roses is a gift) and therefore trigger agreement with the head noun even in English (Brems Reference Brems2003; Keizer Reference Keizer2007: 151; see also Langacker Reference Langacker1991a: 89).

The fact that agreement systems are much more nuanced than usually believed (even inside one language, that is, disregarding the obvious complexity revealed by typology, see Corbett Reference Corbett2006) is evident in the attraction data involving determiners. The Franck et al. (Reference Franck, Vigliocco, Antón-Méndez, Collina and Frauenfelder2008) data referred from above are particularly informative. Franck et al. wanted to manipulate the morphophonology of both the noun and the article in complex NPs, as this differs across languages in its transparency/cue reliability. They compared Italian, Spanish, and French. Thus, for instance, Italian nouns tend to be highly reliable as the o/a distinction that codes masculine vs. feminine can be observed in 80% of the nouns in that language. By contrast, determiners are less reliable since when the noun starts with a vowel, the article is contracted independently of the noun's gender (e.g. l'armadio ‘the wardrobe-masc’; l'aria ‘the air-fem’). In such cases, (about 25%), the article is ambiguous with respect to gender in its morphophonology. In Spanish, with many nouns like animal ‘animal-masc’, adolescente ‘adolescent-masc/fem’), or felicidad ‘happiness-fem’, the o/a contrast is informative in only some 68% of the nouns. Then there are a few nouns like mano ‘hand’, which end in -o but are feminine and even nouns which end in -a but tend to be masculine (e.g. agua ‘water’). By contrast, determiners are to be trusted almost without exception, e.g. l aman o ‘the hand-fem’, está suci a ‘the hand is dirty-fem’. What Franck et al. found was a picture of almost perfect opportunism: encoders tended to trust the noun preferentially in the three languages, but that preference was neatly modulated by its reliability: they were sensitive to the nouns alone in Italian, both to the nouns and the articles in Spanish, an they were particularly observant of the articles in French (as nominal endings have largely eroded in this language). That attraction is so opportunistic adds another dimension to the fixity of the semantics vs. syntax debate in agreement by making this classic tug-of-war only part of the story.Footnote ¹⁹

4.4 The fast track of agreement in Romance

We now turn to the Acuña-Fariña et al.'s (Reference Acuña-Fariña, Carreiras and Meseguer2011) comprehension findings in Spanish mentioned briefly in Section 4.2. A relevant aspect of that work was the sheer co-indexation speed. Recall that the existing data from comprehension in English (Pearlmutter et al. Reference Pearlmutter, Garnsey and Bock1999) revealed the classic mismatch effect, but only in relatively late measures (that is, in either regressions or post critical regions). Acuña-Fariña (Reference Acuña-Fariña2009) searched for cross-linguistic differences in the processing of agreement in the psycholinguistic literature (reviewing especially eye-tracking and event-related potential (ERP) studies) and concluded that fast, shallow, non-strategic co-indexations were highly prominent in Spanish and Italian, but less so in English. For instance, evidence from the processing of epicene nouns in Italian (Cacciari, Carreiras & Cionini Reference Cacciari, Carreiras and Cionini1997) showed that they preferentially bond with (Garrod & Terras Reference Garrod and Terras2000) pronouns of the same morphological gender, even in contexts strongly biassed pragmatically to favour the other gender (e.g. the noun vittima ‘victim-fem’ preferred a feminine pronoun to a masculine one even in the context of a certain boy having a car crash).Footnote ²⁰ He observed that sometimes blind obedience to alliterative form (form continuity, as in l acas ablanc ay suci a ‘the white dirty house’) even led readers to garden path in Spanish because sometimes form continuity crosses a phrasal boundary and illegally binds non-constituents together). The Acuña-Fariña (2011) attraction study in comprehension confirms the following general pattern: in richly inflected languages, form clues are privileged by parsers, which must deal with them at a rate of about ten per second. This happens – let us not forget – precisely in comprehension, that is, a domain where such form clues reach minds before anything else (like meaning) does. Such speed makes unification theories less appealing, as it is difficult to see how the posterior check that unifies independently assigned featural marks can be conducted within such a short time window. The gender data are particularly revealing, since effects are registered as if in advance, that is, at a region before the co-indexation site (the morphology of the adjective) appears on the scene. The same can be said about the idea that attraction (and agreement in general) is a matter of reactivating a previously encountered referent only at the point where a consistent cue instructs language users to do so (a cue-retrieval (memory) process, as in Solomon & Pearlmutter Reference Solomon and Pearlmutter2004; Badecker & Lewis Reference Badecker and Lewis2007; Wagers et al. Reference Wagers, Lau and Phillips2009). If anything, the speed revealed here and elsewhere (Acuña-Fariña Reference Acuña-Fariña2009) in Spanish comprehension is reminiscent of pro-active strategies in the gap-filling literature (Stowe Reference Stowe1986; Tanenhaus et al. Reference Tanenhaus, Boland, Garnsey and Carlson1989; Boland, Tanenhaus & Garnsey Reference Boland, Tanenhaus and Garnsey1990; see also Betancort, Carreiras & Acuña-Fariña Reference Betancort, Carreiras and Acuña-Fariña2006). Is Marking and Morphing (Eberhard et al. Reference Eberhard, Cutting and Bock2005; Bock et al. Reference Bock, Eberhard, Cutting, Meyer and Schriefers2001, Reference Bock, Butterfield, Cutler, Cooper Cutting, Eberhard and Humphreys2006) the solution then? Ignoring the significance of objects interfering in S–V attraction the problem with this theory is that, since gender tends to be arbitrary in Spanish, it is difficult to see how such rapid gender effects may arise in a system (designed on the basis of the semantically grounded and formally conspicuous English-style number feature) where morphing interacts with a previous marking stage based on the conceptual features of the message. The feminine in mesa ‘table’ cannot be grounded in marking/meaning. In sum, only blind copying operations of the kind that have recently been discarded in the attraction literature (Bock et al. Reference Bock, Eberhard, Cutting, Meyer and Schriefers2001; Eberhard et al. Reference Eberhard, Cutting and Bock2005) offer any scope for accommodating fast (indeed pro-active) gender co-indexation bindings in the Romance languages in comprehension (Hawkins Reference Hawkins1994, Reference Hawkins2004). This takes us back to two factors that interact opportunistically: morphology (the Romance languages) and the direction of encoding (comprehension vs. production).

4.5 Opportunistic interactions

The role of morphology and the direction of encoding are also evident in the third conspicuous aspect of the Acuña-Fariña et al. (Reference Acuña-Fariña, Carreiras and Meseguer2011) findings – the absence of a distributivity effect. At first sight, this is surprising given that using the same materials and the same language Vigliocco et al. (Reference Vigliocco, Butterworth and Garrett1996a) did find such effects. Recall that even English showed those effects once materials were carefully controlled (Eberhard Reference Eberhard1997). This invalidates the thesis that distributivity is to be expected more in the Romance languages than in English because the rich morphology of the former provides more opportunities to consult meaning representations activated by the frequent cues via (distributed) unification (Vigliocco et al. Reference Vigliocco, Butterworth and Semenza1995, Reference Vigliocco, Butterworth and Garrett1996a, Reference Vigliocco, Hartsuiker, Jarema and Kolkb). In fact, given our current understanding, the very opposite idea makes perfect sense, namely, that the richer the morphology the stronger the encapsulation of agreement operations from the interference of conceptual properties of the message.Footnote ²¹ To the best of my knowledge this idea has its origins in the Berg (Reference Berg1998) study mentioned above (Section 4.3). As noted earlier, Berg carried out a completion study in English and German in which he showed that English grammaticalises interference far more than German does in agreement operations (e.g. a series.sg of odd events.pl were.pl the result of that). The reason for this, he suggested, is that the strong morphosyntactic component of German grammar keeps semantics at bay. Lately, Marking and Morphing theoreticians (Lorimor et al. Reference Lorimor, Bock, Zalkind, Sheyman and Beard2008; Foote & Bock Reference Foote and Bockin press; see below) have further explored this idea in an attempt to test how their model works and to refine it, specifically in the reconciliation process, where morphology acts as a counterweight to meaning: the more morphological specifications there are, the smaller the effect of marking (the contribution from notional number). There are two factors to consider. One is the direction of encoding, since whatever is to be kept at bay must be activated before morphosyntactic operations may be affected by it. In production, what is activated first is meaning. Hence the Vigliocco et al. (Reference Vigliocco, Butterworth and Garrett1996a) findings of distributivity effects in Spanish in a production study and the absence of such effects in the Acuña-Fariña et al. study in comprehension (same materials, same language). The second factor is the semantic blocking component (‘morphological filtering’ in Foote & Bock Reference Foote and Bockin press), i.e. the rich morphology:

Acuña-Fariña (Reference Acuña-Fariña2009) showed this logic to be correct in a comparison of English and Spanish with materials similar to those in Berg (Reference Berg1998) . The same conclusion can be applied to the findings of a production study by Lorimor et al. (Reference Lorimor, Bock, Zalkind, Sheyman and Beard2008) on attraction in Russian, another richly-inflected language. This language was less, rather than more, affected by semantic interference than English:Footnote ²²

Haskell et al. (Reference Haskell, Thornton and MacDonald2010) examined English corpora and concluded that English does tend to grammaticalise semantic interference. They point out that a process similar to structural priming contributes to the error asymmetry of English attraction data via speakers' past experiences with related agreement constructions which contain singular phrases that agree in plural with predicates (like Berg's (Reference Berg1998) category 3 construction, e.g. a series of/a bunch of+PLURAL, mentioned in Section 4.3 above).Footnote ²³ Given the role of morphology and the direction of encoding, the absence of notional effects in the Acuña-Fariña et al. (Reference Acuña-Fariña, Carreiras and Meseguer2011) study is in fact doubly right: first, Spanish must be less affected by notional interference due to its strong, well-oiled morphosyntactic component; and secondly, in a task which starts with redundant form continuity (comprehension), the conceptual properties of the sentential message cannot be active before the form comes, but must be necessarily made possible by that pre-existing form. Hence, meaning has fewer opportunities to interfere.

The last piece of evidence for this view of attraction and agreement (involving morphology, direction of encoding and architectural opportunism) comes from a recent production attraction study by Foote & Bock (Reference Foote and Bockin press), in which they contrasted two varieties of South-American Spanish – Mexican and Dominican – with English, which acted as the baseline for comparison. It turns out that whereas Mexican Spanish has preserved the rich morphology of European Spanish, Dominican Spanish is losing it (like French or Brazilian Portuguese). For instance, syllable-final -s is now reduced or weakened (or simply elided), and this can eliminate distinctions between the second and third person singular forms in almost all tenses and moods. Furthermore, syllable-final -n can also be weakened or elided, making it difficult to distinguish between third person singular and plural forms (see Lunn Reference Lunn, Lee, Geeslin and Clements2002). The morphology of number on determiners and adjectives is also being lost (Holm, Lorenzino & De Mello Reference Holm, Lorenzino, Mello and Ortiz López1999; Pérez-Leroux Reference Pérez-Leroux and Ortiz López1999; Toribio Reference Toribio2000). In their first experiment Foote & Bock (Reference Foote and Bockin press) observed larger distributivity effects in English and in the Spanish variety with poorer morphology (Dominican). In the second, when participants produced sentences with null subjects (and therefore did not repeat the morphologised preambles), there were large notional effects in the two varieties of Spanish, suggesting that it is the morphological realisation on subject NPs in particular, and not general language type, that drives agreement computations. This suggests a far more nuanced account of agreement operations than has been contemplated by any existing model or theory. It also suggests a much more opportunistic and therefore effective and realistic account in that probabilistic results are inescapable even inside the same language. This line of work brings hope that, by adequately measuring the role of the morphological component in agreement, very precise cross-linguistic and even cross-construction predictions may be possible.

4.6 Competence and performace

In sum, the varying cross-linguistic levels of semantic affectedness/encapsulation cannot be predicted by any theory (linguistic or psycholinguistic) which does not incorporate the role and size of morphology and the direction of encoding as essential parameters. The filtering of meaning effects by an exuberant morphology (measured, for instance, as the magnitude of distributivity effects) is congruent with the fact that morphological transparency (on whatever constituent this resides the most) boosts the signal and so promotes accuracy. It also automatises the creation of phrasal packages even at the risk of garden paths. Since agreement is not the only form of clause-building, languages which have less of it must compensate for their attrition by making use of more direct conceptual influence. There is nothing counterintuitive in that, given that the whole purpose of language is the conveyance of concepts from one mind to another to start with, so occasions to bypass form are desirable as long as there is no information loss. Lately, a consensus seems to be forming around the idea that ‘semantic agreement appears when grammatical agreement fails’ (Baker Reference Baker2008: 24; see also Wechsler & Zlatić Reference Wechsler and Zlatić2003; Wechsler Reference Wechsler2011; Wechsler & Hahm Reference Wechsler and Hahm2011). Since having a rich or a poor morphology is a matter of degree, the perennial tug-of-war between semantics and encapsulation in agreement – with its typical cross-linguistic specificities – is naturally accounted for. Both processing and the grammaticisation of processing routines must necessarily be conveniently opportunistic. This logic applies both cross-linguistically and intra-linguistically (as it also depends on the direction of encoding). A clear general prediction is that notional effects should be stronger in English-style languages than in Spanish-style ones. Another is that they should be less strong in comprehension than in production even in the same language. I have used the word ‘opportunism’ throughout, but this is simply to emphasise the latitude each language has in shaping its biases and its grammar within a system/menu that is assumed to be the same for all languages (Foote & Bock Reference Foote and Bockin press).

These conclusions are intended to be of relevance to any linguist interested in the grammar of agreement in particular and in the grand design of the language faculty in general. They have to do with basic dynamics, such as the direction of encoding and the relationship between meaning and form, and they show ways of incorporating the new research on attraction in comprehension into a debate dominated by the facts of production. However, in the introduction we made it clear that lately various theoretical linguists working under the aegis of the Minimalist Program have boldly targeted experimental work on agreement (including attraction) as a propitious ground for testing their specific linguistic theories. It is well-known that competence and performance work in their own mysterious ways (Phillips, Wagers & Lau Reference Phillips, Wagers, Lau and Runner2011); however, some connections of the findings reported here with specific linguistic theories are too tempting to ignore. I cautiously proceed to that now finally, noting that this work is not primarily meant to be a contribution to minimalist grammar.

The three main conclusions reached in this paper are as follows:

1. i. Notional effects should be stronger in English-style languages than in Spanish-style ones (as morphology cannot ‘contain’ semantic interfacing).

2. ii. They should be less strong in comprehension than in production even in the same language (as form, not meaning, is the initial processing cycle, and alliterative form is given even more processing advantage in the rich morphology languages).

3. iii. Given the varying degrees of ‘morphologisation’ in the world's languages, the perennial tug-of-war between form and meaning (and the many idiosyncratic systems) in agreement is thus naturally accounted for.

Given these conclusions and generative ideas about encapsulation of syntactic operations (cycles; Chomsky Reference Chomsky, Roberts, Michaels and Uriagereka1999, Reference Chomsky2001; Miyagawa Reference Miyagawa2009: Chapter 2), the parametrisation reported here sits well with models of agreement which are not rigid. For instance, Bobaljik (Reference Bobaljik, Ackema and Neeleman2002b: 16–17) argues that Pollock's (Reference Pollock1989) proposal to split the traditionally doubly-headed inflection phrase IP into a series of functional projections, giving agreement and tense separate status as heads, and requiring that each projects to a full maximal category, as in (16b) below, is appropriate for rich-inflection languages like Icelandic. For languages with a poor morphosyntactic component, like English or Danish, he proposes the tree in (16a) instead.

1. (16)

   1. (a) English

      

   2. (b) Icelandic

      

The basic idea behind this flexibility is that the inventory of functional projections in the IP domain is subject to cross-linguistic variation. This idea, originated in Johnson (Reference Johnson1990), and pursued also in Thráinsson (Reference Thráinsson, Abraham, Epstein, Thráinsson and Jan-Wouter Zwart1996) and Bobaljik & Thráinsson (Reference Bobaljik and Thráinsson1998), among others, constitutes what is known as the Split Infl Parameter. Although it is meant to explain inflection–movement correlations in Germanic, it may be recruited here to argue that a rich tree with abundant functional projections such as (16b) would stave off semantic interference (which we have attested) more than a poorer tree (with less arborecence) because there is specifically a head labeled Agr in it. The information contained in this node is predictable/routinisable, as are all the verb endings in Romance. In languages like Galician/Portuguese, inflected infinitives dissociate agreement and tense rather clearly too. Conversely, in English, for instance, the past tense is not compatible with agreement in the third person (*walked-s) nor with auxiliary verbs (*he may goes), so in this language Infl can only house one element, either the past tense or the third person suffix, but nothing more elaborate. The less dense tree in (16a) captures that. Another reason why semantic interfacing might be more costly in a configuration like (16b) is that percolation pathways would be more firmly canalised in such finely-articulated trees. They might even require more cognitive resources to deploy, making interference less advantageous. A tree menu like (16a–b) assumes an Agree architecture (Chomsky Reference Chomsky, Roberts, Michaels and Uriagereka1999, Reference Chomsky2001), not a Spec–Head one (Koopman & Sportiche Reference Koopman and Sportiche1991; Koopman Reference Koopman and Griffin2003, Reference Koopman and Boeckx2006). In the context of the current debate among minimalist theoreticians as to whether one framework is superior to the other, the present findings would appear to be more consistent with the former model.Footnote ²⁴

Another reason why an Agree model may be more realistic than a Spec–Head one has to do with the geometrical dynamics it affords. Under Agree, the probe c-commands the entire NP while this is inside the vP. Normally, it would target the first noun in it (the head) based on locality (Chomsky Reference Chomsky2001), but if the process goes wrong for some reason, the second noun would still be related to the probe via c-command. However, in a Spec–Head system, the second noun would not be in any c-command relation with T.Footnote ²⁵ Now, this would be a perfect system because it would prohibit attraction mistakes, but the actual fact is that attraction mistakes do occur and are in fact very frequent. As noted in the introduction, and as Eberhard et al. (Reference Eberhard, Cutting and Bock2005) point out, attraction occurs in a stunning 13% of the materials in experiments. Intuition suggests that normal everyday use follows suit (see Pfau Reference Pfau and Griffin2003). So the Agree architecture could actually explain these frequent mishaps better than an error-free Spec–Head set-up. Naturally, these considerations only make sense if we assume a close connection between our parsing system and grammar.Footnote ²⁶

Finally, there is another dimension where the results reported here might relate to research in Minimalism, and it has to do with another current debate central to this framework, namely, the role played by morphology in grammar. Basically, there are two positions on this. Proponents of Distributed Morphology (Halle & Marantz Reference Halle, Marantz, Keyser and Hale1993; Embick & Noyer Reference Embick, Noyer, Ramchand and Reiss2007) assume that the presence of Agr nodes is subject to post-syntactic language specific well-formedness conditions, rather than being licensed in the narrow syntax. That is, on this view, agreement is a ‘PF-ish’ phenomenon (Bobaljik Reference Bobaljik2002a; Putnam Reference Putnam2007) implemented via the ‘late-insertion’ of ‘dissociated’ morphemes. The idea is that syntax concatenates abstract morphemes and these are subsequently provided with phonological exponents (vocabulary items) via post-syntactic vocabulary insertion or morpheme realisation rules. In Bobaljik (Reference Bobaljik, Ackema and Neeleman2002b: 25) words, ‘[s]yntax puts abstract morphemes together, morphology spells them out’. The opposite view is a lexicalist account that favours early insertion of morphemes (Jelinek Reference Jelinek1984; Vikner Reference Vikner and Haegeman1997; Ackema Reference Ackema2001; see also Chomsky Reference Chomsky1995 and Levelt Reference Levelt1989). In an early insertion model of minimalist derivations, words would be morphologised relatively long before the attraction effects can be perceived, therefore making the whole phenomenon of attraction much less probable. By contrast, in a late-insertion model, the plural suffix would enter the scene right before the verb is given morphological form, thereby creating more opportunities for mistakes. Again, given actual rates of attraction errors, and the morphing (formal) processes uncovered by the Marking and Morphing model, late insertion would appear to be a more realistic approach (see Pfau Reference Pfau and Griffin2003 for a similar conlusion based on a corpus of German errors). This is being highlighted by Marking and Morphing researchers in their most recent work (e.g. Lorimor et al. Reference Lorimor, Bock, Zalkind, Sheyman and Beard2008: 773). Note that, as is well-known, late insertion is independently very compatible with the well-known fact that in stranding mistakes (Garrett Reference Garrett and Bower1975), the stems of words exchange while their affixes remain in place (She's already trunked two packs, when She's already packed two trunks was intended). This suggests ‘a point in production when the structural components of words (stems and affixes) are linked to positions in a phrasal frame’, and also ‘that stranded morphemes occupy their syntactically expected positions, unlike their host stems … In I roasted a cook, the past tense -ed is where it should be’ (Eberhard et al. Reference Eberhard, Cutting and Bock2005: 9). It is interesting to note that, as Humphreys (Reference Humphreys1998) has shown, stranded morphemes are much more likely to be inflectional than derivational. Equally interesting is the fact that, contrary to the behaviour of inflections in nouns and verbs, personal pronouns do not seem to leave their grammatical properties behind when they are involved in exchanges, but take their agreement properties with them (e.g. Most cities are true of that when That is true of most cities is intended; see Eberhard et al. Reference Eberhard, Cutting and Bock2005: 10–11, and referentes therein). A dissociated architecture of stem+suffix is, therefore, appealing for nouns and verbs, making an undissociated one the logical choice for pronouns.