Thematic Hierarchies In Argument Realization English Language Essay

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

The discussion of the semantic determinants of argument realization in previous chapters centerebut interactions between arguments that affect argument realization cannot be ignored. These interactions suggest that there are precedence - or prominence - relations among arguments statable in terms of their semantic roles. This chapter examines the possible sources of these relations through an in-depth examination of the construct known as the thematic hierarchy - a ranking of semantic roles chosen. Three major questions are the focus of this chapter: which notion of prominence underlies the ranking that defines a particular formulation of the thematic hierarchy, which linguistic phenomena is it supposed to account for, and can its effects be derived from more basic components of a lexical semantic representation?

We show that two major conceptions of prominence find their way into analyses of linguistic phenomena that appeal to a thematic hierarchy. On the first, the ranking of semantic roles is determined by structural properties of a lexical semantic representation - properties defined over a predicate decomposition or event structure. On the second, the ranking is determined by entailments associated with arguments. Our examination of these two conceptualizations supports the conclusions of earlier chapters that the mapping from lexical semantics to syntax needs to make reference to arguments in both ways. We show that the two notions of prominence are sometimes used to account for distinct phenomena, though the hierarchy as defined by either notion is a derived, rather than primitive construct.

to argument realization. Therefore, untangling the sources of these various

formulations should help shed light on the nature of prominence itself.

Distinct formulations of the thematic hierarchy can be compared only if

the ranking among the roles is given the same interpretation and the ranking

is intended to explain the same range of phenomena. As we show, proposed

hierarchies differ from each other significantly in both respects. The thematic

hierarchy is sometimes used in the algorithm which maps from lexical semantics

to syntax (see section 5.3.2), but, as we discuss in this chapter, thematic

hierarchies are put to other uses. Furthermore, even when different instantiations

of the thematic hierarchy are designed for use in a mapping algorithm, the

rankings of semantic roles defining each instantiation can be compared only

if they are posited in the context of similar assumptions about the nature of

syntax and its relation to a lexical semantic representation.

As we also discuss, in some instances a thematic hierarchy is simply used

to provide a shorthand for capturing a local empirical generalization underlying

a specific linguistic phenomenon; such a hierarchy cannot be taken to represent

a universal construct. We show that sometimes such phenomena are accounted

for in terms of semantic properties other than semantic roles and the thematic

hierarchy effects are only apparent, arising because the relevant semantic properties

distribute across traditional semantic roles in a systematic way. In other

instances, the ranking of roles captures a valid empirical generalization which

is the result of a complex interaction of factors. We conclude that in general

a thematic hierarchy is a notational device that is given various interpretations

and is put to a variety of uses, and, hence, there is no single universal ranking of

arguments which will capture all valid generalizations expressing regularities

in the association of semantic roles with morphosyntactic realizations.

6.1 Thematic hierarchies: appealing, but problematic

Thematic hierarchies have proved appealing because they allow an argument

of a verb to be referred to in terms of its relative position on the thematic

hierarchy, instead of in terms of its semantic role. The advantage of this way

of referring to arguments is illustrated throughout this section and the next.

In previous chapters, we showed that a range of semantic roles can be

associated with both subject and direct object and that even for a given verb

arguments bearing various semantic roles can be realized as subject and object.

Since there is so much variability, formulating general algorithms which determine

which of a verb's arguments is its subject or direct object presents a

real challenge. Fillmore's subject selection rule in (1) represents a well-known

early attempt to confront this challenge.

(1) If there is an A [=Agent], it becomes the subject; otherwise, if there is an I

[=Instrument], it becomes the subject; otherwise, the subject is the O

[=Objective, i.e., Patient/Theme]. (Fillmore 1968: 33)

This rule implicitly establishes precedence relations among the semantic roles

mentioned and could be simplified to (2), if used in conjunction with the

hierarchy of roles in (3) (Bresnan and Kanerva 1989; GivÙŽn 1984b, 2001;

Kiparsky 1985; Grimshaw 1990).

(2) The argument of a verb bearing the highest-ranked semantic role is its subject.

(3) Agent > Instrument > Patient/Theme

Consistent with Fillmore's algorithm, (2) has the effect of realizing a verb's

agent, if there is one, as its subject; in the absence of an agent, it allows an

argument bearing another semantic role to be the subject. The generality of

(2) - for instance, its lack of reference to specific semantic roles - makes

mapping algorithms that appeal to a thematic hierarchy attractive.

Besides subject and object selection, accounts of a range of other phenomena

involving argument realization have also appealed to thematic hierarchies,

including passivization (Bresnan and Kanerva 1989; GivÙŽn 1990; Grimshaw

1990; Jackendoff 1972; Trithart 1979), causativization (Carrier-Duncan 1985;

Dik 1980; Foley and Van Valin 1984; Polinsky and Kozinsky 1992) and other

morphosyntactic processes (Alsina 1994, 1999; Alsina and Mchombo 1993;

Bresnan and Kanerva 1992; Carrier-Duncan 1985; Hawkinson and Hyman

1974; Morolong and Hyman 1977), serial verbs (Baker 1989; Carstens 2002),

compounding (Foley and Van Valin 1984; Grimshaw 1990; Y. Li 1990; Potter

1991), and light verbs (Grimshaw and Mester 1988). For instance, Bresnan and

Kanerva (1989: 27, (64)) propose that passivization involves the suppression

of the argument of a verb that bears the highest semantic role, an idea which

receives similar, though not identical, expression in Grimshaw (1990), while

Carrier-Duncan (1985: 14, (38)) proposes that the rule of reciprocal verb formation

in Tagalog is "assign a reciprocal interpretation to the lowest thematic

role and bind it to the next higher one."

Researchers have also proposed that the thematic hierarchy figures in the

statement of typological implicational generalizations about the range of semantic

roles that can be realized as subject or object in a given language, as well as

in the statement of generalizations concerning which subject or object choices

are most likely to receive a "marked" expression (Asudeh 2001; Dik 1978,

1980, 1997a; GivÙŽn 1984a, 1984b, 2001). In this connection, thematic hierarchies

have figured, often in combination with person and animacy hierarchies,

in generalizations concerning the surface manifestations of argument realization

- what Keenan (1976) calls "coding" properties, including unmarked word

order (K. P. Mohanan and T. Mohanan 1994; Siewierska 1988, 1993; Uszkoreit

1987), agreement (Duranti 1979; Evans 1997; Hawkinson and Hyman

1974; Morolong and Hyman 1977; Trithart 1979), inverse marking (Aissen

1999) and case marking (Aissen 1999; Joppen and Wunderlich 1995). As this

list suggests, thematic hierarchies are invoked in the analysis of phenomena

from a wide range of languages, including Albanian (Sells 1988), Chinese

(Y. Li 1990), Dutch (Dik 1980), English, German (Uszkoreit 1987), Greek

(Everaert and Anagnostopoulou 1997), Hindi (K. P. Mohanan and T. Mohanan

1994), Japanese (Grimshaw and Mester 1988), Tagalog (Carrier-Duncan 1985),

Kwa languages (Baker 1989), and Bantu languages (Alsina 1994, 1999; Alsina

and Mchombo 1993; Bresnan and Kanerva 1989, 1992; Bresnan and Moshi

1990; Duranti 1979; Hawkinson and Hyman 1974; Morolong and Hyman 1977;

Polinsky and Kozinsky 1992; Trithart 1979).

Thematic hierarchies also figure in the accounts of a range of other linguistic

phenomena, including anaphoric relations (Everaert and Anagnostopoulou

1997; Kuno 1987; Jackendoff 1972; Schwartz 1986; Sells 1988; Van Valin

and LaPolla 1997; W. Wilkins 1988) and control (Chierchia 1983; Nishigauchi

1984). We do not, however, discuss these other uses.

Following the methodology set out by Dowty (1991), we restrict our attention to the hierarchy as it is used in argument realization, on the assumption that this focus is most likely to uncover a unified conceptual characterization of the hierarchy as it pertains to one particular domain (though, as we show, even in this restricted domain there is no unified notion of a thematic hierarchy). Furthermore, as far as we are aware, most attempts to appeal to a thematic hierarchy in other domains have been unsuccessful in that the generalizations have been shown either to follow from other factors or to be spurious. For instance, Nishigauchi (1984) argues that the thematic hierarchy helps determine the controller of a PRO subject of an infinitival complement; however, Ladusaw and Dowty (1988) reexamine the data and argue that "the verb's entailments and facts about human actions always determine the controller" (1988: 68).

The apparent relevance of the thematic hierarchy to the characterization of numerous phenomena in diverse languages has made it an attractive theoretical construct. It has been adopted by proponents of a range of theoretical frameworks, including Functional Grammar, Lexical-Functional Grammar, the Principles and Parameters framework, and Role and Reference Grammar. But the attitude towards the thematic hierarchy largely parallels the attitude towards the semantic roles which comprise it. It is convenient to posit a thematic hierarchy when it seems to be implicated in robust linguistic generalizations, but

acknowledgment of its usefulness is tempered by an uneasiness concerning its status. This uneasiness has two sources. First, if, as discussed in chapters 2 and 3, semantic roles are derived constructs, then any hierarchy defined in terms of them should also be a derived construct. The question then arises as to what the ranking of the arguments is derived from. Second, and more problematic, is the lack of agreement over the appropriate formulation of the thematic hierarchy, in terms of both the roles constituting it and the ranking of these roles. These considerable disagreements have led some critics to reject the notion outright (Newmeyer 2002: 65).

In this chapter we review the primary uses of the thematic hierarchy, while

bearing in mind these two important problems. In section 6.2, we illustrate what

we consider its major motivation, namely, context dependence in argument

realization. We point out that thematic hierarchies have been used to account for

very different kinds of generalizations over a variety of domains, although they

all involve context dependence in an abstract sense. In section 6.3 we review a

sample of the thematic hierarchies proposed in the literature, highlighting the

major contrasts among them. In section 6.4 we consider hypotheses concerning

what these thematic hierarchies are meant to represent, what phenomena

hierarchies are meant to account for, and how they have been considered to

be derived from more basic components of lexical semantic representations.

In section 6.5, we review possible sources for the various formulations of

the thematic hierarchy. First, we set aside the generalizations which appear to

implicate a thematic hierarchy but are more likely attributable to other semantic

factors. Then, we show that the thematic hierarchy is put to use in several ways,

giving rise to various formulations; specifically, in a given instantiation, the set

of semantic roles and their relative rankings is determined by the interpretation

given to the hierarchy and by the types of generalizations being captured.

Since the thematic hierarchy has played a key part in the explanation of

numerous complex linguistic phenomena, we cannot survey all its instantiations,

let alone identify the interactions between it and various other linguistic notions

that have been implicated in the explanations of these phenomena. We hope

that the exposition of the various conceptions of the thematic hierarchy and

the discussion of factors that can complicate the comparison of particular

hierarchies will assist future researchers in making progress in the analysis of

specific phenomena.

6.2 The primary motivation for hierarchies: context


As already mentioned, thematic hierarchies allow arguments to be referred to

in terms of their relative ranking. As Baker (1997: 109) and Speas (1990: 73)

note, this property makes them a particularly useful device for dealing with

what we called the "context dependence" pervasive in argument realization:

the options for the syntactic realization of a particular argument are often not

determined solely by its semantic role, but also by the semantic roles borne

by its coarguments. In this section we present several argument realization

generalizations which illustrate context dependence. In the following sections

we demonstrate that these generalizations are not all of the same kind, leading

us to propose that there is no single universal ranking of semantic roles.

Consider first the subject selection paradigm in (4), which was the motivation

for Fillmore's subject selection rule (1), together with the further examples

that fill out this paradigm in (5).

(4) a. The door opened.

b. John opened the door.

c. The wind opened the door.

d. John opened the door with a chisel. (Fillmore 1968: 27, (40)-(43))

(5) a. âˆ-The door opened with the wind.

b. âˆ-The door opened by John.

c. âˆ-The chisel opened the door by John.

No single semantic role is consistently associated with the grammatical relation

subject, and, conversely, no semantic role consistently shows a particular

syntactic realization. Yet, whether an argument bearing a particular semantic

role can be realized as subject depends on the semantic roles of the other

arguments in the sentence, as the contrast between (4) and (5) makes clear. For

instance, a patient cannot be a subject in the presence of an agent, instrument,

or natural force. This pattern follows from the subject selection rule (2), in

conjunction with an appropriate thematic hierarchy, which ranks patient below

agent, instrument, or natural force. The syntactic expression of the patient

depends on this role's position in the hierarchy relative to the roles of other

arguments in the same clause. The patient can only be the subject when it is

the sole argument of the verb (see (5)) or when any other arguments bear roles

lower in the hierarchy, such as location, as in The cat sat on the windowsill.

This account extends to instruments, which also show context dependence: they

can only be subjects in the absence of an agent. Similarly, Speas (1990: 73)

points out that a recipient - the animate goal in an event of physical or abstract

transfer of possession - may be a subject, but only in the absence of an agent,

citing (6).

(6) a. John received a package from Baraboo.

b. Mary sent a package to John from Baraboo. (Speas 1990: 73, (102))

This asymmetry follows if agent is higher than recipient in the thematic hierarchy,

and the argument bearing the highest role is selected as subject. It is

usually assumed that the precise statement of the thematic hierarchy will emerge

from an examination of the interrelations in the realization of all semantic


Context dependence is also found in the realization of arguments bearing

the experiencer role. As discussed in sections 1.2 and 1.5, experiencers can

be expressed as subjects of psych-verbs such as fear (e.g., The toddler feared

the lion) or objects of psych-verbs such as frighten (e.g., The lion frightened

the toddler). The realization of the experiencer depends on the semantic role

of the nonexperiencer argument of the verb, which, as argued in section 1.2,

differs for the two types of verbs. With frighten verbs, the semantic role of

the nonexperiencer argument could be characterized as an agent, instigator,

effector, or cause (Grimshaw 1990; Pesetsky 1995). In formulations of the

hierarchy that include an experiencer (Belletti and Rizzi 1988; Bresnan and

Kanerva 1989; Grimshaw 1990; Speas 1990), the agent-like roles such as agent,

instigator, effector, or cause are ranked the highest and thus have priority over

the experiencer for expression as subject. In contrast, with the fear verbs, the

nonexperiencer argument cannot be analyzed as bearing an agent-like role, but

is better analyzed as bearing the stimulus (or theme) role. If experiencer is

ranked above stimulus in the hierarchy (Belletti and Rizzi 1988; Grimshaw

1990; Van Valin 1990), fear verbs will have an experiencer subject.2

A somewhat different instantiation of context dependence is reflected in

Polinsky and Kozinsky's (1992) appeal to a thematic hierarchy in the analysis

of the applicative construction in the Bantu language, Kinyarwanda. In this

construction, an NP, normally expressed as an oblique, is expressed as an

object, and the verb also has an "applicative" morpheme affixed to it; this

morpheme often identifies the semantic role of the added object. Kinyarwanda

allows NPs bearing a wide range of semantic roles to be introduced as objects

by the applicative morpheme; these include recipients, benefactives, possessors,

instruments, and comitatives. When a sentence has NPs bearing more than one

of these roles, which of these NPs can be expressed as an object, rather than

as an oblique, is dependent on the semantic roles of the other NPs within the

sentence. For instance, when a benefactive and an instrument cooccur, only

the benefactive - and not the instrument - can be expressed as an object.

Polinsky and Kozinsky point out that the asymmetries can be captured using a

thematic hierarchy. The just-described asymmetry would follow if benefactives

are ranked above instruments, and if accessibility to objecthood is determined

by a thematic hierarchy, with the NP in the pair having the higher-ranked role

being the object.

The thematic hierarchy is also used to explain properties of the applicative

construction in another Bantu language, Chichewˆ a. In Chichewˆ a the choice

of added semantic role is apparently constrained by the roles of the original

arguments of the base verb (Alsina 1999: 26-27; Bresnan and Kanerva 1992:

117-18). A verb whose only argument bears the patient/theme role cannot

be found with a benefactive introduced by the applicative morpheme, while

a verb whose only argument bears the agent role can be, as can a verb that

takes both agent and patient/theme arguments. Alsina (1999: 26) and Bresnan

and Kanerva (1992: 117) propose that the applicative morpheme can only

introduce an argument whose semantic role is lower on the thematic hierarchy

than that of one of the verb's original arguments. The distribution of

applicative objects can then be captured with a thematic hierarchy in which

benefactive is lower than agent, but higher than patient/theme. Furthermore,

Alsina (1999: 26-27) and Bresnan and Kanerva (1992: 118) place location

lower than patient/theme based in part on the fact that a verb whose only

argument bears the patient/theme role can be found with a location applicative


Context dependence also motivates reference to the thematic hierarchy in

the statement of crosslinguistic generalizations about argument realization.3

This use is most explicitly and thoroughly explored by Dik (1978, 1980,

1997a, 1997b) within Functional Grammar, but a comparable use is found in

GivÙŽn's work (1984a, 1984b, 1990, 2001). "Deep" ergative languages aside,

all languages allow a verb's agent argument to be expressed as subject, but

differ as to whether they allow arguments bearing other semantic roles to be

expressed as subject in the absence of an agent argument (see section 1.6).

Although the set of semantic roles expressible as subject varies from language

to language, Dik (1978) proposes that the range of variation is constrained

and takes a form that can best be described through reference to a thematic

hierarchy; see GivÙŽn (1984b) for a similar proposal. Languages choose the

semantic roles that can be realized as subject or object from a continuous

portion of the hierarchy starting with the agent for the subject and with the

second-highest role for the object and working downwards, as stated in Dik's

Continuity Hypothesis in (7).

(7) For any language, if Subj or Obj function can be assigned to some semantic

function Sj, then Subj or Obj can be assigned to any semantic function Si, such

that Si precedes Sj in SFH [=Semantic Function Hierarchy, i.e., the thematic

hierarchy] (for Obj assignment, Si

_= Ag). (Dik 1978: 76)

Crosslinguistic differences are restricted to choice of cut-off points for subjecthood

and objecthood on the hierarchy. The result is that, "we can specify the

possibilities of Subj and Obj assignment for a language simply by giving the

last semantic function in the SFH to which Subj and Obj can be assigned" (Dik

1978: 75). Assuming the thematic hierarchy in (8), all languages have agent

subjects, some might have agent and patient subjects, others agent, patient and

recipient subjects, and so on. None, however, would have patient, benefactive,

and instrument subjects only.4

(8) Agt > Pat > Rec > Ben > Inst > Loc > Temp (Dik 1978: 70, (3))

Dik presents evidence from a range of languages to illustrate that all the possible

cut-off points are instantiated. This hypothesis, then, represents an implicational

generalization of the type familiar from work on language universals.

Furthermore, the same hierarchy, with the agent removed, figures in crosslinguistically

valid implicational generalizations about object selection (Dik 1978:

76, 1997: 246, (41); GivÙŽn 1984a: 163) and passivization (GivÙŽn 1990: 566).

For example, according to (8) a language would only allow benefactive objects,

if it allowed patients and recipients also to be objects.

These typological generalizations involve a somewhat different instantiation

of context dependence than the examples discussed earlier. The other

instances involve priorities among the coarguments of a verb with respect

to their potential realization as subject; that is, they involve priorities relevant

to the mapping algorithm. In contrast, Dik's and GivÙŽn's generalizations

apply across verb types and across languages. They define a set of possible

languages with respect to the options for the realization of different semantic

roles as subject and object. For example, while Dik's generalization states

that a language will only allow recipients as subjects, if it also allows agents

and patients as subjects, the other type of generalization involves whether a

recipient may be realized as the subject of a particular verb in a language

which allows such subjects; this possibility is determined in the context of

the coarguments of that verb. The ranking of semantic roles used in the statement

of typological generalizations, then, may turn out not to be the same as

the one that might be used by a mapping algorithm. In sections 6.4 and 6.5,

we show that recognizing that different kinds of generalizations capture context

dependence over different domains is critical for understanding why not

all proposed thematic hierarchies are alike. First, we review the various formulations

of the thematic hierarchy, as a prelude to trying to uncover their


6.3 Formulations of the thematic hierarchy

In this section, we lay out representative formulations of the thematic hierarchy;

additional hierarchies are listed in Newmeyer (2002: 65-67, (25)). This

comparative presentation, which is inspired by those in Baker (1996a: 7-8, (2))

and Macfarland (1991: 105), is intended to highlight the most important differences

among the hierarchies.5 As noted by Baker and Macfarland, much of the

controversy centers around the placement of the patient/theme role with respect

to other roles, particularly the spatial roles such as goal and location. For this

reason, the hierarchies have been grouped according to their treatment of the

patient/theme role, which is italicized for ease of identification. In section 6.5

we discuss the sources of the controversy over the ranking of these roles. The

following abbreviations are used in the hierarchies: Act - actor, Ag - agent,

Be - benefactive, Da - dative, Eff - effector, Ex - experiencer, In - instrument,

G - goal, L - location, P - path, Pa - patient, Re - recipient, S - source,

Th - theme. Certain oblique roles have been omitted from the bottom of Dik's,

Fillmore's (1971a), GivÙŽn's, Larson's, and Speas' hierarchies; Dik's "goal"

role (see note 4) and Fillmore's "objective" role have been relabelled "patient"

to conform to more common usage; Bresnan and Kanerva's hierarchy uses the

label "recipient," which can be understood as a type of goal and is treated as

such in the comparative presentation; see below for further discussion. Some

hierarchies do not rank all roles with respect to each other; roles that are ranked

together - and, thus, are unranked with respect to each other - are separated

by a slash.6

No mention of goal and location:

Belletti & Rizzi 1988: Ag > Ex > Th

Fillmore 1968: Ag > In> Pa

Goal and location ranked above patient/theme:

Grimshaw 1990: Ag > Ex > G/S/L > Th

Jackendoff 1972: Ag > G/S/L > Th

Van Valin 1990: Ag > Eff > Ex > L > Th > Pa

Goal and location ranked below patient/theme:

Baker 1989: Ag > In> Th/Pa > G/L

Baker 1997: Ag > Th/Pa > G/S/L

Carrier-Duncan 1985: Ag > Th > G/S/L

Dik 1978: Ag > Pa > Re > Be > In

Fillmore 1971a: Ag > Ex > In> Pa > G/S/L

Jackendoff 1990b: Act > Pa/Be > Th > G/S/L

Larson 1988: Ag > Th > G

Speas 1990: Ag > Ex > Th > G/S/L

Goal ranked above and location below patient/theme:

Bresnan & Kanerva 1989: Ag > Be > Re/Ex > In> Th/Pa > L

Kiparsky 1985: Ag > S > G > In> Th/Pa > L

GivÙŽn 1984b: Ag > Da/Be > Pa > L > In

Besides the varying rankings of the goal, source, and location roles with

respect to the patient/theme role, a few other differences are worth noting.

First, some hierarchies include semantic roles associated only with arguments,

but others include semantic roles more likely to be associated with adjuncts.

Second, hierarchies differ in the roles which are ranked. A particular hierarchy

may contain a small number of roles, simply reflecting the set of roles relevant

to the analysis of a particular phenomenon, as in Belletti and Rizzi's (1988)

hierarchy, which is motivated by a study of psych-verbs. Differences in roles

ranked may sometimes be linked to the granularity of the analysis: some

hierarchies are based on a rather fine-grained semantic analysis which leads to

the inclusion of a number of roles, while other hierarchies contain fewer roles,

often reflecting a coarser-grained semantic analysis. For example, Bresnan and

Kanerva (1989) distinguish a recipient role from a purely spatial goal role, and

lump all the spatial locative roles, including the spatial goal, under a location

role; others consider the recipient role simply to be a subtype of goal and do

not single it out, while still others distinguish between the spatial goal role

and other spatial roles. Van Valin (1990) makes a distinction between an agent

role, in the sense of an animate, volitional instigator, and an effector role -

the role of a participant that brings something about (Van Valin and LaPolla

1997: 118; Van Valin and D. Wilkins 1996), ranking the former higher than

the latter, while still others do not make this distinction. Van Valin (1990; Van

Valin and LaPolla 1997: 126-27) also posits distinct theme and patient roles,

ranking theme above patient. Others do not make this distinction; they either

include only one of these two roles or rank them together. Some hierarchies

specifically do not rank noncooccurring argument types, while others do.

The proliferation of thematic hierarchies has provoked two reactions.

Newmeyer takes the multiple formulations as evidence that the construct is

simply not useful: "There is reason for strong doubt that there exists a Thematic

Hierarchy provided by UG. That seems to be the best explanation for the fact

that after over three decades of investigation, nobody has proposed a hierarchy

of theta-roles that comes close to working" (2002: 65). Other linguists

apparently assume that there is a universal ranking of semantic roles and that

it should be possible to resolve the conflicting role rankings reflected in the

various statements of the hierarchy in favor of a single ranking once the supporting

data is more carefully scrutinized (e.g., the debate between Bresnan

and Kanerva [1989, 1992] and Schachter [1992]).

Deciding whether it is possible to identify a thematic hierarchy that "comes

close to working" requires determining which generalizations a thematic hierarchy should capture. It does not make sense to choose between two hierarchies which are meant to account for different sets of phenomena. As already

reviewed, a heterogeneous collection of generalizations can be validly captured

using a ranking of semantic roles, and there is no reason to expect to find a

single ranking of roles which works for all of them.

Distinct facets of argument realization are sensitive to different types of semantic notions, which may be

defined at various levels of granularity. In order to understand why different

hierarchies rank roles differently, it is necessary to realize that thematic hierarchies introduced to capture particular generalizations are derivative constructs: they are convenient devices that allow valid linguistic generalizations to be captured, but ultimately they derive from the more basic components of the lexical semantics of verbs and the properties of the NPs which serve as their arguments. A lack of consensus about "the" thematic hierarchy is not, on its

own, a reason to reject any particular hierarchy; rather, it is a reason to probe

the basis for any given ranking of semantic roles. There is much to be gained,

then, from an attempt to understand the various rankings of semantic roles, as

they often do express valid generalizations.

6.4 The grounding of the thematic hierarchy

Since semantic roles are now largely viewed as derivative constructs defined

over the more basic components of the lexical semantic representations of verbs,

the semantic roles constituting the thematic hierarchy and their rankings should

also derive from these components. In chapter 3 we reviewed two ways of

defining semantic roles: in terms of positions in an elaborated lexical semantic

representation and in terms of clusters of lexical entailments associated with

arguments. As we show, each way of defining semantic roles is naturally

associated with a distinct conception of the thematic hierarchy. On either

conception the hierarchy is derivative.

The two interpretations of the thematic hierarchy are characterized by

distinct interpretations of "prominence," as hinted by Bresnan and Kanerva

(1989: 23-24). On the first, prominence is defined "structurally" over the lexical semantic representation of a verb (Baker 1997; Croft 1998; Jackendoff

1990b; Kiparsky 1985). This conception fits well with the view of semantic

roles as labels for positions in an articulated lexical semantic representation,

such as a predicate decomposition or an event structure. The thematic hierarchy

is viewed as a statement of a generalization, but not as an independent construct

(Baker 1996a, 1997; Croft 1991: 186, 1998; Kiparsky 1985; Wunderlich

1997a, 1997b). On the second interpretation, prominence is taken to be akin to

cognitive salience, and the thematic hierarchy is taken to be one of a number

of "natural prominence scales," such as the person, animacy, number, and definiteness hierarchies. The other scales, however, each provide a ranking of the

possible values of a particular attribute of an argument, while the thematic hierarchy does not rank an argument according to the values of a single attribute.

A range of semantic properties determine the salience of an argument, so that

the thematic hierarchy is better viewed as "the cumulative result of a number of

interacting relative prominence relations between semantic entities," to quote

T. Mohanan (1994: 28). There is, then, a natural affinity between a thematic

hierarchy incorporating this interpretation of prominence and the conception

of semantic roles as clusters of properties of arguments. Although some of the

properties which determine the salience of an argument involve its semantic

role, many, in fact, are properties of the NPs which typically fill this semantic

role. They should not, strictly speaking, be considered defining properties of

the semantic role; however, since the properties of the role fillers may align in

particular ways with the associated semantic roles, they may also give rise to

what appear to be thematic hierarchy effects.

6.4.1 Structurally based thematic hierarchies

For a thematic hierarchy derived from a structured representation of verb

meaning, prominence is defined over this representation. When one semantic

role is ranked higher than a second, the first is assumed to be structurally more

prominent than the second. Some researchers pair structural prominence with

the interpretation of semantic roles as labels for positions in a more elaborated

lexical semantic representation, such as an event structure. Prominence, then,

is defined in terms of the geometry of the representation, as discussed in

section 4.2.5. Other researchers take structural prominence to be defined by

the order of composition of arguments with their verb. This order presumably

reflects a verb's lexical semantic representation, though proponents of this

viewpoint usually do not posit an explicit lexical semantic representation to

undergird the posited orders.

When prominence is defined over an event structure, it is taken to be

inversely correlated with depth of embedding in the event structure, or, as

Kiparsky (1997: 484) puts it, "the order of Th-roles is a reflection of their

semantic depth," an idea that he attributes to Bierwisch and that is further

refined in related work by Wunderlich (1997a: 102, 1997b: 44). Prominence

relations defined in this way can as a matter of convenience be expressed by a

thematic hierarchy.

The second structural interpretation of prominence takes the thematic hierarchy

to be a reflection of the order in which a verb composes with its arguments:

the argument that composes first is lowest in the hierarchy, the argument

that composes next is the next lowest, and so on. This view is adopted by

Kiparsky (1985, 1997) and shared by Larson (1988: 382-83). Kiparsky (1997)

and Wunderlich (1997a, 1997b) bring these two positions together: they take

semantic roles to represent positions in an event structure and they propose that

order of composition of arguments can be determined from the structure of this

representation. Order of composition, then, reduces roughly to depth of embedding in a semantic representation. In fact, there is much to recommend this move. As Kratzer (1996: 115-16) argues, from a formal semantic viewpoint, differences in the order of composition of the arguments of a multiargument verb (or predicate) have no semantic consequences; order of argument composition is semantically meaningful only if the relevant arguments are arguments of distinct predicates in an event structure. Therefore, order of composition can provide independent motivation for the statement of the thematic hierarchy, only if order of composition is equated with depth of embedding in a semantic representation.

The structural conception of the hierarchy is usually coupled with a prominence

preservation approach to mapping (see section 5.2). Such a mapping

ensures that the semantic prominence relations among arguments whether

defined by depth of embedding in an event structure or by order of composition

are preserved in the corresponding syntactic structure. For example, in

Principle and Parameters terms, if one position in an event structure is more

prominent than another, then the argument in the syntactic d-structure associated with the first position asymmetrically c-commands the argument in the syntax associated with the second.

The geometry of the lexical semantic representation is claimed to be at the

root of certain generalizations regarding argument realization. For example, in

languages with noun incorporation, dative verbs - verbs with agent, theme,

and recipient arguments - can incorporate the theme, but not the recipient

(Baker 1988: 453-54, n. 13, 1996b: 292-93). Kiparsky (1997: 484) argues that

this asymmetry reflects the structure of the lexical semantic representation:

the theme is closer to the verb than the recipient in this representation, and,

therefore, the verb composes with it first. He captures this generalization by

appealing to a thematic hierarchy and proposing that the argument of a verb

whose semantic role is lowest on the hierarchy has precedence for noun incorporation.

This explanation is extended to asymmetries in the structure of idioms.

English dative verbs are found in a multitude of idioms with fixed themes and

variable recipients (e.g., lend x an ear, promise x the moon, read x the riot act,

show x the ropes), but not in idioms with fixed recipients and variable themes.

Explanations of this and other idiom asymmetries take as their starting point

the assumption that verbs form idioms with items they compose with first, a

property that would be reflected in these items bearing the lowest-ranked of the

semantic roles associated with the verb (Kiparsky 1985; Larson 1988). If this

assumption is correct, the idiom data provide evidence that recipients are ranked

above themes. Idiom asymmetries have figured prominently in discussions of

the hierarchy, particularly with respect to the ranking of recipients and themes,

and we consider this type of evidence and its validity further in section 6.5.3

in the context of a more detailed examination of the ranking of these two roles.

Although a structurally based hierarchy is defined with respect to a verb

and its arguments, it can be broadened to include semantic roles of what are

typically considered syntactic adjuncts, if these "adjunct" roles can be defined

over an "augmented" event structure associated, say, with an applicative verb.

Marantz (1993: 143-44), for example, gives applicative verbs complex event

structures, distinguishing an "inner" - or embedded event - and an "outer"

event, with instruments included in the inner and benefactives in the outer

event. The idea is that the verb composes first with the arguments belonging to

the inner event and only then with those belonging to the outer event, giving

rise to thematic hierarchy effects involving adjuncts, as we discuss further


On the structural conception, the thematic hierarchy is epiphenomenal,

algorithmically derived from an articulated lexical semantic representation or,

alternatively, from whatever determines order of composition of arguments

with their verb. There is no single independently formulated thematic hierarchy

which exhaustively ranks all semantic roles. Rather, each event type gives rise to its own structurally defined hierarchy composed of the semantic roles associated with that event type; a hierarchy, then, cannot rank noncooccurring arguments, such as agent and effector or experiencer and recipient. The semantic role which is most prominent for a given event type shares with the most prominent semantic roles for other event types only the property of representing the argument that is structurally highest among the coarguments. Constraints on

the possible realizations of arguments bearing particular semantic roles derive

from the depth of embedding of the positions in the event structures which

define the relevant roles; the possible realizations are derived from the pool of

possible event structures.

On this understanding of the structurally based thematic hierarchy all constraints

on a verb's argument realization options should arise from differences

in the relative depth of embedding of its arguments in its event structure.

However, some constraints on a verb's argument realization options cannot be

explained in this way since depth of embedding does not always impose an

exhaustive ranking on all arguments of a verb. Often, particular predicates in

an event structure have more than one argument. For example, two-argument

stative verbs, such as have, love, see, or want, have a single predicate in their

event structure, with two arguments associated with it; the same holds of two argument activity verbs, such as rub, wipe, push, and study (e.g., B. Levin 1999;

Rappaport Hovav and B. Levin 1998a; Van Valin and LaPolla 1997). (There

are "technical" ways around this limitation; for instance, extra predicates could

be introduced, but their introduction and the ranking they impose on arguments

can be hard to justify.) As mentioned, Kratzer (1996) points out that the order

of composition of arguments of a single predicate cannot be independently

established, except through their grammatical realization, which is what order

of composition is supposed to explain to begin with. Consequently, depth of

embedding does not prevent a verb like want from expressing its experiencer

as object or a verb like push from expressing its theme as subject; however,

such mappings are unattested. Although the thematic hierarchy is supposed

to preclude them, it is unable to do so, if it is determined by the depth of

embedding of arguments. Thus, such two-argument verbs are problematic.

In order to impose a ranking on the arguments of all verbs, proponents of

structurally based thematic hierarchies sometimes postulate event embedding

within a verb's predicate decomposition, even where it is not clearly justified.

Baker (1997: 123-24, 2003: 79-83), for example, postulates event embedding

in the form of a causative event structure for all transitive verbs to explain

observed prominence relations common to the arguments of such verbs. Yet,

this assumption has never been seriously entertained by lexical semanticists,

and it is explicitly argued against in B. Levin (1999) and Van Valin and LaPolla

(1997: 101-02).

One way of distinguishing among arguments in an event structure is to

take advantage of a distinction among event structure participants introduced

in Rappaport Hovav and B. Levin (1998a) and elaborated in B. Levin (1999),

building on an idea in Grimshaw (1993). These papers propose that roots (see

section 3.2) are associated with some number of participants and that when

roots are integrated into event structures, some or all of these participants are

associated with argument positions in the event structures. With change-ofstate

verbs, the root is associated with a cause and an entity that changes state,

and in the event structure the first of these participants is associated with the

argument of the causing subevent and the second with the argument of the

result subevent, respectively. The roots associated with two-argument activity

or stative verbs are also associated with two participants, but only one of these

root participants is associated with an argument position in event structure. For

instance, the surface-contact verb wipe has a root associated with an agent and

a surface, but its activity event structure only specifies an agent. Its other root

participant is only licensed by the root itself. The result is an asymmetry in the

status of two arguments within the same level of event structure embedding,

which might be reflected in prominence relations in the syntactic realization

of the arguments. This approach to ranking, however, is only viable if "pure"

root participants can be identified in a principled way, but this remains to be

worked out.

Croft (1991: 186, 1998) offers an alternative account of how the thematic

hierarchy derives from the internal structure of events, which provides an

independent criterion for ranking all the arguments of a verb. For him, the

internal structure of an event is defined by the causal chain that makes explicit

force-dynamic relationships among the event participants (see section 4.3). The

position of arguments in the causal chain serves to impose a ranking on the

arguments, in which prominence is defined by force-dynamic antecedence and

is the source of thematic hierarchy effects. For Croft, then, only force-dynamic

antecedence, and not depth of embedding, is the relevant structural notion,

and, furthermore, there is an independently defined notion of antecedence for

coarguments of single predicates. Langacker (1990: 238) also proposes that

"energy flow," as he calls the force-dynamic relationships along a causal chain,

gives rise to a thematic hierarchy. In a similar vein, Siewierska (1991) considers

the thematic hierarchy as grounded in what DeLancey (1981) calls "natural

attention flow," which she writes "refers to the actual development of events

in the real world, the basis for the perception of naturalness being temporal

order" (1991: 105). These views may be related to Primus' proposal (1999:

48-51) that the asymmetries among event participants embodied by the paired

entailments in Dowty's lists of proto-role properties (see section 3.1.1) are

instantiations of a general notion of control. This notion could be interpreted as

a generalization over various force-dynamic relations and might provide a way

to impose ordering where depth of event embedding fails. In what follows, we

make reference primarily to causal order, leaving open whether it subsumes

these other notions.

Since causal order only defines relations of precedence, and not hierarchical

relations, approaches which refer to causal order are typically accompanied by

approaches to argument realization which also do not appeal to hierarchical

structure. This correlation is not necessary. Marantz (1993: 144-45) also shares

this general view of the origins of the thematic hierarchy, though he instantiates

the idea in terms of event embedding. He suggests that the instrument and

the patient/theme belong to an event embedded inside another event, which

includes the benefactive. As the instrument figures among Croft's antecedent

roles, while the benefactive is one of Croft's subsequent roles (see section 4.3),

the hierarchical relations imposed by event embedding mirror causal order.

The question, then, is whether these two ways of understanding structural

prominence - in terms of depth of embedding or order of composition and

in terms of causal order - are related. Prominence relations among arguments

established by depth of embedding in event structure are indeed preserved in the

syntax, and it appears that they take precedence over other semantic properties

implicated in argument realization. For instance, as shown in section 3.1.1, the

inanimate cause arguments of morphological causatives are always realized as

subjects, even in the context of animate causees. The causees are presumably

arguments of an event embedded under the causing event that includes the

cause and, thus, will have a less prominent realization than the cause. But

embedding relations often reduce to causal order, since in a causative event

the least embedded argument, the cause, also force-dynamically antecedes the

other arguments. Can causal order, then, replace embedding relations as the

criterion used to rank arguments?

Causal order sometimes does impose a ranking where depth of embedding

fails. For example, the activity event structure of the surface-contact verb

wipe contains a single predicate, and its arguments, as coarguments of this

predicate, are not distinguished by depth of embedding. Its agent, however,

force-dynamically antecedes the argument representing the surface, consistent

with its realization as subject. But even causal order has apparent limitations.

As noted in section 3.1.1, two-argument verbs with one necessarily sentient

argument, such as fear, love, see, and want, express this argument as subject,

suggesting it is more highly ranked than the nonsentient argument; however,

it can be difficult to independently motivate a particular causal order for the

arguments of all such verbs (see section 4.3). Neither causal order, nor depth

of embedding, then, is adequate for ranking all arguments of a verb for subject

or object selection. Yet, as shown in chapter 3, a mere counting of entailments

as in Dowty's (1991) proto-role approach does not give the right results either,

since the causation entailment overrides all other entailments in determining

subject choice. Perhaps depth of embedding is always respected in the mapping

to syntax, as expected on the prominence preservation approach to mapping.

When depth of embedding does not impose an exhaustive ranking on a verb's

arguments, causal order is appealed to, if applicable. When causal order also

fails to impose a complete ranking, other semantic properties of the verb's

arguments may determine their relative ranking. A thematic hierarchy built

around entailments associated with semantic roles and their fillers is the topic

of the next section.

6.4.2 The thematic hierarchy as a natural prominence scale

As Bresnan and Kanerva (1989: 23-24) point out, some researchers interpret

the thematic hierarchy as a topicality hierarchy; arguments bearing semantic

roles higher on the hierarchy are said to be more likely to be topics than

those bearing roles lower on the hierarchy (Fillmore 1977b; GivÙŽn 1984b;

Hawkinson and Hyman 1974; Trithart 1979). On this conception, the thematic

hierarchy is one of a family of "natural prominence scales." Such scales rank

the possible values of some attribute of a particular linguistic unit - semantic,

conceptual, pragmatic, or morphological - according to an intuitive notion

of prominence, perhaps, characterizable as cognitive salience. As McCarthy

writes, "Natural language is replete with natural scales, with one end more

prominent, in the abstract sense, than the other" (2002: 21). Other natural

prominence scales include the animacy hierarchy, the person hierarchy, and the

definiteness hierarchy; some also include the grammatical relations hierarchy

and the morphological case hierarchy among them, though these seem different

in nature. The ranking represented by each scale essentially prioritizes the

possible values of an attribute for a range of linguistic phenomena, including

facets of argument realization. For this reason, Dik (1997a: 36) calls them

"priority hierarchies."

A given NP in a sentence usually has various attributes and, thus, is associated

with values drawn from several natural prominence scales. The values

of the various attributes associated with a given NP tend to be chosen from

the same ends of the relevant scales. So the agent role is ranked high on the

thematic hierarchy, and agents tend to be human, a value high on the animacy

hierarchy; the agent role is also usually associated with subject, which is high

on the grammatical relations hierarchy, and with nominative case, which is

high on the morphological case hierarchy. Consequently, certain clusters of values

are considered unmarked combinations, and, therefore, natural prominence

scales are central to discussions of markedness. As discussed in section 1.5,

agent-patient verbs are considered the prototypical transitive verbs crosslinguistically,

expressing their agent and patient arguments as subject and object,

respectively. Divergences from this pairing of semantic roles with grammatical

relations typically involve explicit morphosyntactic marking. For instance,

when the patient argument of an agent-patient verb is realized as a subject,

passive morphology is found on the verb in many languages; the marked subject

choice is reflected in the morphologically marked passive verb form. These

effects can be derived using a thematic hierarchy along with the assumptions

that agent is ranked highest on a thematic hierarchy, subject is ranked highest

on a grammatical relations hierarchy, and the mapping from semantic roles to

grammatical relations preserves prominence relations, that is, it shows what

is called harmonic alignment in Optimality Theory (Aissen 1999; McCarthy

2002; A. Prince and Smolensky 1993: 136; see also section 5.3.2).

Languages vary, however, as to how tight an alignment they demand. Some

allow only the very highest values in the various natural prominence scales to

be aligned, while others allow the highest value on one scale to be aligned with

a range of values on a second scale. Each language, then, defines cut-off points

for the purposes of aligning values on the prominence scales, which results in

the implicational generalizations discussed in section 6.2 concerning the range

of semantic roles with access to subjecthood and objecthood in a particular language.

Dik (1978: 76-78, 1997a: 265-69) provides a comprehensive discussion

of the various types of argument realization generalizations that follow from

such uses of the thematic hierarchy both within and across languages (see also

GivÙŽn 1984a: 163). Surveying a number of languages, Dik illustrates that languages

prefer NPs bearing the semantic roles higher in the hierarchy to be

encoded as subject or else object. Usually when a lower-ranked role is chosen

as subject or object, the choice is accompanied by special verb morphology,

such as passive affixes or the applicative affixes of Bantu languages that appear

when benefactive, instrument, or locative NPs are objects. As Dik (1997a: 266)

writes, "as we proceed through the SFH [=Semantic Function Hierarchy, i.e.,

the thematic hierarchy] . . . Subj and Obj assignment become more and more

'difficult', and the resulting constructions become more and more 'marked'."

Although all languages align the more highly ranked semantic roles with

subject, they encode grammatical relations in various ways, most commonly

via word order, morphological case, and verb agreement, but also via specialized

voice morphology (e.g., active vs. passive voice, direct vs. inverse

marking, applicative morphemes; see section 1.6). Other natural prominence

scales, such as the animacy and definiteness hierarchies, are implicated

in this surface expression of arguments; Aissen (1999: 677), K. Allen

(1987: 51), Artstein (1998: 1), Croft (2003b: 128-32), Dik (1997a: 37), and

Siewierska (1988: 29-103, 1993: 831) present fairly comprehensive, though

nonidentical, lists. Languages vary as to which scales, if any, the surface

expression of arguments is sensitive to. For example, many languages show

"differential object marking" (Aissen 2003; Bossong 1991, 1998): an animate

object receives a distinct morphosyntactic expression from an inanimate object,

and, typically, one that is more morphosyntactically marked. This observation

is interpreted in terms of the markedness of the alignment of values on different

scales: an animate patient is a marked choice for direct object of an

agent-patient verb since the prototypical patient is inanimate. As Aissen (2003)

shows, the relevant generalization is best restated in terms of a natural prominence

scale because animacy is not a binary-valued attribute; rather, there are

multiple values for animacy that can be arranged in an animacy hierarchy, ranging

from most to least "animate" (K. Allen 1987; Comrie 1989: 185; Corbett

2000: 55-66; Croft 2003b: 128-32; Silverstein 1976: 116-22; Van Valin and

D. Wilkins 1997: 313-17), and languages with differential object marking can

choose larger or smaller segments of the animacy hierarchy as the domain of

differential object marking.

Often, a marked expression with respect to the thematic hierarchy allows a

less marked expression with respect to another prominence scale. For example,

some languages such as Lummi (Jelinek and Demers 1983, 1994) require

that the NP bearing the patient role be the subject of a passive verb if its

referent is higher on the person hierarchy than the referent of the agent NP; this

requirement allows a preference for first- and second-person subjects to be met

at the expense of a marked association of patient with subject. Thus, a more

marked expression with respect to one attribute allows a less marked expression

with respect to another (Artstein 1998: 1; Evans 1997: 420-21), a situation

which suggests, in turn, that there is a precedence ranking among the natural

prominence scales. Recently, these ideas have been pursued within Optimality

Theory, which includes mechanisms for stating the relevant generalizations

and for delineating the range of crosslinguistic variation in the surface coding

of arguments (Aissen 1999, 2003; Artstein 1998; Asudeh 2001; Lee 2003;

Legendre, Raymond, and Smolensky 1993).

What property of semantic roles organizes them into a natural prominence

scale? And, if there is such a property, is it involved in organizing other natural

prominence scales? As already mentioned, an intuitive notion of cognitive

salience has been said to underlie all these scales. For the scales relevant to

argument realization, Newmeyer writes, "Human NPs are more central to the

human experience than inanimates; agents are more central than instruments;

subjects are more central than objects of prepositions; and so on" (2002: 50).

More generally, the ranking of the various values of the attributes which

define each natural prominence scale is said to reflect the contribution of

the attribute to the topicality of the argument bearing this attribute (Artstein

1998; Hawkinson and Hyman 1974; Jelinek and Carnie 2003; Trithart 1979).

This perspective is reminiscent of GivÙŽn's (1984b: 134) proposal that the

thematic hierarchy ranks semantic roles based on their topicality. Inherent

topicality, however, is not a property of a semantic role, even if it might be

indirectly associated with a role due to the potential t