The neurology of syntax: Language use without Broca's area

This is the unedited penultimate draft of a BBS target article that has been accepted for publication (Copyright 1999: Cambridge University Press U.K./U.S. -- publication date provisional) and is currently being circulated for Open Peer Commentary. This preprint is for inspection only, to help prospective commentators decide whether or not they wish to prepare a formal commentary. Please do not prepare a commentary unless you have received the hard copy, invitation, instructions and deadline information.

Abstract

A new view of the functional role of left anterior cortex in language use is proposed. The experimental record indicates that most human linguistic abilities are not localized in this region. In particular, most of syntax (long thought to be there) is not located in Broca's area and its vicinity (operculum, insula and subjacent white matter). This cerebral region, implicated in Broca's aphasia, does have a role in syntactic processing, but a highly specific one: it is neural home to receptive mechanisms involved in the computation of the relation between transformationally moved phrasal constituents and their extraction sites (in line with the Trace-Deletion Hypothesis). It is also involved in the construction of higher parts of the syntactic tree in speech production. By contrast, basic combinatorial capacities necessary for language processing - e.g., structure building operations, lexical insertion - are not supported by the neural tissue of this cerebral region, nor is lexical or combinatorial semantics.

The dense body of empirical evidence supporting this restrictive view comes mainly from several angles on lesion studies of syntax in agrammatic Broca's aphasia. Five empirical arguments are presented: experiments in sentence comprehension; cross-linguistic considerations (where aphasia findings from several language types are pooled together and scrutinized comparatively); grammaticality and plausibility judgments; real-time processing of complex sentences; and rehabilitation. Also discussed are recent results from functional neuroimaging, and from structured observations on speech production of Broca's aphasics.

Syntactic abilities, nonetheless, are distinct from other cognitive skills, and represented entirely and exclusively in the left cerebral hemisphere. Although more widespread in the left hemisphere than previously thought, they are clearly distinct from other human combinatorial and intellectual abilities. The neurological record (based on functional imaging, split-brain and right-hemisphere damaged patients, as well as patients suffering from a breakdown of mathematical skills) indicates that language is a distinct, modularly organized neurological entity. Combinatorial aspects of the language faculty reside in the human left cerebral hemisphere, but only the transformational component (or algorithms that implement it in use) is located in and around Broca's area.

Keywords

agrammatism, aphasia, Broca's area, cerebral localization, dyscalculia, functional neuroanatomy, grammatical transformation, modularity, neuroimaging, syntax, trace-deletion.

Advances in neuroimaging technology have increased our knowledge of the neuroanatomy of higher functions of the central nervous system: it is now possible to get a glimpse at our brain while it is in action. Yet this progress would not be possible without a greater understanding of the knowledge base and operations that underlie complex behavior. Parallel to the increasing precision of the technology, some progress has been made in our understanding of the cognitive architecture that underlies certain behavioral domains. In the study of brain/language relations, theoretical developments in linguistics have gone hand in hand with imaging, making a joint contribution to behavioral neurology: several decades into the study of language and the brain from a linguistic angle, there is now a relatively dense body of facts that can be seriously evaluated. This target article reviews central results, and use them to motivate some novel conclusions regarding the representation of language in the human cerebral cortex. The discussion will revolve around the choice of unit of behavioral analysis, and its theoretical import. An outlook on language that derives from current linguistic theory can draw a new and more precise picture of language and the brain.

The old Connectionist school - led by Broca, Wernicke, and Lichtheim (cf. Lichtheim, 1885, for a comprehensive exposition), and revived in our time by the late Norman Geschwind (e.g., 1970, 1979) - fortified belief in the existence of cerebral language centers. As clinicians, they mostly emphasized the patients' communicative skills, viewing language as a collection of activities, practiced in the service of communication: speaking, listening, reading, writing, naming, repetition, etc. Their characterization of the language centers derived from this intuitive theory, as for each activity they posited a cerebral center. The resulting theory of localization uses these activities as building blocks and takes them as the essence of human linguistic capacity. This view, based mostly on aphasia studies, is still held, especially in clinical circles, as illustrated by current clinical manuals and textbooks (e.g., Bradley, Daroff, Fenichel, & Marsden, 1996, p. 37). The leading neurology textbook in the United States explicitly espouses the teachings of the great neurologists of the 19^th century, depicting the perisylvian region as the location of language, and providing an internal division as follows (analytic categories of behavior are bolded):

Against this background, psycholinguists have since the 1960s attempted to devise a new perspective. As a first step, they challenged the old outlook on the centers, on the basis of theoretical and experimental tools borrowed from linguistics and psycholinguistics (e.g. Goodglass & Berko, 1960; Goodglass, 1968; Blumstein, 1972; Zurif, Caramazza & Myerson, 1974, Zurif, Caramazza & Myerson, 1972; Gardner & Zu, 1975; Heilman & Scholes, 1976; Zurif & Caramazza, 1976; Caramazza & Zurif, 1976). Focusing on the distinction between linguistic levels of representation, they did not altogether deny the validity of the old approach. Rather, they took new issues to be central, and used linguistic concepts, as well as new experimental techniques, to explore them: for them, language was no longer just a set of activities but a structure-dependent piece of knowledge, divided into levels of analysis, mainly phonological, syntactic and semantic. A variety of experiments in the 1970s proved this approach worthwhile: they obtained surprising results, and demonstrated that the brain made linguistic distinctions that could not be couched in the standard view. Consequently, an attempt was made to "redefine" the centers (Zurif, 1980): the anatomical, center-based conception was retained, yet each center was now said to contain devices used for the analysis and synthesis of language, rather than activities. Roughly speaking, the anterior language area was taken to house syntax (harnessed in the service of both comprehension and production), while semantics was believed to reside posteriorly, in Wernicke's area. Neuroanatomy also witnessed parallel advances: It was becoming increasingly clear that the anterior language area is larger than previously supposed; by then, large samples of patients became available, making comprehensive surveys possible. On this basis, the area subtending mechanisms implicated in Broca's aphasia was now taken to "encompass most of the operculum, insula, and subjacent white matter, exceeding Broca's area" (Mohr, 1976, p. 202). The schema remained localizationist, and although its units of analysis changed, the overall view of cerebral loci supporting linguistic behavior remained the same.

From the early 1980s on, new series of studies began to emerge, drawing on still finer-grained functional distinctions, and using new materials and methods to explore the language areas from a more detailed linguistic perspective (cf., for instance, Bradley, Garrett & Zurif, 1980; Schwartz, Saffran & Marin, 1980; Linebarger, Schwartz & Saffran, 1983; Swinney, Zurif & Nicol, 1989; Kean, 1980; Grodzinsky, 1984a,b, 1986, 1989, 1990; Caplan & Futter, 1986). With neuroanatomical considerations remaining constant, the most natural move was to look deeper into the details of the syntactic disruption in Broca's aphasia: syntax was a natural candidate, constituting a central combinatorial aspect of language; Broca's aphasia was the pathology of choice, because the area damaged in this syndrome was where most researchers believed to be the locus of syntax. Thus, although the view of the syndrome was changing, the traditional diagnostic schema was not abandoned, as its usefulness was proven, time and again (but see section 2.6. below for challenges to this view). I will be using evidence obtained in these experiments, enhanced by more recent findings, to reexamine the "redefined centers" view. I will show that a new, highly abstract and precise approach is necessary, to accommodate not only the fine patterns of performance that emerge, but also, to describe the cross-linguistic variation within the syndrome. This will draw a new picture of the cerebral representation of the language faculty: three decades into this line of research, it appears that the the ability to create and analyze meaningful expressions through rule-based combination is sharply distinguished from other seemingly related mental capacities (such as arithmetic or general intelligence). Mental modularity, moreover, is also a property of syntax itself: the neurology indicates that syntax is not supported by one piece of neural tissue. Within this picture, syntax is entirely represented in the left cerebral hemisphere, but for the most part, it is not located in Broca's area. This cerebral region, so the evidence suggests, has a crucial, highly specific role: it is neural home to mechanisms involved in the computation of transformational relations between moved phrasal constituents and their extraction sites.

This is a radical conclusion (at least when pitted against prevailing neurological traditions and beliefs). It is therefore important to emphasize that the evidence for it is entirely empirical, coming from a thorough survey of the aphasia literature, the literature on other cognitive impairments, and recent work in functional neuroimaging. Broca (together with a few generations of great neurologists) appears to have had the right intuitions; yet he was wrong in certain important respects that could not be understood in his day. Better understanding of the nature of language in our time, coupled with improved experimental techniques, enable us to reexamine old claims. This target article thus begins with a tour through results obtained in neurolinguistic research that have brought about the change.

The movement to redefine the language centers has had its effect in certain circles: the neurolinguistic localizing schema of language perception may not have permeated the clinical literature, yet it is currently accepted in cognitive neuroscience. On this model, syntax is represented in the part of left anterior cortex that receives its blood supply by the superior distribution of the left middle cerebral artery, (i.e., Broca's area and its vicinity as indicated above) while semantics and the lexicon are posterior, located temporo-parietally around the Sylvian fissure (Alexander, Naeser, & Palumbo, 1990; Damasio & Damasio, 1989; Zurif, 1995a). Based on an ever growing experimental record, it is now widely accepted that the speech production problem in Broca's aphasia implicates syntactic mechanisms of some sort (cf. Goodglass, 1976, 1993; Grodzinsky, 1984a; Marshall, 1986; Damasio, 1992; Zurif, 1995a); it is also believed that in comprehension, there is a deficit to receptive mechanisms of grammatical analysis (Damasio, 1992; Goodglass, 1993; Grodzinsky, 1990; Zurif, 1995a). Wernicke's aphasia, by contrast (following a posterior lesion in and around Wernicke's area), is believed to involve the lexical and interpretive components of the language faculty (Goodglass & Kaplan, 1983; Damasio & Damasio, 1992; Zurif, 1995a). Crucially, although other pathological signs are found in the overall description of many, if not most, language disorders, they are not part of the model for brain/language relations, either because they come from nonlinguistic domains (e.g., anosognosia, non-fluency, or dysarthria), or because they are not pathognomonic of any particular syndrome.

This neurolinguistic model was formed on the basis of studies of pathology, which related neuroanatomy to linguistic function by correlating impaired behavior with morphological lesion data (PM, CT or MRI). With the advent of functional imaging, evidence for the same distinctions has been sought from normal language processing. Initial findings seem to corroborate the basic approach: studies of functional imaging (PET and fMRI) have detected involvement of similar regions in syntactic processing (Bookheimer et al., 1993; Mazoyer et al., 1993; Stromswold et al., 1996; Just et al., 1996; Bavelier et al., 1997); electrophysiological studies (ERP) that chart electrical activity during the performance of syntactic tasks have likewise been consistent with this picture (Neville et al., 1991; Kluender & Kutas, 1993; Münte, Heinze & Mangun, 1993; Friederici, 1995). The movement to revise the picture from one of activity-based cerebral centers to centers representing different levels of linguistic analysis, then, has appeared to be on the right track, and has thus become the prevailing view in neuroscience.

A careful examination of the experimental evidence, which includes the more recent results, leads to conclusions that are much less definite. The data even appear contradictory at times: Wernicke's aphasics have some disturbances in syntactic comprehension (Zurif and Caramazza, 1976; Schwartz, Li, Saffran & Pate, 1987; Shapiro et al., 1993; Grodzinsky & Finkel, 1998), whereas Broca's patients, while failing certain tasks that probe receptive syntactic abilities (Goodglass, 1968; Caramazza and Zurif, 1976), have shown success in others (Linebarger, Schwartz, & Saffran, 1983). Taken at face value, these findings cast serious doubts on the model, in which Broca's area (but not Wernicke's) supports receptive syntactic mechanisms. Yet, given certain provisos, the neurological model castill be maintained. Wernicke's patients' failures in syntactic comprehension are rather inconsistent and varied (Zurif & Caramazza, 1976; Shapiro et al., 1993; Zurif, 1995a), and are by and large ignored. In the case of Broca's aphasia, moves have been made to reconcile empirical contradictions: syntactic abilities have been broken down into tasks, and it has been claimed that "syntactic comprehension is compromised" (Martin et al.,1989), and that "grammaticality judgment is intact" (Linebarger, Schwartz, & Saffran, 1983). Common to such analyses is a rather "holistic" approach to the functional deficit, details of cerebral localization being of more concern than linguistic questions. Gross distinctions between form and meaning seem sufficient, and hence, less attention is paid to more detailed structural properties of linguistic stimuli. As a result, the neurological model of language has continued to prevail.

Yet it is quite possible that the apparent inconsistencies in the results discussed above are due to the fact that many analyses lump together complex grammatical systems without distinguishing among syntactic types that are used in experiments. Seeming experimental discrepancies may thus be reconciled upon examining the structural properties of experimental stimuli. If true, this possibility could lead to a new view of the functional role of the language areas. In particular, Broca's (and perhaps Wernicke's) aphasia may become a selective deficit to receptive grammatical (i.e., syntactic) mechanisms, affecting only subsystems of the syntax (in addition to overt problems in speech production). In this view, inconsistencies in experimental results are only apparent, and disappear once the right linguistic distinctions are made.

It is hard to overestimate the implications of these conclusions (if they are valid) to the neurological study of language: data on a partial syntactic deficit not only lead to a refinement of our view of brain/language relations; they also call for an experimental linguistic approach, involving extensive and systematic use of large varieties of sentence types as test materials. Aphasia studies, which enable tests of the most fine-grained aspects of language and their neural representation, should play a central role in this kind of research program. A review of the current experimental record, and the conclusions it leads to, is thus in order. I will argue that the move from activity-based descriptions of brain/language relations to task- and linguistic-level-oriented ones is insufficient. The evidence suggests that the main language areas of the brain follow the particulars of linguistic theory. The precise functional neuroanatomy of these regions must therefore be more specific linguistically. A review of the current state of the evidence (described, of necessity, in linguistic terms) is thus in order.

2.1. Language Comprehension without Broca's area: The centrality of syntactic movement

Broca's aphasia is best known for the non-fluent, "telegraphic" speech associated with it. The comprehension problem in this syndrome is less noticeable, and harder to detect. This is probably why Broca's aphasia had initially been thought to be only a speech production problem (with obvious consequences for the diagnostic schema); not until the 1970s, when controlled experiments on comprehension began to be conducted, did a deficit in this modality become apparent. Indeed, the standard diagnostic batteries, e.g., Goodglass & Kaplan (1983) have not even acknowledged a receptive disorder in Broca's aphasia. These authors merely noted that at the early stages, Broca's aphasics "may be confused by more complex spoken messages" (p. 55). Yet, when experiments began to be conducted, they revealed - much to the surprise of investigators - a disruption in syntactic comprehension. An experiment that is taken as a landmark (Caramazza & Zurif, 1976) compared "semantically irreversible" and "reversible" object relative clauses, such as the ball that the boy is kicking is red, and the cat that the dog is chasing is black. While Broca's aphasics were successful in comprehending the former, using semantic cues to get around their deficit, they failed on the latter, indicating that they were unable to use the relevant syntax to get at the correct interpretation. This result clearly did not fit the old Connectionist model that distinguished production and comprehension anatomically, locating the former anteriorly in Broca's area, and the latter in and around Wernicke's area.

This surprising finding, and others that ensued, triggered a theoretical attempt to put the newly discovered receptive problem on a par with its expressive counterpart: as the speech production deficit was already thought to disrupt syntax, it was initially claimed that Broca's aphasics also suffered "asyntactic comprehension" (Zurif & Caramazza, 1976; Caramazza & Zurif, 1976), and that their deficit was parallel in both domains, "overarching" both production and comprehension (cf. Bradley, Garrett & Zurif, 1980; Zurif, 1980 for elaboration).

This claim shaped the debate in the early 1980s. The need to come up with more precise statements, and make distinctions among linguistic types, was becoming apparent, and thus initial attempts were made to capture all aspects of the grammatical aberration in one descriptive statement that would, moreover, span all modalities. The belief in Broca's area and its surrounding tissue as the sole locus of syntax led to a hope that patterns of impairment and sparing in speaking, listening, reading and writing would all fall under the same generalization. Thus, labor was invested in obtaining such a generalization - a unified deficit analysis of comprehension and production in agrammatism (Zurif, 1980; Zurif & Grodzinsky, 1983; Grodzinsky, 1984a,b).

However, it quickly turned out that this outlook had been overly optimistic. The pattern of selective impairment was more intricate than this account could allow, and comprehension had to be set apart from production. New experimental results were coming in, creating a rather dense body of data that was drawing a new picture, in which the comprehension deficit seemed more restricted than previously supposed: on testing, Broca's aphasics showed near-normal abilities in comprehension and grammaticality judgment on many syntactic constructions and hence did not appear to have "asyntactic comprehension". There was a disruption, but it was restricted to certain aspects of syntax. It was becoming clear, then, that a distinction between different levels of linguistic analysis would not suffice, and that distinctions within syntax were needed to account for the comprehension deficit, just as they were in speech production. Much of syntax, then, was intact in comprehension, as became evident through experiments that mostly required interpretation (and as will be shown later, the production deficit also turned out to be more selective, though in a very different manner). The one clear exception, which had actually stood out since the beginning of the experimental investigations in the late 1960s, was transformational movement in the syntax, as indicated by marked comprehension deficiencies on structures derived by such operations (cf. Goodglass, 1968; Caramazza & Zurif, 1976; Schwartz, Saffran & Marin, 1980; Ansell & Flowers, 1982 for early results). These basic results have since been fortified by massive evidence, coming from different laboratories which use diverse experimental techniques. For almost all these patients, anatomical and pathologic data are availab: common to all is a positive diagnosis as Broca's aphasics on a standardized test batteries (i.e., BDAE, WAB for English, AAT for German and Dutch), and a focal lesion to the left cerebral hemisphere, caused in the majority of cases by occlusion of the left Middle Cerebral Artery. And the repeated finding is clear: patients that are diagnosed as Broca's aphasics do suffer a receptive disorder of syntax, yet a highly restricted one: There is clear evidence for near-normal performance in moother domains of syntax, coupled with sharp failures on structures containing transformational operations.

We will now examine the experimental record in detail, for it underscores the centrality of grammatical transformations in the comprehension deficit of Broca's aphasia, and as a consequence, delineates the functional role of Broca's area in language rather precisely. First, in comprehension, Broca's aphasics can construct basic syntactic trees (phrase structure) for simple sentences that do not contain intra-sentential dependency relations, such as actives (cf. for instance, Grodzinsky, Pinango, Zurif & Drai, in press, for a review); they are also near normal in detecting violations of phrase structure rules (Linebarger, Schwartz & Saffran, 1983; Grodzinsky & Finkel, 1998). Second, the patients seem to have no impairment to their lexicon in comprehension, namely, the part of the lexicon that interfaces with sentence grammar is intact. This is demonstrated by their ability to detect violations of subcategorization (Linebarger et al., 1983) and argument structure (Grodzinsky & Finkel, 1998), and by the normal time-course of their lexical processing when argument structure is at issue (Shapiro & Levin, 1990; Shapiro et al., 1993). Third, when required to carry out tasks that involve thematic (theta-)assignment, they are successful. The theta-part of the grammar is concerned with thematic roles that a predicate assigns its arguments, and the manner by which they are linked to positions in the sentence where these arguments are realized. Simply put, each position (i.e., Subject, Object) is associated in the lexical entry of a predicate with a theta -role (Agent, Theme, Goal, Source, and Experiencer), from which the semantics of this sentential position can be recovered (cf. Haegeman, 1991, for a review). In this domain, we know that Broca's aphasics have intact abilities. They know the theta-roles of predicates, and are able to assign them directly to positions. This is evident from their normal performance in comprehension tasks that involve direct theta-role assignment in simple structures such as active sentences (cf. Grodzinsky, 1990 for an exposition). They also never violate constraints on the thematic structure of sentences (i.e., the theta-criterion) when they construct syntactic representations (Lapointe, 1985). Fourth, these patients can even compute some (but not all) intra-sentential dependencies and are able to detect violations of rules that govern them. One piece of evidence in this respect is their ability to detect case (Nominative, Accusative, Dative, and the like) on noun phrases. Mechanisms for case assignment are conditions that license the presence of NPs in their base positions. In certain languages, these conditions have overt consequences, through case markers that are realized phonetically. Broca's aphasics have virtually intact abilities in this domain, as can be seen from their performance in case marking languages, e.g., Serbo-Croat (Lukatela, Crain & Shankweiler, 1988; Crain, Shankweiler, Gorell & Tuller, 1989). They can also represent other complex intra-sentential dependencies that are defined over tree-structures. Specifically, they can handle the formal aspects of binding relations, which are the conditions that define the anaphoric relations between pronouns and reflexives, and their antecedents (cf. Grodzinsky, Wexler, Chien, Marakovitz & Solomon, 1993; See also Crain & Shankweiler, 1985). Some relations among pronouns, reflexives and their antecedents are impaired, yet these have to do with discourse-related aspects of pronominal reference, not with syntax (cf. Grodzinsky & Reinhart, 1993).

In sharp contrast with the above evidence for intact receptive abilities, we see a rather severe deficit in the comprehension of constructions derived by transformational movement, as indicated by the following familiar patterns of results, taken from dozens of experiments that investigated aphasics' interpretive abilities. In these experiments, the patient typically listens to a ("semantically reversible") sentence that contains two argument NPs and is required either to select one picture among several (including a critical foil that depicts a reversal of the actors), or to make a truth-value judgment regarding the match between the sentence and a single picture (in which the roles are either correct or reversed, see Grodzinsky, 1990, for a review). This task is believed to probe the subjects' ability to assign theta-roles to positions where the manipulated variable is syntactic structure. Several tokens (10-20) are usually presented for each sentence type, allowing the evaluation of the patients' performance relative to chance, or guessing:

This list is compiled from a large number of studies: for the basic active/passive contrast (1a)/(2a) there are results taken from 17 different studies, with at least 42 patients, all diagnosed positively as Broca's aphasics on standard tests; for the subject/object relatives (1b-c)/(2b-c) there are at least four studies of 16 patients (see Grodzinsky et al., in press, for a detailed review); three studies on 7 patients have documented the contrast between subject and object cleft sentences (1d)/(2d) (Ansell & Flowers, 1982; Caplan & Futter, 1986; Hickok & Avrutin, 1996). The contrast between (1) and (2) manifests not only in performance levels - it is also structural. As we shall see, the sentences in (2) are derived by a transformation in a way that hampers the patient's comprehension, whereas those in (1) do not. Note also that passive constructions feature in both (1) and (2), as emphasized by italics. We return to this issue later. At any rate, the contrast (1e-f)/(2e) was obtained in one study of 5 Broca's aphasic patients (Grodzinsky, Pierce & Marakovitz, 1991, see end of section 2.3.1 below). Critically, the contrast between (1) and (2) cannot be explained by appeal to familiarity, or frequency, since this type of account would need to show that (1f), for instance, is less frequent than (2a), or that (1c) is more frequent than (2c). Such demonstrations are unimaginable.

It thus appears that most aspects of syntax, whether pertaining to basic relations or to the more intricate dependencies, is intact in the comprehension of Broca's aphasics, with one salient exception: Syntactic movement - grammatical transformations.

A grammatical transformation is a complex operation over structural representations of sentences in natural language. Very roughly, it involves the copying of a constituent to another position in a sentence and the substitution of the material in the original position by a trace - a phonetically silent, yet syntactically active category that plays several important functions, two of which are relevant here: First, it is through the link between the trace and its antecedent that theta -roles (which are always assigned to canonical positions) are transmitted indirectly. Uniformity considerations force verbs to assign their assoctheta -roles in the same direction, regardless of the ordering of constituents around them. Hence, when a constituent is moved, it is through the link between it and the trace that its theta-role is transmitted. Thus in (3a) the role of Theme (recipient of action) is assigned by the verb like to its right hand constituent (=t), as it would be in an active sentence; the mutual index shared by the trace and its antecedent which man is the vehicle for theta -transmission. As a result, the latter becomes Theme:

The mediating function of traces has major ramifications for the comprehension of structures with movement that will be discussed, and become especially salient in cross-linguistic contexts. Second, constraints on movement are sometimes formulated as permissible relations between traces and their antecedents - the ungrammaticality of (3b) follows from an upper bound set on the (structural) distance between a trace and its antecedent, which was exceeded in this case. We shall see how this involvement of traces in the determination of grammaticality also impinges in important ways on the patients' metalinguistic skills. Note that the description of this syntactic relation is deliberately generic, in a manner just sufficient for present purposes. Linguistic theory evolves, at times even rapidly, and terms tend to change with theoretical perspective. Certain generalizations remain stable, however, because they deal with basic syntactic phenomena. The relationship between traces and their antecedents in movement operations plays a central role in a large class of theories. In the main, then, the presentation here is compatible with most current theoretical frameworks, including the Minimalist Program (Chomsky, 1995).

So, movement operations are implicated in the comprehension deficit in Broca's aphasia. Their impairment is also manifest in the on-line processing of structures with movement: we will see that these are will be compromised, whereas other complex processes are intact. This disruption may suggest some preliminary clues concerning the underlying cause of the comprehension deficit in Broca's aphasia. Finally, we shall see how therapeutic methods whose theoretical underpinnings are like the foregoing may be surprisingly successful.

One attempt to capture agrammatic comprehension used this general theoretical background to accommodate the main findings. The basic observation was that structures derived by movement (those in (2)) were poorly comprehended by the patients, whereas those without movement (1) yielded normal comprehension performance on tasks requiring thematic (theta -)assignment (put simply - the determination of "who did what to whom"). An initial attempt was made to partition the two structural types (Grodzinsky, 1984b; 1986). It was assumed that in agrammatism in Broca's aphasia, all traces of movement are deleted from syntactic representations. This claim has far-reaching empirical consequences for sentence interpretation, grammaticality judgment and on-line processing in agrammatism. As we shall see, considerable empirical evidence gathered from all these tasks supports this claim, suggesting that the patients suffer a disruption to only part of their syntax.

What performances follow from trace-deletion? What linguistic behaviors would a traceless language user exhibit? In interpretation, it would imply problems in movement-derived constructions and the preservation of all else; in judgment, it would predict that violations of grammaticality would go undetected if traces are crucially involved in the determination of grammaticality; and in processing, it would predict real-time problems in linking antecedents to the positions they vacated. All three predictions are borne out. What remains is to examine the distribution of traces and inquire whether trace-deletion cuts the experimental pie correctly. The demonstration will also tie this deficit to the anterior language areas, by linking this functional impairment to Broca's aphasia.

A first hint comes from an examination of the patients' performance on passive constructions. In (1)-(2), predicates with passive morphology feature (italicized) in both cells, as they generate both guessing behavior (2a,e), and near-normal behavior (1e,f). This finding correlates perfectly with trace-deletion, since passive morphology does not necessarily imply a syntactic movement transformation. Certain passives (known as "lexical" (1e,f)) are base-generated, whereas others ("verbal" (2a,e)) are derived by a transformation, hence their representations contain a trace (cf. Grodzinsky, Pierce & Marakovitz, 1991). Trace-deletion partitions these data correctly (movement <=> comprehension problem). Yet this discussion glosses over important particulars. As we will see, there are still constructions that do contain movement, but where the aphasics perform normally; there are also distinctions among types of erroneous performance that trace-deletion cannot handle. In short, a more detailed analysis of both the syntax and the experimental tasks is in order.

Consider, first of all, interpretive tasks that require theta -role assignment - the basic data set in (1)-(2). If traces mediate the transmission of theta-roles (such as Agent, Patient, Experiencer, etc.) to moved constituents, then the result of trace-deletion would be that moved NPs would lack a theta-role. This may provide a formal means of partitioning the data: impaired structures contain traces, whereas the rest do not. Yet mere partitioning is insufficient: while it points to the line dividing these structures that give the patients trouble from those that do not, the actual performance rates does not follow from just trace deletion: trace-deletion explains, perhaps, why a moved constituent does not have a theta-role, yet it does not imply chance performance on the passives or object relative clauses. On the above assumptions, the rest of the grammar is intact. This should give patients enough information to carry out a thematic assignment task and get around the deficit - they should be able to infer the missing theta -role from the rest of the available information (i.e., the theta -structure of the predicate, the fact that another NP is assigned a theta -role directly, and all other grammatical principles, cf. Grodzinsky, 1990, ch. 5; 1995a).

Another consideration that points to the inadequacy of trace-deletion is internal to the linguistic framework. Current approaches to syntax assume traces in virtually every structural representation. Even in simple active-declarative sentences there is movement of the subject from a (deep) position inside the verb-phrase into the overt position of subject, (cf. (4a), where the VP is inside bolded brackets). This is known as the VP-internal subject hypothesis (Kuroda, 1986; Kitagawa, 1986; Koopman & Sportiche, 1988): subjects are base-generated inside the VP, and are forced to move up the tree. If true, this hypothesis may have the consequence that even actives can receive a theta-role indirectly. If traces are deleted, then subjects of active sentences will have no theta -role, because the trace, under this hypothesis, may be crucial for theta -transmission. Yet such sentences pose no comprehension problems to the aphasics. Trace deletion, then, is an insufficient account of the data. Something must be done to remedy this.

Moved constituents lack a theta-role for aphasics, because of the deletion of the trace. Yet, every NP must have some role in interpretation. It was proposed that moved NPs are assigned a role by a non-linguistic, linear default strategy, which in the cases discussed assigns the Agent role to traceless clause-initial NPs (this is somewhat reminiscent of Bever's (1970) influential proposal, and similar to Jaeggli's (1986) proposal for the objects of by-phrases in derived nominals). The interaction of this strategy with the grammar results in compensation for certain structures, but confusion for the cases that are performed at chance. For example, consider a schematic representation of the verbal passive taken from (1a), and object relative clause (2b):

In (4a), the subject normally receives the role. It moves from the VP internal position, where it leaves a trace. The correct role is transmitted through the trace to the subject, which moves leftward. In Broca's aphasia, this transmission does not occur (as trace deletion precludes theta-transmission), yet the default strategy assigns the subject the Agent role, which happens to be correct, and normal performance follows. In the case of passive (4b), the oblique object (the NP argument of the by-phrase, the girl) gets the Agent role. Crucially, no syntactic movement is involved in this part of the representation, and hence, the agrammatic patient is able to assign it properly. By contrast, the subject of the passive, the boy, is movement-derived in two steps, hence two traces, both linked to the subject: first, it moves out of its base object position (t) and lands in the VP-internal subject position; second, it raises from the latter position [Spec, VP] to its final place [Spec, IP] (cf. Burton & Grimshaw, 1992 for discussion). The link between these positions is the channel for theta-role transmission. In agrammatic Broca's aphasia, both traces are deleted, hence no such channel is available, which means that the subject of a passive sentence receives no theta -role grammatically. At this stage, the strategy kicks in, assigning Agent to this NP. So we have a thematic representation with one Agent in the by-phrase (assigned grammatically) and one in the subject (assigned strategically). This situation does not allow for a unique determination of the Agent of the action, and guessing follows, resulting in chance-level performance. Next, compare the account for subject (5a) vs. object (5b) relatives:

In (5a), the VP-internal trace (t_i)is linked to a trace in subject position (t'_i), which is in turn linked to the head of the relative clause (the boy). The traces are deleted in Broca's aphasia, and the relative head lacks a theta-role. However, the strategy assigns it Agent, which is precisely what it would have received under normal circumstances. The cognitive strategy thus compensates fully for the deficit here (and in other cases in (1)), and the observed above-chance performance follows. In (5b), by contrast, the traces (t_i) and (t'_i)are also deleted and the strategy applies, yet here it gives an undesirable result: There are two Agents in the representation and the result is chance performance.

In sum, the performance rates of Broca's aphasics on the comprehension of all the above constructions is deduced by assuming trace deletion and a strategy. This combined claim has become known as the Trace-Deletion Hypothesis (TDH, whose current formulation, to be precise, is restricted to traces of constituent movement, cf. Grodzinsky, 1995a). The performance of the patients is deduced through either thematic competition, or compensation: The strategy always assigns an Agent label to clause-initial NPs. Thus, if a moved constituent is linked to a different theta -role normally (as in passive, object-gap relatives, object clefts and the like), this constituent now becomes Agent; and since there is another, grammatically assigned Agent in the thematic representation, the two Agents compete, thereby inducing chance performance by agrammatics. In cases where the moved NP was supposed to be Agent (such as subject-gap relatives, subject clefts, or actives under the VP-internal subjects hypothesis), this role is not assigned normally through the trace due to trace deletion, yet the strategy correctly compensates by assigning that NP the Agent role by default.

Direct evidence for the validity of this strategy comes from an experiment with "psychological" verbs (Grodzinsky, 1995b). When required to assign theta-roles to in sentences containing such verbs, 4 Broca's aphasics (all suffering focal lesions as a result of left Middle Cerebral Artery infarction) performed below chance (i.e., they reversed theta-roles systematically, pointing to the wrong picture most of the time) on passives of psychological predicates such as in (6) (even though they performed normally on their active counterparts):

This contrasts sharply with the chance-level performance of patients on movement-derived structures with agentive predicates. What is different about these verbs (admire, adore, love, hate, etc.) is their thematic structure - their subject is not Agent, but Experiencer (cf., Belletti & Rizzi, 1988; Pesetsky, 1995). This means that the object of the by-phrase in the passive is Experiencer. The interaction of the default strategy with the rest of the representation is different from the other cases of passive, because the nature of the competition in the representation that is created (seen in (6)) differs sharply from the previous passive cases: it is not a competition among equals anymore (i.e., Agent vs. Agent), but rather, between theta -roles with different semantic properties. The surprising contrast between agentive and psychological passive thus follows, strengthening the theoretical account.

2. Syntactic movement, and precisely that, is disrupted along the lines of the TDH:

These TDH-based observations force a new view of the role of these cerebral areas: Broca's area and its vicinity (operculum, insula, and subjacent white matter) support receptive language mechanisms that implement some, but not all, aspects of syntax, namely those pertaining to syntactic movement rules in comprehension (as well as limited aspects of tree-building in speech production). Crucially, the basic combinatorial capacities necessary for sentence processing - structure building operations, lexical insertion, and the like - are not supported by the neural tissue of these cerebral regions. This means that there is very little language in the anterior "language" area. In the next few sections, this conclusion will be fortified with evidence from several experimental angles.

The TDH handles the comprehension of English speaking Broca's aphasics, yet questions immediately arise concerning the manifestation of this disease in other languages. At birth, the brain is presumably one and the same across future speakers of different languages. But are universal grammatical principles, once cashed in as particular grammars of Arabic, Navajo or Thai, represented in adult neural tissue in the same fashion? Does the deletion of traces have identical consequences across languages? Could specific (perhaps parametrically defined) properties of a particular language correlate with the way the deficit is manifested as they do in speech produc? Is the default strategy universal? All these questions call for a comparative approach to agrammatic comprehension, for which we have some early results. The idea, then, is to probe the syntactic abilities of patients who suffer lesions to the same cerebral loci, and have a diagnosis of Broca's aphasia yet who speak languages whose structural properties differ in ways that are relevant to the deficit described by the TDH.

Comparative aphasiology must begin with a choice of dimension alwhich one compares. Languages, after all, vary along many dimensions (as described, for example, by parametric theories of grammatical variation). A reasonable place to start is in basic word-order. This is a property with several manifestations: languages may manifest SVO (Subject-Verb-Object) as their basic word-order, SOV, OSV, etc. Moreover, languages may have secondary orders (for example, an SOV language, in which an active sentence would have this order, may also have an OSV order for active sentences). The mapping from basic to secondary word-order is done, in many cases, through a grammatical transformation. This cross-linguistic variation is thus very pertinent to the description of Broca's aphasia, and to our understanding of the neurology of syntax: The application of the TDH to different structures may depend on basic word order. A comparison between languages with different basic orders - English, Japanese and Chinese - may thus provide clues regarding the cerebral organization of syntax.

Hagiwara (1993) has conducted a series of experiments on the comprehension of Japanese speaking Broca's aphasics. Her impressive findings provide an unusual angle on the deficit and underscore its restrictive nature: They show how movement structure, not the active/passive distinction, determines agrammatic comprehension performance.

The linguistic description of basic Japanese word order has been the subject of debate. One initial puzzle comes from the two kinds of active sentences it has, featuring overt SOV and OSV orders (7a-b). Japanese was initially analyzed as a "non-configurational" language (Hale, 1983), but it is now widely accepted that the SOV order is the basic, "non-scrambled" one (8a), whereas OSV (7b) is secondary, derived by a transformation that moved the object across the subject (Saito & Hoji, 1983; Saito, 1985; see also Fukui, 1993; Miyagawa, 1997). The evidence for this assumption comes from a variety of directions but is mostly based on the behavior of anaphoric expressions when they are in the different positions, which leads to the conclusion that Hanako, the object in (7b), must c-command the Verb Phrase, hence it must have moved to adjoin to a higher projection than that of Taro, the subject:

The scrambled case is derived by a movement transformation, hence its representation contains a trace. What would be the prediction of the TDH regarding agrammatic performance in Japanese? The configurations of the active sentences in (7) are given in (7'), together with the linked theta-roles for agrammatics:

If we assume that the trace is deleted, and that the operation of the strategy is like English, then it follows that in the scrambled case, the object, moved from its base position, would not have a theta -role due to trace deletion. The strategy would thus link Agent to it, by virtue of its sentence-initial position. The resulting thematic representation will be as in (7'b), and chance performance will follow. This is precisely what Hagiwara and Caplan (1990) obtained in a sentence-to-picture matching test.

The split performance on actives in Japanese shows that chance performance can be obtained for agrammatics on "simple" active declarative sentences without passive morphology or relativization, but with movement. Next, we move to the opposite case in this language - sentences with passive morphology (-(r)are), with and without movement. These are the "direct" and "indirect" passives, tested by Hagiwara (1993) in 10 Broca's aphasic patients. Again, movement is the sole determinant of performance.

According to Hagiwara, the sentence in (8a) is a standard, "direct" passive, derived by a movement transformation. The representation is annotated, and contains a trace of movement which is co-indexed with its antecedent - the moved NP Taro (with its nominative case -ga). The by-phrase is Hanako-ni. Trace deletion should result in the subject Taro not being able to receive its theta -role through the chain of movement. It is thus subject to the Default Strategy that assigns it the Agent theta-role; the by-phrase Hanako-ni is intact, just as in the English case, hence Hanako receives the theta-role of Agent. The result is a thematic representation with two Agents, and hence chance performance is predicted. Broca's aphasics indeed perform at chance on this structure, as do their English counterparts. By contrast, the "indirect" passive in (8b) is not derived by movement (as Hagiwara argues, following Kubo (1990) and others). Although it has passive morphology on the verb (-are), this construction has several properties that lead to the conclusion that it is base-generated rather than transformationally derived. For example, the by-phrase in (8b) is not optional, whereas in a regular passive it is. Following the general prediction of the TDH (movement <=> comprehension problem), Broca's aphasics perform above chance on this structure, although it contains passive morphology, is more complex (having a tree with more nodes), more loaded semantically (having two predicates), and has more words than constructions that lead to chance performance, such as the "scrambled" active, or the "direct" passive.

The results from the comprehension abilities in Japanese Broca's aphasics provide a comparative perspective that demonstrates the stability of this syndrome across languages as well as the generality of the proposed account. Moreover, it underscores the claim that the deficit in Broca's aphasia is not related to passive morphology. The data cluster in a way that cuts across this factor: Broca's aphasics score above-chance on basic actives and indirect passives (with passive morphology), and at chance on derived (scrambled) actives and direct passives, as the following table shows:

Since Broca's fail only on the bottom horizontal line, it follows that the sole determinant of the performance of Japanese Broca's aphasics is whether or not the structures contain transformational movement. This result has already been obtained for English. Specifically, it has been shown that Broca's aphasics comprehend without difficulty English sentences with passive morphology which are not derived by transformations (cf. Grodzinsky, Pierce & Marakovitz, 1991). These are sentences like (1e-f) above (the man is interested in the woman; the woman is uninspired by the man), for which a transformational analysis does not apply, for various reasons. This conclusion rules out simplistic approaches, according to which comprehension and its failure are "cue"-driven, namely, that sentence comprehension, which normally depends on overt "cues" such as passive morphology (-en) or the preposition by, is impaired in Broca's aphasics because they are insensitive to these overt markers. Finally, these results fortify the syntactic accounts of Japanese word-order which use movement rules for its analysis (cf. Kayne, 1994; Miyagawa, 1997 for recent discussions).

Yi-Ching Su (1994) has investigated the comprehension of relative clauses in Chinese in 2 Chspeaking agrammatic Broca's aphasics. Of interest is the peculiar structure of Chinese relative clauses: while the overt basic word order is SVO (as seen in the bracketed relatives in (10)), heads of relative clauses, as well as complementizers (bolded in (10)), follow the relative, contrary to English (11); this clear contrast in phrasal geometry correlates, rather interestingly, with the performance of Broca's aphasic on relative clauses in the two languages - whereas object relatives (11b) are impaired in English, and subject relatives are preserved (11a), Chinese relative clauses give the exact opposite pattern (10):

b. Object: [The man]_i who_i [the woman pushed t_i] was tall chance

The performance of Broca's aphasics is predicted only by the location of the trace, and its interaction with the strategy. Most importantly, the performance of aphasics cannot be specified through reference to construction type (i.e., subject- vs. object-relative clause) as the contrast between English and Chinese shows. This conclusion holds, despite debates regarding word order in Chinese (e.g., Huang, 1982; Travis, 1984). This is reminiscent of the discussion of Japanese, where the data could not be accounted for through a specification of constructions like active and passive, or a morphological distinction between them. The comparative examination has shown, then, that the determinants of the comprehension deficit in Broca's aphasia are phrasal geometry, the location of the trace, and the interaction of trace-deletion with the strategy. Finally, if the comprehension deficit is part of the diagnosis, parametric considerations (in this case - word order parameters) must be taken into account (see also Beretta, Hurford, Patterson & Pinango, 1996, for comprehension data on Spanish speaking Broca's aphasics that support this view). An "extensional" diagnosis of Broca's aphasia, i.e. one based on an enumeration of construction types on which the patients succeed and fail, is ruled out.

The results that have just been reviewed provide factual support for the TDH, yet they come from a single experimental perspective - that of direct tests of sentence comprehension. However, the TDH has a broader potential and can be coupled with other experimental methods. For example, traces are also involved in the operation of constraints on syntactic movement, as we saw in (3). The TDH predicts that violations of grammaticality, in which the trace is crucially involved, would go undetected by agrammatic aphasics, as the claim is that these traces are deleted from the representation, and hence cannot participate in the determination of the grammatical status of a string. Such a finding, if obtained, would provide strong evidence for the claim that the deficit manifested in trace-deletion is indeed representational, and does not follow from a deficit to the theta -module.

Schwartz, Linebarger, Saffran and Pate (1987) have obtained results that provide preliminary clues regarding this issue. They conducted a "plausibility judgment" study in which their patients were asked to make judgments about whether or not sentences made sense. Their design intersected syntactic movement with semantic plausibility, and with length: regarding the semantic factor, the patients were given semantically implausible sentences in which the source of the implausibility was either one of the arguments (12a) or two (12b); syntactically, there was an argument that was either in situ (12a), or moved transformationally (12c-d); lengthwise, the semantic violation was also nested in a "padded" sentence that contained many words (12e), but no movement. Naturally, each violation had a plausible counterpart:

They tested 5 patients, all diagnosed as Broca's aphasics on standardized tests. Of these, 4 had lesions that resulted from occlusion of the left Middle Cerebral Artery, and one had temporo-parietal closed head injury. The results were remarkable: the patients performed relatively well on most conditions, detecting and rejecting implausibility, and accepting plausible sentences. They failed (and did so rather badly) only on a subset of the sentences that contained transformational movement. These were sentences in which the source of implausibility was one transformationally moved argument (12d). By contrast, the aphasics were able to detect plausibility violations if an offending argument was not moved by a transformation (12a,c). Thus in (12a) the cause of implausibility is the puppy, an argument which is not moved; in (12c), even though one source of implausibility, the table, is moved (hence undetectable), the detection of the anomaly is possible through the other argument, the spoon, which is also semantically anomalous, but which is not moved by a transformation.

The Schwartz et al. study, while important, assessed judgment abilities only indirectly, as it involved a mix between lexical semantics and movement in a task that necessitated interpretation, rather than pure judgment of grammatical well-formedness. Its findings are consistent with the TDH but are open to other interpretations as well. It is possible, for instance, that traces are not missing from patients' representations, but rather, are inaccessible to interpretive tasks (theta -transmission), and that this is the reason for the patients' failures. Schwartz et al.'s result, then, is inconclusive as far as the theory is concerned. Thus, in order to investigate this matter directly, a study was recently conducted, aimed at assessing grammaticality judgment abilities in aphasia, where the structures presented were violations of constraints on movement and a large group of controls (Grodzinsky & Finkel, 1998). We tested aphasic sensitivity to violations of constraints on movement of constituents (NP and Wh-movement), each case coming with a set of its own grammatical controls (cases 1 through 4 in (13)). We further investigated aphasic ability to detect violations of constraints on head (verb) movement (cases 7-8 in (13)), for which we had independent evidence: agrammatic aphasics are capable of representing traces of such movement (cf. Lonzi & Luzzatti, 1993, among others). Finally, we also included control conditions - violations of other grammatical principles (cases 5-6) tmake sure that our patients were able to carry out the task:

We tested 4 non-fluent, agrammatic Broca's aphasic patients, all diagnosed as such, with lesions in and around Broca's area, including white matter deep to it, ranging from the operculum, to the anterior limb of the internal capsule, to the peri-ventricular white matter. Our results uncovered a fine, highly restricted deficit, which follows directly from the TDH: the patients were very alert to grammaticality in general, yet they failed when required to compute constraints on the movement of phrasal constituents (conditions 1-4). In each of these conditions, the patients made around 40% errors. All others (condition 5-8) were intact (around 10% error). Interestingly, a recent study of grammaticality judgment in Serbo-Croat reported results that went in the same direction (Mikelic, Boskovic, Crain, & Shankweiler, 1995). Finally, Wernicke's patients, while producing a slightly different pattern, were not particularly successful in making grammaticality judgments, despite the fact that their deficit is usually thought of as a semantic, not syntactic. The conclusions that follow are clear:

A fourth, critical angle on the deletion of traces is that of real-time processing. Information regarding the time-course of language comprehension in aphasia is obviously crucial to our understanding of the underlying pathology. Indeed, extensive investigations of aphasic behavior in real-time have been conducted in recent years (e.g., Bradley et al., 1980; Milberg & Blumstein, 1981; Swinney, Nicol & Zurif, 1989; Friederici, 1985; Shapiro & Levine, 1990; Shapiro et al., 1993, to mention just a few). Shapiro and his colleagues have provided a compelling demonstration of the claim that the overall processing capacities of Broca's (but not Wernicke's) aphasics in comprehension are intact. They conducted a series of studies on their lexical abilities, and their manifestation in real-time sentence processing (Shapiro & Levine, 1990; Shapiro et al., 1993). They showed that Broca's aphasics are normal in the way they handle verbs, in that their on-line processing routines make fine distinctions among verb types exactly the way normal speakers do. This was not the case in Wernicke's aphasia, however. Posterior lesions, then, unlike those in Broca's area, do cause damage to the lexicon.

It is by now well established that normal language users demonstrate trace-antecedent relations in real-time tasks (e.g., Swinney et al., 1988; Stowe, 1986; Swinney & Nicol, 1989; Bever & McElree, 1988; McDonald, 1989; Nagel & Shapiro and Nawy, 1994; Tanenhaus, Carlson & Trueswell., 1989; Love & Swinney, 1996; Swinney & Zurif, 1995). The typical experiment exploits priming effects to uncover antecedent reactivation. The leading idea is that the link between a trace and its antecedent means that in the course of comprehension, the antecedent is reactivated at the trace. Thus in (14), the drink will be active when heard (namely, at 1), will then decay (2), but will get reactivated following the verb (3), due to its link to the trace.

(14) The priest enjoyed the drink¹ that the caterer was² serving t³ to the guests

This is precisely what on-line experiments on normal language users have discovered. Through different methods - Cross Modal Lexical Priming (CMLP) being a central one - reactivation of antecedents in the position of their traces has been shown for subject- and object-relatives (Swinney et al., 1982; Swinney & Nicol, 1989; Swinney & Osterhout, 1990; Swinney & Zurif, 1995), for passive (McDonald, 1989), and for other structures. The experiments take the drink as prime, and while the sentence unfolds auditorily, a target is flashed on a screen at points (1)-(3). The expected finding, then, is that if a target word, say, juice, is presented visually to subjects at points (1), (2) and (3) when they are listening to the sentence, and the subjects have to make a lexical decision on it, priming effects will be documented at (1) and (3), but not at (2).This is what is found: priming effects are obtained only in (1) and (3).

Now, consider the TDH and the expected real-time behavior of Broca's aphasics in CMLP. Deleted traces mean no reactivation at the trace. This means that only in point (1) would a priming effect be obtained. Decay would explain the lack of an effect in (2), and the correlate to trace-deletion would be a lack of priming in (3). Conducting such experiments is quite difficult, yet this is precisely the result of a series of carefully controlled studies of both subject and object relative clauses (Zurif et al., 1993; Swinney & Zurif, 1995). Importantly, Broca's aphasics do prime, even if not in a fully normal fashion (e.g., Swinney, Zurif & Nicol, 1989; Shapiro & Levine, 1990; Shapiro et al., 1993)). Yet, when faced with a task that involves priming within a movement-derived construction, they are seriously impaired. Finally, this failure is not characteristic of all aphasics, nor is it necessarily related to general comprehension skills: Wernicke's aphasics with posterior perisylvian lesions perform normally on this task, even though their comprehension abilities are severely compromised.

We have gone through various sorts of evidence, among which tests of real-time syntactic analysis. This review led to a new delineation of the deficit. But can this lead to an unambiguous statement regarding the underlying cause? An answer to this question would lead to an explicit theory of the function of the anterior language areas. The question, obviously, is whether or not this area supports a device dedicated to syntactic analysis of transformationally moved constituents, and if so, what this device is. One possibility is that the comprehension deficit follows from a general disruption to "working memory", not from a language mechanism. A recent PET finding (Jonides et al., 1997) indicates that in nonlinguistic memory tasks that require the subject to relate two non-adjacent members in a list, Broca's area gets activated in a secondary fashion. To some this may sound like proof that this area supports a memory cell, the disruption of which entails the syntactic deficits in this syndrome, yet such a conclusion is a bit hasty. Although it may turn out to be truin some future, "final" analysis, such a conclusion is not warranted on the basis of the available evidence. To argue for a disruption to a generalized memory resource, one must show that this resource makes contact with formal constraints on the inner workings of this memory store from which constraints on syntactic movement would follow. Moreover, we have seen that other real-time performances of Broca's aphasics show how intact their language processing device is. Lexical access and insertinto sentential positions - found intact by Shapiro and his colleagues - demonstrate directly how specific the impairment is. We may then move on, and consider other, language specific, interpretations, namely: Is Broca's area home for a processing device whose disruption precludes antecedent reactivation at the right time? Or is it, perhaps, a representational medium that makes silent categories disappear? It is not possible, at this point, to decide between these options, yet either way, we have delineated the deficit and have shown that any deficit analysis, hence any theory of the role of Broca's area, must have the consequences of the TDH.

The foregoing discussion has been based on aphasia results, which allow for a precise characterization of the role of Broca's area through deficit analyses. Recent technological developments in neuroimaging are beginning to make it possible to see pictures of the brain in action. The current neuroimaging literature makes only tenuous connections to the lesion-based body of knowledge, and focuses, for the most part, on semantic and phonological aspects of the mental lexicon. Fewer works concentrate on the most salient aspect of language - its combinatory nature. Fewer than 10% of the works presented in a recent Neuroimage conference (10/101) investigated the computation of combinatorial operations in language processing. Still, there are some, perhaps preliminary, experiments on functional imaging of sentence processing (Mazoyer et al., 1993; Stromswold et al., 1996; Just et al., 1996; Bavelier et al., 1997). The evidence they present is fully consistent with the new picture presented here.

Consider, first of all, the most linguistically detailed study available - Stromswold et al.'s PET investigation. They measured blood flow during visual exposure to sentences whose plausibility the subjects were requested to judge. The relevant conditions included several experimental variables: relative clauses that were either (plausible or implausible) right-branching subject gaps (15a-b), or (plausible or implausible) center-embedded object-gaps (15c-d):

What can be made of this result regarding the functional role of this cerebral region? An examination of the stimulus materials reveals three dimensions: plausibility, type of relative (center-embedded vs. right branching), and gap location (subject vs. object position). Broca's area lit up as a consequence of an interaction between all three, hence no conclusion that separates these three factors can be made. As a consequence, the meaning of this result is not entirely clear, nor is there an apparent connection to a theory. Still, one conclusion can be prudently drawn: the PET finding is consistent with the claim that the computation of transformational relations is done in Broca's area, because one of the experimental variables was the location of the gap.

Stromswold et al.'s study is more linguistically sophisticated than any of the other available investigations of functional imaging of language activity. Other studies do not make detailed claims possible. Thus Just et al. (1996) used fMRI to test neuronal activity during language comprehension. They presented sentence pairs as in (16), to which the subjects were expected to make a truth-value judgment. The experimental sentences were chosen by a "complexity" measure that the authors do not specify. They included three types of fairly complex stimuli: actives with conjoined VPs (16a), center-embedded subject relative clauses (16b), and center embedded object relatives (16c), each followed by a question (italicized) to which the subjects were requested to respond by pushing a yes/no button:

Several results are reported: first, while signals were picked up in Broca's and Wernicke's areas of the left hemisphere upon exposure to all types of stimuli, their right hemispheric homologs remained relatively silent; second, Broca's and Wernicke's areas were equally activated; third, there appeared to be a difference in number of activated voxels among the three sentence types. Stimuli (16a,b,c) engaged an increasing number of voxels in this order. All cases involved very complex materials: apart from a (transformationally derived) question in every case, there was a transformation in both types of relative clauses, and a complex construction, involving VP-conjunction, in the third. Comparisons are thus far from being straightforward. Still, it is important to emphasize that while these results do not lead to specific conclusions, they are consistent with the position espoused here.

Next, Mazoyer et al. (1993) looked at the brains of normal French speakers during language understanding. They exposed their subjects to stories in unfamiliar language (Tamil), to French word-lists, well-formed sentences containing non-words, semantically anomalous sentences, and stories in French (composed of sentences whose grammatical properties are not given). Right and left temporal regions were activated in most conditions, yet only left inferior frontal regions, in and around Broca's area, lit up on exposure to words and stories in the subjects' language (French). Listening to stories in French, moreover, activated the left middle temporal gyrus significantly more than its right hemispheric homolog.

Finally, Bavelier et al. (1997) also conducted an fMRI study, with stimuli described as "short declarative English sentences (mean length = six words, range = four to nine words) and...consonant strings of equivalent length to the words used (mean length - four letters)." Another control was a film of a speaker of American Sign Language, signing the same sentences. Here, again, it was found that the right hemisphere remained silent, while the left was active on exposure to sentences in a familiar language.

Since the choice of stimuli for this experiment was not based on structural considerations, only general conclusions regarding the cerebral representation of language mechanisms are possible. Vague as they are these results are fully consistent with the more detailed claims that are made on the basis of lesion studies. The same is true of all the imaging studies we reviewed: they are all consistent with the claim that the linguistic factor associated with "hot spots" in Broca's area is not syntax as a whole, but rather, mechanisms that underlie syntactic movement. Critically, there is nothing in the functional imaging results that contradicts the above conclusions, or speaks against the TDH-based view of the limited role of Broca's area in language processing.

The degree of generality of the foregoing is appreciable: dozens of studies were reviewed here, including many neurologically intact subjects as well as brain-damaged patients, tested in different laboratories at different times, in different countries and languages, on a broad variety of sentence types. The tasks reviewed come from comprehension in several languages, grammaticality and plausibility judgment, and real-time processing. The results converge. Still, it is important to note that sstudies yield seemingly contradictory results. These divide in two: failures to replicate previous results, which have sometimes led to the claim that the syndrome of Broca's aphasia gives an altogether inconsistent picture (e.g., Martin et al., 1989; Druks & Marshall, 1995; Berndt et al., 1996); and those that document performances that run contrary to the TDH but accept basic premises. (e.g., Hickok, Zurif & Canseco-Gonzales, 1993; Hickok and Avrutin, 1995). Both types of presumed inconsistencies must be taken very seriously: an unstable syndrome (and certainly a non-existent one) is the wrong object of inquiry; likewise, a false hypothesis is, most likely, the wrong one to follow. It must be revised, perhaps even abandoned, when confronted with data that cannot be accounted for. Thus, an old debate has been revived in recent years, in which the coherence of the clinical categorization, as well as the validity of the have TDH been challenged (cf. Badecker and Caramazza, 1985; Miceli et al., 1989 for attacks on the syndrome-based conception; also Hickok et al., 1993; Lukatela, Shankweiler & Crain, 1995; Berndt, Mitchum & Wayland, 1997, for critiques of the TDH on the basis of new data). Many (if not most) objections have been explained away: in some cases patient selection was the cause of the inconsistency, as patients that were not Broca's aphasics were included in the experimental group; in others, there were problems with experimental procedures or design; in still others, the results may have been misanalyzed or misinterpreted (cf. Caplan, 1986; Zurif, Gardner & Brownell, 1989; Grodzinsky, 1991; Zurif, 1996).

Still, even after this type of cleanup, a certain amount of inter-subject variation persists. Grodzinsky, Pinango, Zurif and Drai (in press) have recently conducted a survey of comprehension scores of Broca's aphasics in two contrasts that pertain to the TDH: actives vs. passives and subject-gap vs. object-gap relative clauses. For both sentence types there are multiple studies: the review covers 17 different studies of active/passive, with 42 different patients, and 4 studies of subject/object relatives and clefts, with 17 different patients. All the comprehension experiments we analyzed had a binary choice design (containing the correct response and its thematic reversal). The expectation from the results of the active sentences was a pattern approaching 100% correct. For the passive case, chance performance was expected; thus, since patients are said to be guessing, the performance of each patient should be equated with a series of (unbiased) coin tosses, as such series are exactly the expression of chance behavior. As such series are known to distribute binomially, with a median of around 50% correct, we expect our patients to follow this distribution. Moreover, we expect a statistically significant difference between the performances on actives and passives. The analysis of the actual data reveals that this is the case: the actives cluster around the 100% mark; the passives distribute binomially with a mean and median of 55%; the two differ significantly from one another. Figure 1 shows the distribution of performances for actives (dashes) and passives (dashes and circles), where the latter is compared to a computer-generated model of an averaged binomial distribution (full line):

Figure 1: the distribution of performances for 42 Broca's aphasics (# of patients vs. % correct) for active (dashes) and passive (dashes and circles)sentences. Full line is a computer-generated simulation of a binomial distribution (adapted from Grodzinsky, Pinango, Zurif & Drai, in press).

An analysis of subject/object relative clauses and clefts produces similar results. The data thus fit the theoretical expectations, and the analysis further demonstrates that constrained within-group variation is permissible, as long as the group picture leads to scores that, on precise measurement, yield the expected results. This analysis also shows why single-case studies can be very misleading: Extreme cases, like Druks and Marshall's (1995) patient who performed worse on actives than on passives (cf. Zurif, 1996 for a critique) cannot be evaluated seriously outside a group. It is only in this context that one can judge whether they conform to an expected pattern. Outside this context, they are seen as exceptional, which may not be the case. Our analysis of the current empirical record suggests that an approach to aphasia that allows constrained variation seems to be on the right track: the weight of the evidence indicates that the TDH has had, as far as is known, the broadest empirical coverage. Still, certain core cases have forced reformulation.

Moreover, it has been found that in certain cases, Broca's aphasics are not at chance on the comprehension of certain (transformationally derived) passives, giving rise to yet another surprising asymmetry (5 patients in Saddy, 1995 and Balogh & Grodzinsky, 1996). We have already seen how Broca's comprehend non-transformational passives at near-normal levels. At issue here is a case of a different type: when the subject of a passive is a quantified expression (18a), near-normal performance is yielded by the very same patients who are at chance on a "regular", agentive passive sentence (18b), again, contrary to the TDH, which is indifferent to the properties of the antecedent of the trace:

Both findings are inconsistent with the TDH as stated. It has, therefore, been revised to accommodate them (cf. Grodzinsky, 1995a). The account builds on a common property that both cases share. Specifically, the observation is that both antecedents (who in (17a) and every boy in (18a)) have a common semantic property that sets them apart from regular NPs. These two subjects fall under the default strategy that assigns the Agent role to moved phrasal constituents that have no theta -role. So the idea of the revision is to restrict the scope of the strategy and condition its application to a theta-less NP - an antecedent of a trace - on the semantic properties of that NP. The proposal is that this semantic property (which relates to the way these elements interact with discourse semantics) precludes the strategy from being applied in such cases, on semantic grounds. Whether or not this new account is valid remains to be seen. What is important, though, is that it is this type of account that leads to the reformulation of new empirical questions, and thus to a systematic enrichment of the data base. Indeed, the overaexperimental picture has gradually built more structure into the relationship between linguistic behavior and the neural tissue that supports it.

Finally, a question arises regarding the syndrome-specificity of the results. In particular, is the deficit as characterized by the TDH specific to Broca's aphasia? Can it be found in Wernicke's aphasia, or, perhaps, subsequent to damage to the right hemisphere? That the latter is not the case will be shown in section 5.1. Regarding Wernicke's aphasia, the record seems somewhat mixed: Wernicke's aphasics do not always carry out syntactic tasks successfully, yet their performance does not fall under the TDH. That is to say, the way they fail is much less uniform, and in most cases very large variation is observed for them. are, admittedly, fewer studies dedicated to the syntax in Wernicke's than in Broca's aphasia, but the available results are compelling (see Swinney & Zurif, 1995; Shapiro et al., 1993 for Broca's/Wernicke's difference in real-time sentence processing; Grodzinsky & Finkel, 1998 for judgment; Zurif & Caramazza, 1976, Grodzinsky, 1984b for comprehension). There are, perhaps, some syntactic abilities in the posterior language area, yet they are less tangible and characterizable than those in the left frontal cortex, which lead to clear conclusions, as we have seen.

The focus of this paper is comprehension. Yet Broca's area, so the evidence suggests, subserves not only receptive mechanisms of language, but also, language production. In fact, the most salient feature of Broca's aphasia has always been effortful, non-fluent and telegraphic speech. Below, I review some recent results in the production of language in Broca's aphasia, because they emphasize 2 points: first, they show the difference between the deficits in expressive and receptive mechanisms of language; second, they again show the need for an abstract grammatical approach to language and the brain.

Consider, first of all, a salient cross-linguistic difference in the behavior of agrammatic Broca's aphasic speakers, and the way they reveal how linguistic distinctions are honored by the brain. The omission of functional elements in speech production has always been the hallmark of the diagnosis of agrammatism in Broca's aphasia (e.g., Marshall, 1986; Adams & Victor, 1993, p. 417; Goodglass, 1976, 1993). Yet it is well established that patients either omit or substitute inflectional elements, but only if the morphology of the ambient language permits such omissions, as the patients observe rules of lexical well-formedness, and never produce non-words or word-parts (Grodzinsky, 1984a, 1990): in both verbal and nominal inflected elements omissions of inflectional morphemes are observed if a bare stem is a real, licit, word (+zero morphology); otherwise (i.e., in words that are -zero morphology, where omission of inflection is illicit and results in non-words), substitution errors occur:

The same cerebral pathology manifests differently, depending on the morphology of the language spoken: in English and Japanese we observe the familiar pattern of omission of the inflectional morphology, whereas in the rest of the languages we see an inflectional element substituted for another, with a resulting grammatical aberration, but lexical structure is unimpaired. Zero-morphology is thus relevant: Broca's aphasia in a language where stems are legal words presents with omissions; otherwise, there are substitutions. Grammars (or sub-grammars) align themselves accordingly:

(20)	+zero morphology	-zero morphology
	English	Hebrew
	Japanese	Russian
		Italian
	*omission*	*substitution*

An account of these phenomena must be grammatical and abstract. If we assume that inflectional features (theta-features) are underspecified in the syntactic representation of agrammatic Broca's aphasics (Grodzinsky, 1984a, 1990), we get errors of inflection, whose type depends on the ± zero-morphology property of a given language: Underspecified features in a +zero-morphology language would result in omission, whereas in a -zero-morphology language the result would be substitution. There is, then, a varied manifestation of the syndrome, which correlates perfectly with an identifiable linguistic property:

Since omission has always been a critical diagnostic sign of agrammatism in Broca's aphasia, clinicians cannot ignore the variation just presented if they are interested in a precise and universal diagnosis. They must accordingly enter the conditional statement in (21) into the clinical diagnosis of agrammatic Broca's aphasia to allow for variation in error type as a function of grammatical (morphological) properties of the language spoken. Similar to the cross-linguistic TDH, the description of speech production errors across languages leads to a parametric definition of the syndrome, one in which grammatical parameters (imported from a theory of possible grammatical variation) are embedded. We return to this issue below.

PASSIVE MORPHOLOGY MOVEMENT	-	+	PERFORMANCE LEVEL
-	basic (unscrambled) active	indirect passive	above-chance
+	derived (scrambled) active	direct passive	*chance*

(13)	CONDITION	+Grammatical	-Grammatical
	1. NP movement	It seems likely that John will win It seems that John is likely to win John seems likely to win	*John seems that it is likely to win
	2. Wh-movement/ that-trace	Which woman did David think John saw? Which woman did David think that John saw? Which woman did David think saw John?	Which woman did David think that saw John?*
	3. Superiority	I don't know who saw what	*I don't know what who saw
	4. Adjunct/complement	When did John do what?	*What did John do when?
	5. Filled gaps	Who did John see? Who saw John?	Who did John see Joe? Who John saw Joe?
	6. Bad complements	The children threw the football over the fence The children sang	The children sang the football over the fence* *The children threw
	7. Place of auxiliary	They could leave town Could they leave town? They could have left town Could they have left town? They have left town Have they left town?	*Have they could leave town?
	8. Negation	John has not left the office John did not sit	John did not have left the office John sat not