Island Constraints and Online Dependency ...

Island Constraints and Online Dependency Formation Processes Masaya Yoshida, Nina Kazanina, Leticia Pablos and Patrick Sturt Northwestern University, University of Bristol, Leiden University, and The University of Edinburgh

1.

Introduction

Since the discovery of island constraints (Ross, 1967), there has been a long standing controversy regarding the nature of islands, namely whether the island effects should be understood as the formal constraints on syntactic structures and syntactic operations (Chomsky, 1981, 1986, Huang, 1982, Lasnik and Saito, 1992, Phillips, 2006, 2010, Rizzi, 1982, 1990, Sprouse, Wagers and Phillips, 2010 among others) or as the epiphenomenon resulting from general cognitive constraints on language processing (Dean, 1991, Grosu, 1972, 1981, Hawkins, 1999, Hofmeister and Sag, 2010, Kluender and Kutas, 1993, Kluender, 1998, 2005, Pritchett, 1992, Pritchett, 1991 among others). The aim of this paper is to examine the prediction that processing complexity-based accounts make for islands (henceforth the complexity account) (Hawkins, 1999, Hofmeister and Sag, 2010, Kluender and Kutas, 1993, Kluender, 1998, 2005) by crucially referring to the processing of cataphoric constructions. We predict that the processing of cataphoric dependencies be constrained by islands if the basic logic of the complexity account is adopted. Through a word-by-word self-paced moving window experiment that used the gender mismatch paradigm, we show that the online cataphoric dependency formation is not constrained by islands, and therefore the processing of cataphora raises a challenge for the complexity account. 2.

The Complexity Account and the Processing of Cataphora

The basic logic behind the complexity account can be summarized as follows (Phillips 2010, Sprouse, Wagers and Phillips 2010): island effects arise because the postulation of a gap inside an island incurs a high processing burden on the parser. Under the complexity account, there are two main sources for the processing complexity: (i) a longdistance dependency and (ii) a complex element that intervenes between the head and the tail of the dependency (Hawkins, 1999, Hofmeister and Sag, 2010, Kluender and Kutas, 1993, Kluender, 1998, 2005). Simply put, during the processing of a long-distance dependency, the dependent element that must be held in the working memory (Fiebach, Schlesewsky and Friederici, 2001, Fiebach, Schlesewsky and Friedrici, 2002, Gibson,

Yoshida, Kazanina, Pablos and Sturt

1998, Gibson and Pearlmutter, 1998, Pickering and Barry, 1991), and the complexity of intervening elements (Gibson, 1998, Hofmeister and Sag, 2010, Warren and Gibson, 2002) both consume memory resources and they cause a processing overload. Under the complexity account, the unacceptability of the following example is accounted for by the complexity attributed to holding a wh-filler in memory, and the processing difficulty of the intervening elements. In (1), a wh-filler-gap dependency is formed across a relative clause in subject position (presumably one of the least acceptable types of islands). (1)

*We wondered what [RC the clinic that reassured Gregory about _ ] informed the police?

The greatest appeal of the complexity account is that processing factors that explain island violations also explain a range of other seemingly unrelated phenomena, such as the acceptability gradation of multiply center embedded sentences (Gibson, 1998 among others) and the difficulty of object relative clauses as compared to subject relatives (Carreiras et al., 2010, Hsiao and Gibson, 2003 for recent review) among others. Because the complexity account relies on general processing complexity considerations, it is not stipulated to apply only to some specific constructions, such as wh-filler-gap dependencies. Thus, the complexity account predicts that the parser should not be able to form any long-distance dependency into an island due to the combined processing burden. However, somewhat problematically for the complexity account, it has long been recognized that islands do not hinder dependencies between pronouns and their antecedents (Chomsky, 1977, Ross, 1967). For example, in Ross’s original formulation of islands, islands are stipulated as constraints on the representation that involve a “gap” which, in turn, is the result of the “chopping rule” (Ross, 1967, see also Hornstein, Lasnik and Uriagereka, 2007 for a recent discussion). Thus, structures that do not involve a gap are not subject to island constraints. In (2), for example, an acceptability judgment shows that his may be interpreted as Gregory, confirming Ross’s original formulation. (2)

His1 agents were certain that [RC the clinic that reassured Gregory1 about the privacy protection] had informed the police.

Crucially, an example like (2) shares a number of features with an example like (1). First of all, there is a dependency between the pronoun his and its antecedent Gregory. Furthermore, the elements that constitute a relative clause intervene in this dependency. Therefore, if the complexity account outlined above is correct, we expect an example like (2) to be equally unacceptable as (1): the combination of the complexity attributed to the pronoun-antecedent dependency formation and to the interveners should result in a processing overload. A note is in order. It is possible that the application of the complexity account to cataphoric dependencies is not as simple as illustrated above. An alternative possibility could be to stipulate that the processing complexity of wh-filler-gap dependencies and the cataphoric dependencies are different. Obviously, the elements that form cataphoric dependencies (the pronoun and the antecedent DP) and those that form wh-filler-gap dependencies (the wh-filler and the gap) are different. However, such stipulation is obviously not desired. First, it is not clear whether there should be a difference in the

Processing Complexity and Dependency Formation

processing complexity of these two types of dependencies. It has been observed in the literature that cataphoric dependencies incur processing complexity much like filler-gap dependencies (Filik and Sanford, 2008). Furthermore, it is not clear whether such a distinction of two types of dependencies stems from general cognitive constraints. Crucially, if we do not find a good motivation for such a distinction in terms of general cognitive constraints or in terms of sentence processing mechanisms, a position that adopts such a distinction is (almost) equivalent to a position that assumes two different grammatical (either syntactic or semantic) dependency types which are subject to different grammatical constraints. 1 Once such a distinction is introduced without an independent motivation, the complexity account looses its appeal, such that grammatically stipulated constraints are reduced to general cognitive constraints. The acceptability of an example like (2) could potentially be counter evidence against the complexity account. However, we need to be cautious in assessing the predictions made by this account since it could alternatively predict that cataphoric dependencies across islands are acceptable if the following reasoning was entertained. Since the original predictions of the complexity account are formulated based on how dependencies are formed in the real-time processing of sentences and not on the final grammatical status of the sentence, the contrast in judgment acceptability between (1) and (2) does not necessarily constitute direct counter evidence for complexity accounts. In the case of the cataphoric dependency in (2), it is possible that the example is judged acceptable because the dependency is formed in the later stage of processing as a result of ‘coercion’: if the parser cannot find other antecedent for the pronoun in the most immediate available input, the parser could forcefully link the pronoun to the only available noun phrase. It has been observed that the formation of referential dependencies can indeed be delayed (see Garrod, Freudenthal and Boyle, 1994, Garrod and Terras, 2000, Sanford, Moar and Garrod, 1988). Therefore, islands - assumed to affect primarily the online dependency formation under the complexity account - may not block the late formation of the cataphoric dependency. As a result, an example like (2) is predicted to be acceptable. Therefore, to test the prediction of the complexity account through cataphoric dependencies, it is crucial to examine whether cataphoric dependencies are formed immediately during the online processing of the sentence, and whether the online cataphoric dependency formation is affected by island constraints or not. 3. 3.1

Online Formation of Cataphoric Dependencies and the Complexity Account Active Dependency Formation

It is well known that the processing of wh-filler-gap dependencies involves the so-called active dependency formation mechanism (Clifton and Frazier 1989). Upon encountering the wh-filler, the parser actively searches for its licensor (i.e. a verb, a preposition or a gap). Thus, once the dependent element is encountered, the parser attempts to complete 1

It seems obvious that they are different types of dependencies since studies on theoretical linguistics have observed differences between these two types of dependencies. However, what is crucial in this context is whether such a distinction is motivated by the consideration of general cognitive constraints or sentence processing mechanisms.

Yoshida, Kazanina, Pablos and Sturt

the dependency as soon as possible. As a result, the parser posits a gap in every relevant position even before all bottom-up information confirming the presence of the gap becomes available (Crain and Fodor, 1985, Frazier, Clifton and Randall, 1983, Frazier and Flores d'Arcais, 1989, Phillips, 2006, Stowe, 1986, Tanenhaus et al., 1989, Traxler and Pickering, 1996). The parser’s preference to create the shortest dependency has often been linked to the parser’s tendency to minimize the cost of holding the dependent element in working memory (Chen, Gibson and Wolf, 2005, Gibson, 1998, Pickering and Barry, 1991). Stowe demonstrated that the parser tries to posit a gap in the object position of the verb bring in (3), even though the gap can be posited after the preposition to at a latter position in the sentence. (3)

My brother wanted to know who Ruth will bring us home to __ at Christmas.

Stowe observed a reading time slowdown at the pronoun us which occupies the potential gap position when compared to the same pronoun in a sentence without a wh-dependency. This slowdown is referred to as the Filled-Gap Effect (FGE) in the literature. Importantly, Stowe’s observation suggests that the parser tries to posit the gap and complete the dependency generated at the wh-word who as soon as possible. Stowe further demonstrated that the parser does not try to posit a gap in an island, even if the positing of the gap inside an island lead to the early termination of the dependency formation. Thus the dependency formation effects, such as FGE (Stowe 1986) or plausibility effects (Phillips, 2006, Traxler and Pickering, 1996) are not seen in islands (Bourdages, 1985, McElree and Griffith, 1998, Stowe, 1986, Traxler and Pickering, 1996). The observation that dependency formation processes are affected by islands is exactly what the complexity account predicts. Because island domains induce processing complexity, the parser does not attempt to posit a gap inside an island. For the complexity account, it is crucial that the filler-gap dependencies are formed actively and immediately, and that the parser will try to terminate the dependency as soon as possible.2 Previous studies showed that the processing of a cataphoric dependency also involves an active dependency formation mechanism (Kazanina et al., 2007, van Gompel and Liversedge, 2003). Making use of a Gender Mismatch paradigm, Kazanina et al., examined the processing of cataphora constructions as in (4). (4)

2

a. b. c. d.

He announced the war strategy… She announced the war strategy… His advisor announced the war strategy… Her advisor announced the war strategy… …while the brave, noble king gathered up the troops, but James/Diane thought…

Note that, the same effect is predicted by the formal account of islands. Under the formal account, the parser does not try to posit a gap in an island because it has immediate access to the grammatical constraints, which prevent it from constructing a grammatically illicit representation.

Processing Complexity and Dependency Formation

In (4), male and female pronouns are located in the nominative position (he/she) and genitive position (his/her). Kazanina et al., observed that the embedded subject DP king was read significantly slower in (4d) than in any of the other three conditions. In other words, the embedded subject DP yields a slower reading time when its gender information mismatches with the genitive pronoun (the so-called Gender Mismatch Effect: GMME (van Gompel and Liversedge, 2003)). This result suggests two important aspects of the processing of cataphoric pronouns. First, the parser tries to link the pronoun to the closest potential antecedent. As can be seen in (4), in this design, there is always a grammatical antecedent (James/Diane) available in a later position in the sentence. However, the GMME is observed at the first embedded subject DP, king. Thus, much like the processing of Filler-gap dependencies, the parser also tries to minimize the length of the dependency in cataphoric contexts. This suggests that the active dependency formation mechanism is also present in the processing of cataphora constructions and that this effect is motivated by working memory considerations, i.e., holding a pronoun in memory incurs a cost. This view is further supported by the observation that a cataphoric pronoun also induces a processing cost (Filik and Sanford, 2008). Second, the fact that a gender-mismatching nominative pronoun does not yield GMME indicates that the active search process is constrained by a structural condition such as Binding Condition C (Chomsky, 1981), since the only difference between (4a/b) and (4c/d) is the structural position of the pronouns in the sentence. In (4a/b), the pronouns are located in the nominative position, and thus they c-command the embedded subject king. On the other hand, in (4c/d), the pronouns are in the genitive position, and they do not c-command the embedded subject. GMME is thus observed when the closest potential antecedent is not c-commanded by the pronoun, i.e. when the configuration is not subject to Binding Condition C. 3.2

The Predictions of the Complexity Account on Cataphora Processing

As we have seen, under the complexity account, the unacceptability of the island violation is understood as the processing overload resulting from the combination of the cost associated with the dependency formation process (holding the dependent element and associating it to the licensor) and the cost of processing the intervening elements (Hawkins, 1999, Hofmeister and Sag, 2010, Kluender and Kutas, 1993, Kluender, 1998, 2005). The complexity account, thus predicts that any dependency that crosses an island (or a sequence of complex elements) should yield a processing overload and the island effect. In the previous discussion, we have made clear that the processing of cataphora constructions involves the active dependency formation mechanism. In other words, the processing of cataphora constructions involves the following three components: holding the dependent element (the pronoun) in memory; searching for the licensor (the antecedent DP); and associating the dependent element to the licensor. Thus, if this is the case, the complexity account predicts cataphoric dependencies to be affected by islands in a configuration like (5). (5)

His1 agents were certain that [RC the clinic that reassured Gregory1 about the privacy protection] had informed the police.

Yoshida, Kazanina, Pablos and Sturt

In this configuration, the dependency between the pronoun his and the antecedent Gregory is intervened by complex elements such as a definite DP and the relative pronoun. Thus the combination of holding the dependent element, the pronoun, and the cost of processing these interveners should result in a processing overload. Furthermore, the processing overload should affect the online processing of cataphoric dependency formation. In order to test this prediction, we conducted a word-by-word self-paced moving window experiment, making use of the Gender Mismatch Paradigm. 4.

Experiment

Participants Seventy-two undergraduate students at Northwestern University participated in the experiment for course credit. All participants were native speakers of English and gave informed consent for participation. Design and Procedure Twenty-eight sets of sentences such as (6) were created using a 2×2 design with factors PRONOUN CASE (Genitive vs. Nominative) and GENDER SPECIFICATION of the pronoun and the embedded subject (gender-match vs. gendermismatch). (6)

a. b. c. d.

His agents were certain that [RC the clinic that reassured Gregory Morton about the privacy protection] had informed the police, but Margaret didn't believe it. Her agents were certain that [RC the clinic that reassured Gregory Morton about the privacy protection] had informed the police, but Margaret didn't believe it. He was certain that [RC the clinic that reassured Gregory Morton about the privacy protection] had informed the police, but Brandon didn't complain about it. She was certain that [RC the clinic that reassured Gregory Morton about the privacy protection] had informed the police, but Margaret didn't complain about it.

In the Genitive conditions (6a/b), the first potential antecedent for the pronoun, Gregory Morton¸ is located inside a relative clause. If the parser considers such a relation at an early stage of processing notwithstanding that the antecedent is located inside a relative clause, longer reading times should be found at the NP Gregory Morton in the genderincongruous (6b) as compared to the gender-congruous (6a). Conversely, if no such differences are detected, it would suggest that the parser skips the positions inside a RC during its active search for an antecedent, as predicted by the complexity account. Nominative conditions in which the pronoun c-commands the NP were included. The coreference of the pronoun and Gregory Morton violated Binding Condition C in these conditions. If Condition C is respected during online processing as Kazanina et al, observed, then Gregory Morton should not be taken as a potential antecedent at the earliest stage of dependency formation. If so, we expect no difference in the reading time of the NP Gregory Morton regardless of the gender specification in (6c) and (6d). The critical NP Gregory Morton was located inside a relative clause in all conditions, but only in the Nominative conditions was it additionally subject to the Binding Condition C.

Processing Complexity and Dependency Formation

The gender-match and mismatch sentences differed only in the gender of the pronoun in the first clause, which either matched or mismatched in gender with the subject of the second clause (Gregory Morton). This subject was the same across four conditions and it was always a gender-unambiguous proper name. Additionally, to ensure that the cataphoric pronouns received a grammatical antecedent in every case, the target structures were embedded further in a sentence introduced by the conjunction but. The gender of the third clause subject was chosen such that each sentence had a unique grammatical antecedent for the genitive and nominative pronouns. Twelve English speakers rated the acceptability of experimental materials using a 1 (not acceptable) to 5 (fully acceptable) scale. Mean ratings did not significantly differ from one another (2×2 ANOVA with factors CASE (Genitive vs. Nominative) and GENDER (gender-match vs. gender-mismatch) all Fs