Recency and Lexical Preferences Recency and Lexical Preferences in Spanish

One experiment provided evidence in support of Gibson, Pearlmutter, Canseco-Gonzalez, and Hickok's (1996) claim that a recency preference applies to Spanish relative clause attachments, contrary to the claim made by Cuetos and Mitchell (1988). Spanish speakers read stimuli involving either two or three potential attachment sites in which the same lexical content of the two-site conditions appeared in a di erent structural con guration in the three-site conditions. High attachment was easier than low attachment when only two sites were present, but low attachment was preferred over high attachment, which was in turn preferred over middle attachment, when three sites were present. The experiment replicated earlier results and showed (1) that attachment preferences are determined in part by a preference to attach recently/low, and (2) that lexical biases are insu cient to explain attachment preferences. Recency and Lexical Preferences 3 Recent debate about the properties of the human sentence processor has focused both on the purported universality, or cross-linguistic applicability, of various parsing principles (e.g., Brysbaert & Mitchell, 1996; Cuetos & Mitchell, 1988; Frazier, 1987a, 1987b; Mitchell & Cuetos, 1991; Schriefers, Friederici, & Kuhn, 1995), and on the degree to which such principles might be replaced by lexicallyand contextually-determined preferences (e.g., Boland, Tanenhaus, Garnsey, & Carlson, 1995; MacDonald, Pearlmutter, & Seidenberg, 1994; Trueswell & Tanenhaus, 1994). While some of the proposals which are intended to account for evidence about lexical and contextual e ects also have implications for universality issues (e.g., Tabor, Juliano, & Tanenhaus, 1997; cf. Frazier, 1995), the majority of the work examining one issue has not directly examined the other (cf. Schriefers et al., 1995). One proposed universal parsing principle which has received recent attention with respect to the universality issue is Frazier's (1979, 1987a) Late Closure principle (see also Kimball's (1973) Right Association and Gibson's (1991) Recency Preference). Late Closure speci es that the parser should attach new material to the clause or phrase currently being processed, when grammatically permissible. This principle accounts for the preference in (1a) to attach yesterday to the recently-processed verb phrase (VP) died rather than the earlier VP said, and for the preference in (1b) to interpret the noun phrase (NP) a mile as the direct object of the current VP (headed by jogs), instead of as the subject of the upcoming matrix clause. In (1b), the result of the Late Closure preference is measurable di culty (Frazier & Rayner, 1982) when a mile must be reanalyzed as the matrix clause subject. (1) a. John said Bill died yesterday. b. Since Jay always jogs a mile seems light work. While Late Closure intuitively applies in English, Cuetos and Mitchell (1988; Mitchell & Cuetos, 1991; Mitchell, Cuetos, & Zagar, 1990) argued that a preference for attachments to recent over non-recent material (as in (1a)) was not universal. For Spanish in particular, they showed that comprehenders preferred high (less recent) attachment in constructions like (2a), where the relative clause (RC) que tuvo el accidente (who had had the accident) can either attach high to hija (daughter) or low to coronel (colonel). English speakers preferred low attachment for corresponding ambiguous English stimuli, as in (2b). (2) a. El periodista entrevisto a la hija del coronel que tuvo el accidente. b. The journalist interviewed the daughter of the colonel who had had the accident. Mitchell and Cuetos (1991; Cuetos & Mitchell, 1988) provided on-line evidence for a Spanish high attachment preference by using disambiguated versions of the construction in (2a), as shown in (3). Spanish comprehenders read the prepositional phrase con su marido (with her husband) more slowly in (3a) than in (3b), because the RC que estaba en el balc on (who was on the balcony) is forced to attach to the low site actriz (actress) in (3a), whereas it attaches to the high site criada (servantFEM) in (3b). Recency and Lexical Preferences 4 (3) a. Alguien dispar o contra el criado de la actriz que estaba en el balc on con su marido. \Someone shot the servantMASC of the actress who was on the balcony with her husband." b. Alguien dispar o contra el criada del actor que estaba en el balc on con su marido. \Someone shot the servantFEM of the actor who was on the balcony with her husband." More recent examinations of the ambiguity in (2) and (3) have replicated these ndings (e.g., Carreiras & Clifton, 1993, 1998; Gilboy, Sopena, Clifton, & Frazier, 1995), and the preference for non-recent attachment in this construction has also been found in German (Hemforth, Konieczny, & Scheepers, 1997) and Dutch (Brysbaert & Mitchell, 1996), whereas Italian speakers appear to show the same low attachment preference as English speakers (de Vincenzi & Job, 1995). Because of the variation across languages, Mitchell and Cuetos (1991; see also Cuetos, Mitchell, & Corley, 1996) proposed the tuning hypothesis: that parsing preferences are tuned on a construction-by-construction basis within a language, with comprehenders attending to the relative frequencies of the alternatives with which an ambiguity is resolved. On this proposal, comprehenders' preferences for recent or non-recent attachments in a language result from the history of the comprehenders' exposure to the language and speci cally to similar constructions in the language, rather than from any structurally-de ned simplicity or ease-of-processing metric like Late Closure. Contrary to Mitchell and Cuetos' (1991) claim, however, Gibson, Pearlmutter, CansecoGonzalez, and Hickok (1996) presented evidence that Late Closure (or a similar recency property) does apply in Spanish RC attachments. They measured reading times using an on-line grammaticality judgment task with items like (4), in both Spanish and English. (4) a. la l ampara cerca de la pintura de la casa que fue da~ nada en la inundaci on b. the lamp near the painting of the house that was damaged in the ood Whereas there are only two possible attachment sites for the RC in (2), (4) contains three possible sites (l ampara, pintura, casa; lamp, painting, house) for the RC que fue da~ nada en la inundaci on (that was damaged in the ood). This di erence resulted in a change in the cross-linguistic pattern: Gibson, Pearlmutter et al. (1996) found that both Spanish and English readers preferred attachment of the RC to the lowest site (casa; house), followed by attachment to the highest site (l ampara; lamp), with middle attachments (to pintura; painting) most di cult. Thus when only two attachment sites for the RC were present, in Cuetos and Mitchell (1988), Spanish and English di ered; but when a third site was also present, comprehenders in the two languages showed the same pattern, preferring to attach to the most local site. In order to account for the di ering cross-linguistic preferences in twoand three-site ambiguities, Gibson, Pearlmutter et al. (1996) proposed two principles. One of these, Recency (closely related to Late Closure), applied uniformly across languages and speci ed an increasing cost associated with increasingly distant attachments. The second principle, Predicate Proximity, speci ed a xed cost associated with attachments which were not as close as possible to a predicate phrase (typically a VP). Furthermore, they hypothesized that the relative strength of Predicate Proximity varied across languages, and that it was Recency and Lexical Preferences 5 weak in English but strong in Spanish. (See the General Discussion for more detail on this hypothesis.) As a result, in English, Recency dominates even for 2NP ambiguities, and becomes stronger as the number of sites increases, because more distant attachments become increasingly more costly by Recency, while Predicate Proximity is constant over distance. In Spanish, however, Predicate Proximity is strong enough to dominate in 2NP ambiguities, even though Recency dominates for more extended ambiguities (three or more sites). Although Gibson, Pearlmutter et al.'s (1996) proposal provides one possible account of the RC attachment data, there are other ways to interpret the results. In particular, recent constraint-based lexicalist theories (e.g., MacDonald et al., 1994; Trueswell & Tanenhaus, 1994) have argued that attachment preferences are determined by a combination of lexically-based frequency biases and broader contextual biases such as plausibility, rather than structurally-based principles. Because the stimuli in the 2NP and 3NP studies di ered in lexical content and were not independently controlled for plausibility, these factors might be able to explain the preferences in both the 2NP and 3NP structures. Related accounts have been proposed for similar ambiguities, involving attachment of a PP (instead of an RC) either to multiple possible NP sites (Gibson & Pearlmutter, 1994; MacDonald & Thornton, 1996) or to a VP or an NP site (Sch utze & Gibson, 1996; Spivey-Knowlton & Sedivy, 1995; Taraban & McClelland, 1988; cf. Clifton, Speer, & Abney, 1991; Rayner, Carlson, & Frazier, 1983). On some lexically-based approaches, PP and RC attachments can be treated similarly, and thus just as some nouns are more likely than others to take a modi er with a particular preposition (e.g., the book about. . . vs. the brick about. . . ), nouns might di er in their frequency of modi cation by an RC. If this determines attachment preferences (i.e., RCs prefer to attach to whichever noun site is most often modi ed), then RC attachment preferences should follow particular nouns, regardless of the nouns' relative structural positions. The present experiment examines this possibility in Spanish using 2NP and 3NP constructions as in earlier work. However, the speci c NPs used for the two sites available in the 2NP construction were also used as the middle and low attachment sites in the 3NP construction, and the same RC was used for all conditions. Thus if preferences follow the particular nouns, the high and low sites in the 2NP construction should show the same pattern as the middle and low sites in the 3NP construction. If instead a non-lexically-based principle determines preferences, then the results should replicate the Spanish pattern in the literature: For 2NP constructions, the high site should be preferred, whereas for 3NP constructions, the low site should be preferred, followed by the high site and then the middle site. Constraint-based lexicalist approaches (and others) also raise the possibility that plausibility could determine attachment preferences. In cases of ambiguous RC attachment, attachment of the RC to one site might create a more plausible interpretation than attachment of the RC to another site. For example, in the lamp near the painting that was ickering in the window, the lamp is a more plausible subject of was ickering than is the painting. Although the RCs in previous studies examining both 2NP and 3NP constructions were designed to be plausible modi ers prior to the disambiguating region, subtle plausibility di erences among the di erent attachments might nevertheless have determined the pattern Recency and Lexical Preferences 6 of results. If such plausibility di erences were responsible for the earlier patterns, then the same set of predictions holds as for lexically-based biases: Attachment preferences should follow the particular nouns, because the attaching RC is the same across the previous 2NP and 3NP constructions. In addition to testing lexical-bias and plausibility predictions, the present experiment will allow us to determine whether three other potentially important di erences between the earlier 2NP and 3NP experiments can explain the di erent patterns of results. The rst of these is that the type of preposition used to link the attachment sites di ered across the 2NP and 3NP experiments. Gilboy et al. (1995) showed in a forced-choice questionnaire study that the nature of the linking preposition for 2NP constructions in uences attachment preferences, such that high attachment is easier when a thematically-vacuous preposition like de (of ) intervenes between the two sites, compared to when a preposition like con (with) intervenes, which assigns a thematic role independent of the head noun it follows (see also de Vincenzi & Job, 1995, for similar results in Italian). Cuetos and Mitchell (1988) and most of the studies examining 2NP constructions used a form of de (as in (2a): la hija del coronel) to link the NPs, whereas Gibson, Pearlmutter et al. (1996) used a variety of prepositions, the majority of which were locatives (e.g., cerca de (near), as in (4a)) or speci ed other adjunct relationships assigning thematic roles independent of their preceding head nouns, including the preposition con (with). Hence Gibson, Pearlmutter et al. may have found a low attachment preference because they used a larger proportion of thematic-roleassigning prepositions than the 2NP experiments (see Cuetos et al., 1996, for discussion). The present experiment controls for this possibility by always using the preposition de to link the attachment sites. A second potentially confounding di erence between the 2NP and 3NP Spanish preference patterns is that the Gibson, Pearlmutter et al. (1996) 3NP study used a word-by-word timed ungrammaticality judgment task, whereas studies of the 2NP ambiguity used either o -line judgments of ambiguous stimuli or region-by-region self-paced reading (e.g., the RC might be presented all at once). Thus one possibility is that the on-line ungrammaticality judgment task arti cially in ates low attachment preferences (e.g., because processing resources normally devoted to maintaining the availability of higher sites in memory are instead required to perform the additional judgment task). Cuetos et al. (1996; cf. Gilboy & Sopena, 1996) discuss related concerns. The study below uses word-by-word self-paced reading, with no grammaticality judgment component, for both 2NP and 3NP stimuli. A third di erence between Gibson, Pearlmutter et al. (1996) and all of the 2NP ambiguity studies is that Gibson, Pearlmutter et al. used complex NP fragments (as in (4)) rather than complete sentences (as in (2a) and (3)). Gibson, Pearlmutter et al. e ectively assumed that Note that the current experiment cannot rule out a potentially even more subtle e ect of plausibility, arising from the possibility that inserting a third (high) site alters the plausibility of attaching to the other two sites. For example, in the counter in the kitchen that was spotless, modi cation of counter by that was spotless is quite plausible; but it is much less so in the grime on the counter in the kitchen that was spotless. Although we cannot de nitively rule out this type of plausibility explanation, it seems unlikely given our stimuli (see Appendix). Recency and Lexical Preferences 7 readers would treat the fragments as subjects of sentences, although a matrix verb was never presented. In the 2NP ambiguity studies, however, the rst of the two sites was a direct or indirect object and thus the entire complex NP followed the sentence's matrix verb. If maintaining an NP structure (e.g., a subject NP) in working memory prior to encountering a licensing verb involves some cost (e.g., Gibson, 1991), then the fact that Gibson, Pearlmutter et al.'s NPs occurred without verbs might have created an additional bias to attach to very recently processed material. The complex NPs in the current experiment were all either direct or indirect objects, so a replication of the Gibson, Pearlmutter et al. result would indicate that the subjectversus object-position placement of the complex NP does not interact with attachment preferences within the NP.