Derivational Order in Syntax: Evidence and Architectural Consequences

generative grammars have traditionally been studied under the assumption that they are implementation independent. It is assumed, for example, that eventually we will be able to write software or build hardware to implement the human language system in computers that differ from human ‘hardware’ in many respects. As such, this is taken as evidence for implementation independence. This, however, is clearly an empirical question. It is a real and attractive possibility that computational modelers of the future might succeed in that effort, but it is far from having been accomplished at this time. The fact that some grammars that approximate fragments of human language have been digitally implemented does not settle the question. A relevant model should minimally be completely descriptively adequate, and one might reasonably want the model to approximate human linguistic abilities in more ways than simply classifying (un)acceptable sentences. Such models currently inhabit the realm of thought experiments. Here we disagree with Neeleman & van de Koot (2010), who agree with us on the importance of implementation (in)dependence, but take it as quite significant that some natural language grammars could be implemented in different artificial devices, concluding that human language is implementation independent. In the meantime, we argue that the emphasis on the implementation independence of generative grammars is misplaced if the purpose of the endeavor is to understand the nature of the human system. If a research community makes the choice to carve out one sub-area of the study of human language abilities, e.g., classifying (un)acceptable sentences, and then finds that this sub-area can be described in implementation independent terms, then this surely does not entail that the sub-area corresponds to a privileged, implementation independent sub-area of the human language faculty. There is no doubt that it is interesting and useful to develop explicit computational models of human language, but such models cannot show whether human grammatical abilities are, in fact, independent of their cognitive or neural implementation. Regardless of what one’s personal hunch is about the likelihood of successful non-human implementation of language, a more interesting question for our purposes is whether it is implementation independent within humans. According to the standard formulation of the principled extensionalist view, speakers of a language have knowledge of the (un)acceptable sentences of their language, and this knowledge can be described at an abstract level by a generative grammar. Importantly, it is assumed that this abstract knowledge is implementation independent, and that speakers can put their knowledge to use in different ways in activities such as speaking, understanding, and internally generated speech. This means that for any given well-formed sentence structure defined by the grammar a speaker may have a number of different ways of mentally assembling the structure. If this assumption is supported, then the principled extensionalist is justified in separating speakers' knowledge of what is well formed from their knowledge of how to assemble wellformed structures. But if speakers do not, in fact, have multiple ways of constructing 2 Neeleman and van de Koot describe a thought experiment in which a pair of computers with radically different underlying architectures are both able to pass a kind of Turing Test for human language, such that their performance is indistinguishable from human native speakers of the language. They argue that one would want to say that both computers speak the relevant language, and conclude that this is because the computers and the humans would share an abstract, implementation-independent grammar. We agree with the intuition behind this thought experiment, but it seems to reveal more about the common usage of predicates like “speaks French” than about the nature of the human language faculty. Despite this misgiving, we find much to like about Neeleman and van de Koot’s discussion of the challenges involved in understanding grammatical theories at different levels of description. Derivational order in syntax Colin Phillips & Shevaun Lewis 16 the same representations, then human language appears to be more implementation dependent. The evidence on whether speakers have multiple ways of constructing the same sentence representation is not extensive, but there are a number of reasons to think that speakers have one and only one way of assembling any individual representation. There is almost no evidence for the alternative view that speakers have multiple ways of building the same representation. First, speaking and understanding proceed in the same (roughly) left-to-right fashion. Although they have different goals, they have a great deal in common, appear to construct the same representations, and plausibly do so in the same order, although this topic has not been investigated in great detail. Second, in comprehension and production there is much evidence that speakers build structures and interpretations incrementally, in roughly the order in which words are presented. We are unaware of evidence that speakers are able to construct the same interpretation in different orders. For example, reading backwards is a task that lies somewhere in the difficult-to-impossible range, despite its correspondence with the derivational order assumed in many generative grammars. Third, in cases of reanalysis in sentence understanding, where comprehenders are led into a ‘garden path’ and must then reorganize their initial parse, some evidence suggests that speakers repair by returning to the error point and simply re-parsing the sentence in exactly the same order that it was presented (Inoue & Fodor, 1995; Grodner et al., 2003). We therefore adopt the working hypothesis that natural language grammars are implementation dependent with respect to how sentences are assembled: there is only one algorithm for structure-building across all human speakers (of the same language). We would certainly welcome more systematic evidence, but currently the evidence for the alternative implementation independent position is practically non-existent. The implementation dependent position is the simpler and more falsifiable hypothesis, and hence should be preferred until proven otherwise. Consequently, we think that the principled extensionalist position is unwarranted, and that the motivations for developing abstract generative grammars are more pragmatic than principled. We have no problems with pragmatic motivations, and we recognize the value of focusing on characterizing good/bad sentences as a way of making headway in describing human language. But this is very different from the position that the characterization of good/bad sentences is a fundamentally separate enterprise from understanding realtime language processes. Ultimately we seek theories that capture how sentences are put together, and not just what their final form is. In discussions of these issues we sometimes encounter an objection in the form of a 'slippery slope' argument. If human language is better described in terms of real time cognitive processes rather than abstract functions, so the argument runs, then why stop there – why not continue all the way down to the level of neurochemical processes in brain cells or beyond? As with other slippery slope arguments, the expectation is that we should find this consequence appalling, and hence should drop the entire argument. We acknowledge the concern, but disagree with this argument. First, we are not arguing against the usefulness of abstract descriptions in the study of language. We find them exceedingly useful. We are simply arguing that there is no privileged level of abstraction – the level occupied by most current generative grammars – that is 3 Although we have focused our attention here on the interpretation of derivational grammatical theories, the issue of implementation (in)dependence is equally relevant to interpreting grammatical theories that assume no derivations. Non-derivational syntactic theories, which are often presented as preferable due to their order-neutrality, imply implementation independence. If syntactic representations are not, in fact, implementation independent, then that should count against nonderivational theories, just as it counts against bottom-to-top derivations. STiL – Studies in Linguistics Vol. 6 17 exempt from considerations of psychological implementation. Second, it should be emphasized that implementation (in)dependence is not an all-or-nothing property of an abstract system. Rather, a system's implementation (in)dependence must be evaluated at each successive degree of abstraction. Our focus in the current discussion is on the relation between symbolic descriptions of the structure of sentences and symbolic descriptions of the procedures for assembling sentences in real time. This leaves open the possibility that human language is implementation independent at a lower level. For example, there are interesting arguments that the basic notion of hierarchical structure in language must be implementation independent with respect to its neural encoding, because of the very different demands of immediate vs. long-term encoding of sentences in memory (Jackendoff, 2002; van der Velde & de Kamps, 2006). Briefly, the most plausible method for long-term information storage in the brain – through changes in synaptic connectivity – is too slow for the millisecondscale processes needed for real-time language use. The need for dual encodings of the same structures suggests that structured representations are not always neurally encoded in the same fashion, and hence that they are implementation independent. It is far from certain that this argument goes through, because it is not clear that immediate and long-term encodings of sentences in memory are genuinely isomorphic to one another. But this is the type of argument that might one day lead to a clear finding of implementation independence for human language (for one specific level of implementational detail). In sum, the slippery slope argument is misplaced in the current case. We should also emphasize that we do not claim that a grammatical theory is inherently preferable if it can be transparently linked to real time processes. Standard considerations of descriptive adequacy (i.e., generating all and only the acceptable sentences of a language) are as relevant as ever, and all theories should be accountable to that standard. A grammatical theory that achieves impressive descriptive adequacy using mechanisms that are opaque to real time implementation is an interesting theory nonetheless. However, the descriptive success of such a theory begs for further analysis of why it works so well, and whether its success crucially depends on the opaque mechanisms. This analysis should ideally lead to development of a more psychologically transparent version of the theory. We take up this challenge below for arguments that have been presented in the syntax literature in favor of bottom-to-top derivations. 4. Psychological Aims of a Real-time Grammar Having examined the psychological commitments of standard generative grammars, it is appropriate to apply similar scrutiny to grammars that adopt roughly left-to-right derivations and that aspire to be models of real time processes. Here it is important to address some possible misconceptions. (One of us bears some blame for spreading some of the misconceptions.) First, the slogan ‘the parser is the grammar’ (Phillips, 1996) sounds nice enough, but it is unfortunately misleading, as it too closely identifies the grammar with the task of language comprehension. It would probably be more appropriate to regard a real-time grammar (the ‘structure builder’) as one important component of a parsing/comprehension system, but certainly not the whole system. We envision a structure building system that combines words and phrases to form sentences and meanings in real-time, in essentially the same manner in comprehension and production. This system is task-neutral, and it could even operate in the absence of external input or a communicative goal, i.e., neither comprehension nor production is necessary. The system is, of course, put to use in speaking and understanding, where the structure that it builds is constrained by the external input (comprehension) or by the speaker’s message (production), but these tasks require far more than just a structure building system. Parsing requires some way of using incoming sounds and words to determine which specific structure building operations are appropriate. Derivational order in syntax Colin Phillips & Shevaun Lewis 18 Production requires some way of mapping the elements of an intended message to operations in the structure building system. It appears that the human parser is fairly good at mapping external input onto appropriate structure building operations, and the human production system appears to be similarly effective at selecting structure building operations that are appropriate to the speaker's communicative goals. But the assumption that natural language grammar has the form of a real-time structure building system does not logically entail the success of either parsing or production. Consequently, some caution is needed when using evidence from parsing or production to assess the nature of the structure builder. Second, and closely related to the first point, the claim that the mental grammar should be understood as a real-time structure generation device does not guarantee that it is part of a perfect parsing device. When we claim that the real-time structurebuilding device is the mental grammar, we predict that the representations that this device constructs should be grammatically well formed, and that it should incorporate whatever detailed grammatical machinery we would normally expect of a grammar. This means that the system should not construct the rough-and-ready representations that some have argued to be created in the service of rapid, efficient comprehension (e.g., Ferreira & Patson, 2007; Townsend & Bever, 2001). However, the claim that the representations that are built on-line are grammatically precise does not entail that they are the same ones intended by the speaker. In extreme cases where a listener is distracted or in a noisy environment, he might use his mental grammar to construct a perfectly reasonable representation that bears only a weak relation to the sentence that the speaker uttered. This would show that the listener is failing to make full use of the input, but it would not license any clear conclusions about the nature of the listener’s real-time structure building system. Consequently, when we examine studies of online satisfaction of grammatical constraints, the key prediction for current purposes is that the comprehender constructs grammatically well-formed representations, even if those representations are not a grammatically possible parse of the incoming sentence. We will have more to say on this point in a moment. Third, the claim that the grammar has the form of a real-time structure building system is independent of long-standing psycholinguistic questions about how speakers resolve syntactic ambiguities in language comprehension. Structural ambiguities arise when the input to the comprehension system has two or more well formed structural analyses, i.e., they are cases where the grammar alone cannot decide. Ambiguity resolution has enjoyed a prominent position in psycholinguistic research on sentence comprehension, and there have been long and heated debates over which types of information are brought to bear in resolving ambiguity (e.g., Frazier, 1987; MacDonald, Pearlmutter, & Seidenberg, 1994; van Gompel & Pickering, 2007). There is, in fact, broad consensus that simpler parses are favored in ambiguity resolution. The controversy surrounds what it means to be ‘simpler’ (e.g., structurally simpler, more frequent, more semantically or pragmatically natural, etc.) and what it means to be ‘favored’ (i.e., the unique parse vs. highest ranked parse among multiple parses pursued in parallel). These discussions are interesting, but they are orthogonal to our claims about the form of the grammar. Fourth, we have been surprised at how often claims about the procedural nature of the grammar are interpreted as claims that grammatical phenomena are epiphenomenal. In linguistics and psycholinguistics one often encounters proposals that some phenomenon that has traditionally been analyzed in formal terms can instead be explained in terms of constraints on parsing or production (e.g., Kluender & Kutas, 1993; Hawkins, 1994; Harris & Bates, 2003; Hofmeister & Sag, 2010). These reductionist accounts of linguistic phenomena are often referred to as ‘processing accounts’, as a way of contrasting them with ‘formal’ or ‘grammatical’ accounts. They are claims that the phenomenon in question does not fall under the purview of the mental grammar. Reductionist analyses of grammatical phenomena are interesting, and we find some types of evidence more compelling than others (for discussion see STiL – Studies in Linguistics Vol. 6 19 Sprouse, Wagers, & Phillips, 2012; Phillips, 2013ab), but they are quite different than the models that we are advocating here. Since reductionist accounts are explicitly nongrammatical in nature, they clearly have little in common with the claim that grammatical derivations follow an order that happens to be well suited to comprehension and production. Finally, we should clarify the reason why we repeatedly describe grammatical derivations as proceeding in a roughly left-to-right order. Sentences are spoken and heard in a strict left-to-right order. (Strictly speaking this is a tautology, as left-to-right order is merely a conventional representation of the temporal order of words in speech.) However, it is probably not the case that mental structure building operations perfectly follow the linear order of a sentence, whether in comprehension or production. To take just one example, in a head-final language such as Japanese it may be necessary for the structure building system to create a position for the head of a phrase before it has completed the arguments and adjuncts that precede the head. More generally, structure building in comprehension is probably not entirely synchronized with the appearance of words in the input. There is growing evidence that comprehenders often build structural positions in their parses before encountering the words in the input that phonologically realize those positions (Aoshima, Phillips, & Weinberg, 2004; de Long, Urbach, & Kutas, 2005; Lau et al., 2006; Mazuka & Itoh, 1995; for review see Lau, 2009), and some evidence for related effects in production (Momma, Slevc, & Phillips, 2013). The upshot of this is that it may not even be desirable to insist upon a strict left-to-right order for grammatical derivations, since the operations of the real-time structure builder may not proceed in a strict leftto-right order. If it is the case that there is a single structure-building system that assembles sentences in a strict order, then it is likely that this order will turn out to be only roughly left-to-right. What matters is whether the order of structure building operations is consistent. 5. Is On-Line Structure Building Grammatically Precise? If the representations that are built during real-time comprehension and production differ from those that are motivated by standard grammatical analysis, then we face prima facie evidence for multiple structure building systems, and thus evidence against our claim of a single procedural grammar. Fortunately, there is now much experimental evidence that bears on this question, and it mostly supports the notion that real-time processes assemble syntactic representations that are the same as those motivated by grammatical analysis. We should reiterate here that the question of how grammatically sophisticated real-time structure building is must be kept distinct from the question of how accurately the parser analyzes the input. A parsing system could, in principle, build wonderfully rich grammatical representations that are nevertheless rather poor analyses of the incoming strings of words. Many classic psycholinguistic studies from the 1960s and 1970s addressed the question of whether on-line language processes create representations with the detailed structural properties that were proposed in the newly emerging field of generative grammar. This body of work is most often remembered for its failure to find correlates of the transformational derivations found in theories of that time, but a 4 The actual findings from these studies were less definitive than is typically claimed in historical reports (e.g., Townsend & Bever, 2001, ch. 2). The tests focused on a rather narrow linking hypothesis, and many of the conclusions relied on specific grammatical analyses that have not stood the test of time. Tests of the 'Derivational Theory of Complexity' (DTC) examined the narrow linking hypothesis that the perceptual complexity of a sentence is best predicted by the number of steps in its transformational derivation. But given the many different factors that contribute to the perceptual complexity of a sentence, there is little reason to expect that derivation length should be the primary Derivational order in syntax Colin Phillips & Shevaun Lewis 20 more enduring outcome is that the same studies often found experimental support for the surface structure representations of the time (for reviews see Fillenbaum, 1971; Fodor, Bever, & Garrett, 1974; Levelt, 1974). For example, hierarchical clustering analyses of speakers’ relatedness judgments for word triads from a sentence yielded a good approximation to the surface structure of a sentence (Levelt, 1970), and studies on the perceptual mis-location of non-speech clicks played during a sentence confirmed a difference between object control sentences like The general defied the troops to fight and exceptional case marking constructions like The general desired the troops to fight (Bever, Lackner, & Kirk, 1969). Far more recently, studies that use highly time-sensitive measures such as eventrelated brain potentials (ERPs) have made it possible to track how quickly comprehenders are able to detect different types of anomaly in the linguistic input. This work has shown that speakers detect just about any linguistic anomaly within a few hundred milliseconds of the anomaly appearing in the input. Different types of grammatical anomalies elicit one or more from among a family of different ERP components, including an (early) left anterior negativity (‘(e)LAN’; Neville et al., 1991; Friederici, Pfeifer, & Hahne, 1993) or a P600 (Osterhout & Holcomb, 1992; Hagoort, Brown, & Groothusen, 1993). Many questions remain about what the different components reflect and what determines which components are evoked in any individual situation (Hagoort, 2003; Friederici & Weissenborn, 2007; Lau, Phillips, & Poeppel, 2008; Federmeier & Laszlo, 2009; Gouvea et al., 2010; Brouwer, Fitz, & Hoeks, 2012), but for current purposes the most relevant outcome from this research is that more or less any grammatical anomaly elicits an ERP response within a few hundred milliseconds. If the on-line analyzer is able to immediately detect any grammatical anomaly that it encounters, then it is reasonable to assume that it is constructing representations that include sufficient grammatical detail to detect those anomalies. Another body of on-line studies has examined whether on-line structure building respects various grammatical constraints, i.e., whether the parser ever creates grammatically illicit structures or interpretations. Many studies have found evidence of immediate on-line effects of grammatical constraints, such as locality constraints on wh-movement (Stowe, 1986; Traxler & Pickering, 1996; Wagers & Phillips, 2009), and structural constraints on forwards and backwards anaphora (Kazanina et al., 2007; Nicol & Swinney, 1989; Sturt, 2003; Xiang, Dillon, & Phillips, 2009; Lewis, Chow, & Phillips, 2012; Dillon et al., in press). These findings extend to complex cases that present apparent challenges for incremental application of grammatical constraints, such as constraints on backwards anaphora in Japanese, where the constraints must apply before any verb has appeared in the input (Aoshima, Yoshida, & Phillips, 2009), and constraints on parasitic gaps inside complex subjects in English, where the parasitic gap precedes its licensor (Phillips, 2006). Findings such as these imply that the structures created on-line include sufficient structural detail to allow the constraints to be applied during parsing. The many different types of evidence for on-line grammatical sensitivity do not, of course, strictly require that the structures that are built on-line are exactly those that predictor of processing difficulty. Nevertheless, early evidence based on some transformations that have stood the test of time, such as passivization and subject-auxiliary inversion, provided surprisingly good support for the DTC's predictions, and these findings were not effectively challenged, despite reports to the contrary (for review see Phillips, 1996, ch. 5). Subsequent failure to find perceptual complexity increases caused by such operations as the transformation that converts a full relative clause ("the house that was red") into a pronominal adjective ("the red house") are probably not surprising. We would not want to argue that the DTC was substantiated, but the reports of its defeat strike us as somewhat stylized history. STiL – Studies in Linguistics Vol. 6 21 are sanctioned by the grammar, or that the system that builds them is a procedural grammar. It is always possible that the on-line structure builder is not, in fact, identical to the grammar, but instead is a very effective ‘covering grammar’ for the true grammar of the language. But to the extent that the on-line structure builder constructs exactly the right representations to capture both on-line behavior and standard acceptability judgments, then we see little motivation for postulating an independent grammar that yields no additional empirical coverage. There are, however, a number of findings in the psycholinguistics literature that have been taken to indicate divergence between the structures created on-line and those motivated by traditional grammatical analysis. We will briefly review three types of evidence, from misinterpretations, syntactic priming, and grammatical illusions. Comprehenders frequently misinterpret the sentences that they encounter. Fernanda Ferreira and colleagues have argued that this is a desirable property for a comprehension system, and that the misinterpretations are the result of ‘good enough’ (GE) representations, which contrast with representations that are “detailed, complete, and accurate with respect to the input” (Ferreira & Patson, 2007, p. 71; see also Ferreira, Bailey, & Ferraro, 2002). One recent event nicely illustrates the intuition behind Ferreira’s argument. On October 2 2008 a record TV audience (70 million in the US alone) watched the debate between the vice presidential candidates, Senator Joseph Biden of Delaware and Governor Sarah Palin of Alaska. In a segment on energy and climate change Palin said, “I’m not one to attribute every activity of man to the changes in the climate”. Probably most viewers immediately recognized what Palin intended to convey with this quote (that climate change is not primarily caused by human activity), and few would have even noticed that the literal interpretation of her statement is quite different (that human activities are not exclusively caused by climate change). Palin’s slip-of-the-tongue was (roughly) a classic exchange error, and comprehenders successfully recovered Palin’s message despite the error, presumably because they had a good idea of what the message was likely to be. Ferreira and colleagues argue that situations like this show GE representations in action, and straightforwardly illustrate one of the benefits of an interpretive system that is not a slave to the precise surface form of incoming sentences. Ferreira and colleagues offer additional evidence of systematic misinterpretations from experimental studies. For example, they tested what interpretations speakers take away from garden path sentences in which they initially misparse and must then reanalyze, e.g., While Anna dressed the baby played in the crib (Christianson, Hollingworth, Halliwell, & Ferreira, 2001). When questioned after the sentence, speakers reliably agreed that the sentence stated that the baby had played, but they also agreed on a substantial proportion of trials that the baby was dressed, suggesting that in their final interpretation the single NP the baby filled two conflicting thematic roles. They refer to this as a ‘lingering’ garden path effect (see also Sturt, 2007). Similarly, Ferreira reports that comprehenders often incorrectly judge that the surface subject is the ‘do-er’ in passive sentences like The dog was bitten by the man (Ferreira, 2003). We agree with Ferreira that the misinterpretation of good sentences and the successful repair of speech errors are important phenomena that show that on-line interpretation does not always deliver the correct meaning of an incoming sentence. But as we have 5 In this study comprehenders were probably not simply following a plausibility-based heuristic, as they also showed very similar error rates in judgments of non-reversible passives, such as The cheese was eaten by the mouse. Only in the most difficult sentence types (object clefts) did error rates increase in the manner predicted by a plausibility-based heuristic. Derivational order in syntax Colin Phillips & Shevaun Lewis 22 emphasized above, there is a difference between failure to correctly analyze the incoming string of words and failure to build a grammatically well-formed representation in response to the input. As psycholinguists we are certainly very interested in the question of how effectively people parse, but this is separate from the question of whether the on-line structure builder generates only grammatically wellformed representations. The evidence reviewed by Ferreira does not show that comprehenders assemble grammatically ill-formed representations. We suggest that in the case of ‘lingering interpretations’ of garden path sentences what happens is that comprehenders incrementally update their interpretations over the course of the sentence, but that interpretations are not labeled based on the pieces of syntax that generated them. Consequently, when a syntactic parse undergoes reanalysis to a complete and well-formed structure, rescinding of the incorrect syntactic analysis does not lead to rescinding of any interpretations that were previously generated by that syntactic analysis. Interpretive repair of speech errors, as in the example from the vice presidential debate, presents a slightly different situation, since the repairs occur in situations where the comprehender has a good idea of the intended meaning and hence the speech input is redundant. In such cases the interpretive system likely completes its task before the syntactic parse has finished. It is interesting that the interpretive system is able to do this, but it tells us little about the nature of the syntactic representation that is generated in these situations, which may be perfectly grammatically well-formed. A related observation that may be relevant here is that different types of speech errors appear to impact language comprehension in different ways. In sentences that are syntactically appropriate but semantically garbled, comprehenders are sometimes able to recover the intended interpretation without even noticing the error, as in the example above. In contrast, when speech errors lead to syntactic deformation of a sentence, comprehenders are generally able to recover the intended form, but the error does not pass unnoticed. This suggests that comprehenders may be able to skip detailed interpretation of an incoming sentence, but that they cannot skip syntactic analysis, even if it is not needed for interpretation. A second potential argument for a mismatch between on-line structure building and the structures sanctioned by the grammar comes from syntactic priming. In language production, many studies have shown that the use of a particular syntactic structure increases the likelihood that the same structure is used in subsequent utterances, even when there is no lexical overlap between the utterances that have the same structure (Bock, 1986; Pickering & Ferreira, 2008). This finding has led to interesting questions about what pairs of structures count as the ‘same’ for purposes of syntactic priming. An influential study by Bock and Loebell (1990) showed priming between VPs containing PPs that differ in their thematic roles, e.g., VPs with a locative PP such as The wealthy widow drove the Mercedes to the church primed VPs with a recipient PP such as The wealthy widow gave the Mercedes to the church, and VPs with a locative by-phrase such as The 747 was landing by the control tower even primed full passives such as The 747 was alerted by the control tower. (Note that we use lexically matched examples here for illustrative purposes only. The adjacent prime-target pairs in these studies did not overlap in this way.) In addition, many other studies have shown that syntactic priming does not depend on overlap in the function words (including prepositions) between primes and targets (Bock, 1989; Fox Tree & Meijer, 1999). Taken together, the evidence therefore suggests that relatively coarse-grained syntactic parallelism is sufficient to cause structural priming. This must somehow be 6 This account clearly begs the question of how the relevant interpretations are accessed or generated, if not through detailed compositional interpretation of the sentence structure. Investigation of this question may well lead to the conclusion that the language comprehension system has multiple ways of using cues to generate interpretations. However, this is quite different from the conclusion that there are multiple representational systems or multiple real-time structure building procedures. STiL – Studies in Linguistics Vol. 6 23 reconciled with the evidence that motivates linguists to postulate fine-grained structural distinctions between superficially similar sentences. Different responses to this challenge are possible, and they remain to be resolved. The structural priming evidence could be taken as support for grammatical models that make less finegrained structural distinctions (Culicover & Jackendoff, 2005). Or it could be taken to show that the representations involved in structural priming are not those defined by the grammar (we do not favor this option, but it is a logical possibility). Alternatively, it could be taken to show just that the structural priming paradigm is a relatively blunt tool for investigating structure, because relatively coarse-grained similarity between structures is sufficient to cause structural priming. This would leave open the possibility that on-line processes build fine-grained structures. A third potential motivation for distinguishing the structures built on-line and those sanctioned by the grammar comes from grammatical illusions, cases where comprehenders appear to fleetingly accept structures that are judged bad after more reflection. The most notorious case of a grammatical illusion involves comparative constructions such as More people have been to Russia than I have, which are semantic gibberish but initially sound remarkably good. Townsend and Bever (2001) argue that such cases make an interesting case for a system that distinguishes a roughand-ready initial analyzer from the more fine-grained analyses of the grammar. A number of additional cases of illusory acceptability have emerged in recent years. One case involves the spurious licensing of the negative polarity item (NPI) ever by non c-commanding negation, as in The bills that no senators have supported will ever become law (Drenhaus, Saddy, & Frisch, 2005; Xiang, Dillon, & Phillips, 2009). Another case involves illusory agreement licensing in which a plural-marked verb is judged to be acceptable in the vicinity of a plural NP that is not its syntactic subject, as in The runners who the driver see ... or The key to the cabinets probably are ... (Clifton, Frazier, & Deevy, 1999; Pearlmutter, Garnsey, & Bock, 1999; Wagers, Lau, & Phillips, 2009). Yet another case involves evidence that during the processing of pronouns comprehenders fleetingly consider a clause-mate subject NP as a potential antecedent, in violation of Binding Principle B (Badecker & Straub, 2002; Kennison, 2003), although findings are mixed, and a number of other on-line studies report immediate effects of Principle B (Clifton, Kennison, & Albrecht, 1997; Lee & Williams, 2006; Nicol & Swinney, 1989; Runner, Sussman, & Tanenhaus, 2006; Lewis et al., 2012). For a more detailed review of where grammatical illusions do and do not arise see Phillips, Wagers, & Lau (2011). Grammatical illusions pose a challenge to our hypothesis of a real-time procedural grammar to the extent that they justify a mismatch between on-line structure building processes and processes that operate under less time pressure. However, demonstrations of illusory acceptability do not automatically show such a mismatch. Grammatical illusions could have a number of sources other than the existence of multiple structure building systems. They could reflect mis-parsing, where the comprehender builds a perfectly grammatical representation that happens to not match the input. Alternatively, they could show that the normal workings of the structure building system are slow enough that on-line methods can probe the intermediate steps of the computation. Wellwood and colleagues have recently argued that illusory comparatives reflect mis-parsing rather than construction of an ill-formed internal representation. In acceptability rating studies they find that the illusions are more robust with predicates that are ‘repeatable’ such as go to the gym than with predicates that are ‘non-repeatable’ such as won the lottery yesterday, and suggest that this is 7 These comparative illusions, which some have referred to as Escher sentences, were first pointed out by Montalbetti (1984), although he presents them as a curiosity and does not offer an analysis of the phenomenon, which has not been systematically studied until recently. Derivational order in syntax Colin Phillips & Shevaun Lewis 24 because the sentences are mis-interpreted as event quantification rather than as quantification over individuals (Wellwood et al., 2013). Notably, English grammar allows the syntax of individual quantification to be interpreted as event quantification in certain contexts, as in The Washington DC metro carries more than 200 million passengers per year, which is a claim about person-trips rather than about distinct individuals. The other cases of grammatical illusions that we have listed here are probably not amenable to a mis-parsing analysis, but might nevertheless reflect the normal operations of a grammatically accurate structure building system. The temporary consideration of local antecedents for pronouns that violate Principle B (as found in some, but by no means all on-line studies of Principle B) may be directly related to a natural grammatical implementation of the constraint, applying as a filter that marks candidate referential dependencies as illicit, rather than as a constraint that prevents the generation of illicit candidates. Since pronouns may be associated with a wide range of syntactic and discourse antecedents, a mechanism that generates candidates and then excludes inappropriate candidates may be the most feasible way of applying Principle B, and it may be that the fleeting consideration of illicit antecedents found in some studies reflects this mechanism in action. More broadly, ‘generate-and-filter’ mechanisms are familiar from many grammatical theories (e.g., Chomsky, 1981; Legendre, Grimshaw, & Vikner, 2001), and hence are plausible components of a realtime grammar. The case of illusory NPI licensing remains poorly understood, as few NPIs and only a narrow range of syntactic and semantic contexts have been tested to date, so it is unclear how far illusory NPI licensing extends beyond the specific environments examined so far. However, Xiang and colleagues have attempted to show how the illusions could reflect the operations of the same pragmatic licensing mechanisms that are widely held to be responsible for normal NPI licensing (Xiang et al., 2009). According to this account, the illusions reflect inappropriate pragmatic inferences, rather than use of rough-and-ready licensing mechanisms. In the case of illusory agreement licensing Wagers and colleagues have argued that the illusions arise from the use of grammatically fully appropriate feature retrieval operations in an architecture with noisy memory representations (Wagers et al., 2009). We should emphasize that systematic investigations of grammatical illusions are relatively new and that none of these accounts are definitive. However, they show that it is possible to account for the phenomena using mostly standard structure-building and interpretation mechanisms, without recourse to independent on-line and off-line systems. One clear advantage of the single-system accounts is that they predict that only certain types of temporary illusions should be possible. Accounts that invoke a separate rough-and-ready structure building system are less constrained. 6. Bottom-to-top derivations We have argued that human grammars are implementation dependent, at least with respect to how syntactic structures are assembled, and that it is desirable and plausible to view grammars as procedural systems that can be understood in terms of actual real-time mental computations. However, most contemporary syntactic theories seem incompatible with this view: they assume that grammatical derivations proceed in a strictly bottom-to-top order, progressively combining words and phrases starting from the most deeply embedded elements. For most languages this yields derivations that proceed in a mostly right-to-left order, i.e., the opposite of the order in which comprehension and production operate. We must therefore address the evidence for bottom-to-top derivations, to determine whether it undermines our approach. We have found, to our surprise, that the evidence is neither extensive nor particularly well known. The assumption of a bottom-to-top order of derivation seems to have arisen from some reasonable intuitions about human language. However, it is an assumption that has rarely been revisited over the years. Bottom-to-top is widely regarded as the only possible order of derivation, whereas it ought to be considered as just one among several possible ways to account for certain facts about language. STiL – Studies in Linguistics Vol. 6 25 Probably the most influential intuition underlying this pervasive assumption is that lexical items are inserted into thematic positions. That is, a structure “begins” as a direct representation of thematic relations, which is subsequently transformed to satisfy other syntactic requirements. This intuition has its origins in the kernel sentences of Syntactic Structures (Chomsky, 1957), and it was codified in the Standard Theory as Deep Structure (later ‘D-structure’ or DS), a level of representation that “defines grammatical relations and functions in a straightforward manner” (Chomsky, 1970). It was the point in the derivation at which the argument selection requirements were satisfied (the ‘Theta Criterion’). Empirically, DS was crucial to the distinction between raising and control. In raising constructions like (1), the subject bears a thematic relation to the embedded verb, but not the main verb. In control constructions like (2), the subject has a thematic relation to both verbs. It was claimed that raising constructions are formed by movement from a thematic position in the embedded clause to a non-thematic position in the main clause, whereas control constructions involve an empty category in the embedded clause that is not created by movement, but rather is created by equi-NP deletion or, in later parlance, the empty category PRO. (1) The mani seemed ti to enjoy his ice cream. (2) The mani tried PROi to enjoy his ice cream. D-Structure was argued to be unnecessary as part of the Minimalist Program: the Theta Criterion was a requirement that might as well be satisfied at LF (Chomsky, 1993). Nevertheless, the new theory essentially reinstated it at the point of lexical insertion by requiring thematic roles to be assigned only at external Merge (Chomsky, 1995; but for a different view see Hornstein, 2000). Applying the theta criterion at the point of lexical insertion is admittedly a fine way of capturing these data, and a bottom-to-top order of derivation follows from this. In derivations of this kind, satisfaction of the thematic requirements of a lexical item is guaranteed, and all further transformational operations serve to satisfy other syntactic requirements (encoded as “features”). However, this approach is by no means required. Stripping away theory-particular terminology, the aim is to account for the fact that arguments enter into multiple syntactic relations within a single structure. Usually only one of those relations (the one lowest in the structure) is thematic, except in the case of control constructions. These facts could just as well be captured by the opposite restriction: arguments must merge into a non-thematic position that satisfies syntactic requirements such as case, agreement or scope marking, and subsequent transformations could relate that argument to one or more lower positions in the structure where thematic and other requirements are satisfied. Arguably such operations are required under either framework: even in bottom-to-top theories, the original thematic relations must be re-established at LF for interpretation. A second intuition underlying bottom-to-top derivation is that the endocentricity of syntactic structures is established by heads “projecting” phrases. The features of a head are said to be “passed up” to the maximal projection. Again, while such terminology is a convenient way of describing endocentricity, it is certainly not required. Bare Phrase Structure theory (cf. Chomsky, 1995) highlights the possibility that different levels of projection simply encode the relations between the head (a lexical item) and other parts of the structure. In the X’-style structure in (3), it might seem necessary to ensure that the features of the head ‘travel’ up the tree to give content to the otherwise empty categorial nodes. In the bare phrase structure in (4), the maximal projection of loves has no need to inherit features from the lexical item: it is that lexical item, represented in relation to other lexical items. In this approach, the labeling of levels of projection is merely a mechanism for identifying the dominant node, which determines how the phrase behaves as a whole. Endocentricity is a condition on the structure as a whole, and need not be explained by requiring the Derivational order in syntax Colin Phillips & Shevaun Lewis 26 projection of features from bottom to top. Mechanisms for regulating the match between the features of heads and maximal projections are familiar from constraintbased grammatical formalisms such as Head Driven Phrase Structure Grammar (HPSG), and can be straightforwardly implemented without the need for an ordered derivation.