(Bresnan, 2001:46)
3.
Similarities and differences between LFG, P&P and HPSG
Lexical functional Grammar (LFG) is a syntactic framework which aims to provide a computationally precise and psychologically realistic representation of language (Sells, 1985: 135). In pursuing this goal, LFG uses a model with multiple levels of representation, each with their own architecture, vocabulary and constraints. These levels are not derived from each other, as they are in approaches such as P&P[1]. Instead, they are parallel and linked through mappings constrained by principles of correspondence, which ensure that the information in different levels is compatible (Austin, 2001: 2). Thus, LFG is constraint-based and non-derivational. Another characteristic of LFG is that, unlike P&P, it uses no movement (Nordlinger, 1998: 10). Instead, it resorts to different accounts of the phenomena (such as the long distance dependencies, described in section 2.3.3 below) that are explained by movement in P&P. LFG also differs from P&P in its assumption that notions such as subject and object are primitives. This assumption and the presence of multiple levels of representation are related to the view held in LFG that syntax is not based only on structure (as put by Sells (1985: 137), ‘there is more to syntax than you can express with phrase structure trees’). The fact that the level responsible for phrasal-structure (namely, constituent-structure, discussed below) is only one of many levels reflects this view (for more details about the parallel levels refer to section 2).
LFG began to be developed by Joan Bresnan and Ronald Kaplan in the early 1970’s because they were dissatisfied with the approaches available at the time (Austin, 2001: 2). Kaplan and Bresnan (1982a) disputed the Chomskyan transformational views[2] arguing that it did not provide (nor attempt to provide) means to model the psychological reality of language. In another paper in the same volume, Kaplan and Bresnan (1982b) published an important seminal paper, which laid out the main views and formalisms of LFG.
In
addition to their concern with psychological reality, linguists working with LFG
have also maintained, since its beginning, that the framework should be
typologically grounded (Austin,
2001:2-3). Bresnan (1994:2-3)
argues that much of syntactic theory is based on comparisons between languages
that are too similar, which allows for a detailed level of comparisons, but
provides a weak basis for inferring deep properties of universal grammar.
She maintains that universals have been based on observations of a small number
of similar languages and then inappropriately applied in attempts to fit other
dissimilar languages. In contrast to that, LFG has found motivation for
its architecture in a variety of types of languages from English to Australian
nonconfigurational languages (see section 2.3
for a discussion of these) (Bresnan,2001:4).
Thus, both universality and variability of language are part of the design
principles of LFG. This is dealt with by having a level encoding
properties that vary across languages (c-structure)
and another which concentrates on universal properties (f-structure).
These levels will be discussed in section 2.
This section will provide a brief description of the LFG formalism. This should serve as an expository introduction to help the reader before consulting other sources (see references and recommended readings) in order to acquire more complete and accurate knowledge of the mechanisms of LFG.
As was mentioned in section 1, LFG uses parallel levels of representation, which are linked via correspondence principles. Each level models a distinct aspect of the language, having its own formal properties (Nordlinger, 1998: 10). Examples of levels are f(unctional)-structure, c(onstituent)-structure, a(rgument)-structure, semantic(s)-structure, phonological-structure and thematic-structure. These last three levels deal with logic semantic interpretation, sound systems and discourse pragmatics, respectively (Austin, 2001: 5-6). The first three, which are most relevant to syntactic analysis, will be described below.
F-structure
models the internal (or ‘covert’, Bresnan,
1994: 4) structure of language where grammatical relations are
represented. It is largely invariable across languages[3].
F-structure is formalized though a matrix of attributes associated with values.
Each matrix can be viewed as a mathematical function[4]
mapping
the attributes to the values (Bresnan,
2001: 45-50). (1) provides a simplified example of an f-structure:
(1)
(Bresnan, 2001:46)
Thus,
the f-structure in (1) can be seen as a function that when applied to TENSE, for
example, gives the value PRES (Bresnan,
2001: 48).
Attributes may be syntactic functions such as subject (SUBJ), object (OBJ), topic (TOP) and focus (FOC). They may also be tense/aspect/mood categories (such as theTENSE attribute), nominal categories (such as CASE, NUM (number) and GEND (gender)) or the predicate attribute (PRED) (Nordlinger, 1998: 11). Values may be formalized through symbols (e.g. PL for plural), semantic forms (e.g.‘lion’) or other attribute-value matrixes (i.e. other f-structures) (Bresnan,2001: 47).
F-structure is restricted by the principles of completeness, coherence and uniqueness. Completeness ensures that a predicate and all of its arguments are present in the structure:
Every
function designated by a PRED must be present in the f-structure of that PRED (Nordlinger,
1998: 13).
Coherence
can be seen as the converse of completeness (Nordlinger,
1998: 13). It ensures that all arguments[5]
present in the structure are required by a predicate:
Every
argument in an f-structure must be designated by a PRED (Nordlinger,
1998: 13).
The uniqueness principle ensures that attributes have only one value:
Every
attribute has a unique value (Nordlinger,
1998:13).
It
is important to note that uniqueness allows different attributes to have the
same value (Bresnan, 2001: 48).
For example, it is possible for something to be the value of both the OBJ and
TOP attributes (a topicalized object), but not for the same OBJ attribute to be
associated with two values.
The input of f-structures comes from information encoded in lexical items or from information encoded in the nodes of c-structure (Nordlinger, 1998: 11) (in the form of annotations, discussed in section 2.2).
C-structure
is the level where the surface syntactic form, including categorical
information, word order and phrasal grouping of constituents, is encoded.
This is the level where language variability mainly occurs. C-structure
can be expressed through phrase structure trees, determined by context free
phrasal structure rules, such as those illustrated in (5) and (6) (where ‘C’
stands for any constituent type, and the star indicates any number of
constituents):
(5)
S ®
NP VP (English)
(6)
S ®
C aux C* (Warlpiri) (Bresnan,
2001: 44)
The
phrasal structure may be restricted by a version of X-bar theory (Bresnan,2001:
44-45) that can accommodate for the great amount of cross-linguistic
variation in phrase structures (Nordlinger,
1998:10). For example, as
illustrated above, a language like English may follow different restrictions in
its phrase structure from a language like Warlpiri, a point that will be
discussed further in section 2.3.
In c-structure, each fully inflected word belongs to exactly one node, a restriction known as lexical integrity (Bresnan, 2001: 44). Related to this is the fact that the c-structure is blind to the internal structure of words[6] (Bresnan, 2001: 93) and the processes of word formation and phrasal formation are independent of each other (Nordlinger, 1998: 21-22). As lexical integrity requires words to occupy nodes in c-structure, empty categories cannot usually occupy a node[7] (Austin,2001: 6-7).
A-structures model the grammatically realizable participants of an expression (Bresnan, 1994: 4). They encode the number, type and semantic roles of the arguments of a predicate. A-structures are represented as arrays of predicates and arguments with associated slots filled by semantic roles such as agent, patient and location (Austin, 2001: 5-6). The order of the argument slots reflects their relative prominence (with the most prominent on the left), prominence being determined by a thematic hierarchy[8],such as:
(7) Thematic Hierarchy:
Agent > beneficiary > experiencer/goal > instrument > patient/theme > locative (Bresnan,2001: 307)
A-structures project skeletal f-structures through lexical mapping principles. These will not be fully described here, however it can be briefly said that the mapping principles rely on information about the prominence and type of an argument (available in argument structure). For example, an argument may be mapped onto SUBJ in f-structure if it is initial (thus the most prominent) in the a-structure and it is of a certain type (namely, having a “non-objective” feature). The reader is referred to Bresnan (2001: 302-321) for a thorough discussion. The mappings are constrained by principles such as Function-Argument Bi-Uniqueness and the Subject Condition:
(8) Function-Argument Bi-Uniqueness:
Each a-structure must be associated with a unique function, and conversely. (Bresnan, 2001: 311)
(9) The Subject Condition:
Every predicator must have a subject. (Bresnan, 2001: 311)
The diagram in (10) shows examples of representations of each of the three types discussed here:
(10)
(Bresnan,1994:
73)
The next section discusses the correspondence between these parallel levels.
This section will provide a short description of the correspondence between different levels in LFG. As mentioned above, the correspondence between a-structure and f-structure is done through lexical mapping principles taking into consideration information such as the type and the prominence of arguments listed in a-structure. The following paragraphs introduce the correspondence between the levels of f-structure and c-structure.
In languages like English (a configurational language, see discussion in section 2.3 below), c-structure contains annotations which relate c-structure to f-structure. (11a) presents an example of an annotated c-structure and its corresponding f-structure (11b):
(11)
(Nordlinger,
1998: 12)
The
upward arrow
(
)
denotes the mother of the node to which the annotation is attached, whereas the
downward arrow (
)
notation refers to the node itself (or, as put by Austin
(2001: 8), ‘’
means ‘information about my mother’ and
‘’
as ‘information about me’). Thus, the annotation ‘(
SUBJ) = ’
in (12) means that the f-structure of S (the mother) has a subject attribute,
whose value is the f-structure of NP (the node itself). The annotation ‘=’
indicates that the f-structure of the node with the annotation and the
f-structure of its mother are identified as the same; hence, this annotation
denotes the head relation[9].
For example, in (11), the
f-structure of V is identified with the f-structure of VP, which in turn is
identified with the f-structure of S (Nordlinger,
1998:12). It can also be seen from the annotations in (11a) that
the correspondence between c-structure nodes and f-structure functions is
many-to-one. That is, different nodes (such as S, VP and V) may be
associated with the same S-structure. Another point illustrated in (11) is
that each node in the c-structure is associated with an f-structure function
(i.e., the mapping is a piecewise correspondence) (Bresnan,
2001:50).
These annotations are assumed to be encoded in the phrase structure rules which determine c-structure. (12) provides examples of what those rules would look like:
(12)
(Bresnan,
2001: 56)
The idea that annotations are assigned as part of phrase construction can also be accommodated to the view that phrase structure follows X-bar theory by assuming principles such as those outlined in (13), which would apply to configurational languages like English. Although (13) will not be fully discussed here (the reader does not have to be concerned with the detail of (13) at this stage), it can be said that the result is intuitively similar to that of (12), but it incorporates assumptions of X-bar theory into the construction of phrases:
(13)
Endocentric
Mapping Principles:
a.
C-structure heads: annotate a projecting node in a projection of the same
kind with =.
b.
Specifiers: annotate a non-projecting node in functional category
projections with (
DF)
= .
c.
Functional coheads: annotate
a non- projecting complement node with =[10].
d.
Complements: Annotate a non-projecting complement node dominated by any
lexical category with (
CF)
= .
e.
Adjuncts: Optionally annotate a non-projecting node and its
adjoined-to-sister node with (
AF)
=
and =
respectively.
(Based on Nordlinger, 1998: 15 and Bresnan, 2001: 119)
Note: CF extends for a grammaticalized discourse function, such as TOP, FOC, and SUBJ. DF extends for non-discourse argument functions, such as OBJ and COMPL (complement). AF extends for non-argument functions, such as ADJUNCT, TOP and FOC (see Bresnan, 2001: 102 and Nordlinger, 1998: 15).
The information contained in c-structure annotations (determined by restrictions similar to (12) or (13)) would have to be compatible with the information contained in f-structure for the sentence to be grammatical. This can be summarized (in a simplified manner) as follows. A variable would be assigned to each annotated node in the c-structure tree, which can be seen from (14) (variables are the functions denoted by each ¦n symbol):
(14)
(Austin,
2001: 31)
The
variables would then be resolved by functional equations. For example, for
the subject DP in (14), they would be:
(15)
(¦1SUBJ)
= ¦2
(¦2DEF)
= ¦3
(¦2DIST) = ¦3
(¦2PRED) = ¦4 (Austin,2001: 17)
The equations would be resolved as in (16):
(16)
(Austin,
2001: 17)
The information provided by c-structure would be added to lexical information in f-structure (Austin, 2001). The information from both sources would have to be compatible in a grammatical sentence (that is they would have to unify) (Carnie, 2002: 344). This is described in more detail below.
Unification[11] is the operation used in the manipulation of logical of symbolic expressions such as above. Sells (1985: 149) gives a simplified description of this operation by saying that it can be seen as being similar to the union of sets, making information in those sets identical by merging them. However, differently from simple union, unification fails if the union results in an attribute being specified with conflicting values. The ‘resolution’ of the subject DP (14) seen in (16) was example of unification: (16) represents the union of the values of the different functions seen in (15). Other examples, involving small partial structures are provided in (17):
(17) (a) [NUM SG] unifies with [PERS 3] to give:
(b) [NUM SG] unifies with [NUM SG] to give [NUM SG]
(c) [NUM SG] fails to unify with [NUM PL] (Sells, 1985:149)
There are at least two options in which unification could be implemented in complex expressions, which can be described in broad terms as the calculation of unification in a top-down or in a bottom-up manner. These are associated with the terms ‘outside-in’ and ‘inside-out’ functional application. Outside-in (or ‘normal’, as put by Andrews, 1996: 24-41) unification works by first combining two feature structures at their roots and then combining the result with the remaining nodes in a similar way, in a top-down manner. Inside-out unification begins with two distinct f-structures that share a substructure and combines them in a recursive bottom up manner. Inside-out unification can be implemented using the concept of ‘inside-out uncertainty’ (Halvorsen and Kaplan (1988) and Dalrymple (1993) cited in Andrews; Bresnan (2001: 64-69) and Nordlinger (1998: 62)). In broad terms, this concept can be described as relating the general intuitions here to applications of chains of functions[12],allowing the mathematical implementation of inside-out unification. For a more detailed description, refer to the work cited here (noting that acquaintance with functions is required and that authors often leave out the term ‘inside-out unification’, referring to the concept of inside-out uncertainty instead).
The distinction between these two types of unification is important in that there are analyses which rely on inside-out application in the explanation of some specific linguistic phenomena. These include accounts of some nonconfigurational languages (see Nordlinger (1998), mentioned in section 2.3.1), long distance dependencies such as topicalization (see section 2.3.3), anaphoric binding (Dalrymple (1993) cited in Andrews (1996: 41) and Nordlinger (1998: 62)) and case stacking in Australian languages (Andrews, 1996). Some of these phenomena will be mentioned in this summary, but the differences between the two types of application require a lengthy description, thus it will not be included in this introduction. The reader is referred to the references cited earlier in this paragraph, in addition to Bresnan (2001: 64-69) to learn more about this topic.
This section will provide some examples of the application of LFG to the explanation of linguistic phenomena. Firstly, two phenomena which were important in the motivation of the LFG design, namely nonconfigurationality and movement paradoxes, will be briefly described (Bresnan, 2001: 4). This will be followed by a sketch of how LFG deals with long distance dependencies.
2.3.1.
Nonconfigurationality
The existence of nonconfigurational languages was an important motivation for the LFG architecture (Bresnan, 2001: 5-14). One of the properties of nonconfigurational languages is their relatively free word order[13], illustrated by the following example from Warlpiri:
(18) Ngarrka-ngku ka wawirri panti-rni. (Hale, 1983: 6)
Man ERGAUX Kangaroo spear non-past
Wawirri kapanti-rni ngarrka-ngku.
Panti-rin kangarrka-ngka wairri.
'The man is spearing the kangaroo'.
All of these utterances have exactly the same meaning in which ‘Ngarrka-ngku’ (man) is interpreted as the subject (Hale, 1983: 6). Languages like Warlpiri suggest that grammatical functions such as subject and object may not necessarily be defined by the phrase structure. Instead, the marking of case or agreement (or both) on words seems to be used to indicate what are the subject and the object in the sentence. For example, the word for man in (18) is marked with the ergative case marker (-‘ngku’), which is used to mark subjects in ergative languages. Rich morphology in Warlpiri is an example of a tendency for languages with rich morphological marking to be able to have less restricted word orders, whereas languages with poor morphology tend to have strict hierarchical phrase structures (Bresnan, 2001: 6).
Nonconfigurationality suggests that grammatical relations (e.g. subject, object, etc) may be coded in the shape of words, rather than in phrasal structure. The possibility of encoding function through words or phrases (or both) underlies the LFG design (Bresnan, 2001: 6). It distinguishes it from frameworks such as P&P where the notion of subject and object is defined structurally (the object is the NP occupying the specifier of IP and the object is the NP occupying the complement position of the verb (Baker, 2001: 209)). Broadly speaking, LFG can account for nonconfigurationality by encoding information about grammatical relations such as subject and object in f-structure rather than in c-structure (see Austin and Bresnan, 1996). The representation in (19) illustrates the imperfect mapping between f-structure and c-structure of in aWarlpiri sentence:
(19)
(Austin
&
Bresnan, 1996: 229)
Simplifying the LFG approach to nonconfigurationality to fit the concepts introduced in sections 2.1 and 2.2, it could be said that configurational languages have annotations in c-structure which relate nodes in a tree to relations such as subject and object (as discussed in section 2.2). In nonconfigurational languages, these annotations would not be present in c-structure (Austin, 2001: 17-18); hence, the position of NPs would have nothing to say about their grammatical function. One suggestion for these cases would be that the rich morphology in nonconfigurational languages have similar annotations which build their f-structures (Nordlinger, 1998, cited in Austin). Simplifying Nordlinger’s proposal, she argues that both case and agreement morphology directly construct grammatical structures. Without getting into details, it can be said that her approach makes use of the mechanism of inside-out functional uncertainty, mentioned in section 2.2, in order to formalize her idea that morphology constructs grammatical functions[14]. The mechanism reverses functional relations allowing the case morpheme to directly specify the grammatical function of the f-structure to which they belong (instead of having c-structure determining functions from the top of the tree outside-in, as in configurational languages). Thus, the ergative case marker in (18) above would build a partial f-structure having the SUBJ attribute with the marked NP as its value.
2.3.2.
Movement Paradoxes
Another example of the application of LFG is in the explanation of movement paradoxes. These are cases in which there seems to be a mismatch between the category of a constituent and the position where it is supposed to originate, as in example (20):
(20)
The sentences(20b) and (20c) indicate that the preposition about requires a NP as its complement. Thus, the fact that the topicalized constituent in (20a) is a sentence constitutes a movement paradox. This and other types of movement paradoxes (see Bresnan,2001: 16-24 for other examples) present a case against the ideas that language has the formal properties of categorical phrase structure underlyingly and that their surface form is derived from underlying structure via movement (Bresnan, 2001:19). This is because under such a movement analysis the three sentences would seem to have similar underlying configuration. Nevertheless, about seems to disallow a sentence to function as its object in (20b), but not in (20a).
Mismatches such as this are allowed under LFG assumptions because the correspondence between f-structure and c-structure is not perfect and the different structures have different constraints (Bresnan, 2001: 16-23). So both the NP following about in (20c) and the topicalized sentence in (20a) are objects (values of the OBJ attribute in f-structure), as indicated by the interpretation of these sentences. The ungrammaticality of (20b) can be explained by a constraint in English c-structure requiring the NP occupying the complement structural position in (20 to be nominal (see Bresnan, 1994: 104). The topic sentence in (20a) is base generated in its surface position, thus it does not violate that constraint. It is allowed to have the object function in f-structure, independently from its category or c-structure position (Bresnan, 1994: 104 – 108).
2.3.3.
Long Distance Dependencies
The previous example is related to the fact that LFG does not use movement in its analysis: the lack of movement correctly allowed for the existence of apparent movement paradoxes. However, the lack of a movement assumption pressures LFG to provide alternatives to P&P analyses that use movement to explain linguistic phenomena. This section provides an example of how LFG may provide alternatives to movement explanations by sketching its account of long distance dependencies (in configurational languages), in a further discussion of topicalization, mentioned in the previous section.
Sentences (21) and (22) provide examples of long distance dependencies in which an element in the beginning of the sentence is associated with a grammatical function which usually appears later in the clause:
(21) That person, I’ve never seen you talk to.
(22) Which analysis does Joan seem to prefer?
(Adapted from Austin, 2001: 21)
These dependencies are said to be unbounded in that the distance between the two related positions is unlimited[15] (Austin, 2001: 22). For example, compare (22) to (23):
(23) Which analysis does Joan seem to prefer to write about? (Austin, 2001: 21)
In P&P, these dependencies are attributed to a-bar movement of an element from its underlying position (the position associated with the interpretation of its grammatical function) to the position in which it appears in the surface. As LFG uses no movement, it could not explain this in the same way. In a simplified description, LFG accounts involve the assignment of a functional equation identifying the initial element of the sentence at the same time with a discourse function, such as TOP and FOC (topic and focus) and with a grammatical function, such as SUBJ and OBJ (subject and object), in the later part of the sentence. The identification of these two functions may follow paths (in f-structure) of variable lengths (i.e. embedded in different numbers of COMP (complement) attributes (Austin, 2001: 21-23). This is illustrated by (22) and(23) above: in (23), the function of the wh-element would be associated with the COMP attribute of about which is embedded in the value of the COMP attribute of prefer. In (22), the function associated with the expression topic expression is not as deeply embedded.
This operation of identification of an element as the value of both OBJ and TOP makes use of the concept of inside-out[16] functional uncertainty functional uncertainty, mentioned in section 2.2, a term which describes the mathematical properties of the identifying the functional equation. The interested reader is referred to Bresnan (2001: 64-69) for a mathematical description. In this introductory summary, it will only be said in broad terms that the functional equation is uncertain to allow for the fact that the number of COMP embeddings (a string of functions in a chain) in these dependencies is uncertain (Austin,2001: 22). The identifying operation is determined by (24):
(24) Principle for Identifying Gaps:
Associate
XP 
e with ((x)
DF) =
(Bresnan, 2001: 181)
(Associate
an empty position dominated by XP with the annotation ‘((x)DF)
= ’,
where ‘DF’ stands for ‘discourse function’)
Sentence (25) illustrates the application of (24) to a gap:
(25)
(Adapted
from Bresnan, 2001: 67)
Note: letters subscripts are related to the annotations in (26) below.
The annotation added by (24) acts together with other annotations, such as the ones determined by the Endocentric Mapping Principles in (13). These Principles will annotate c-structure nodes as follows:
(26)
Annotations added to (25) by application of (13):
a.
The NP sister to IP is annotated with (
TOP)
=
(12d)
b.
The NP sister to VP is annotated with (
SUBJ)
=
(12X)
c.
The IP sister of V with (
COMPL)
=
d.
The NP sister to VP is annotated with (
SUBJ)
=
(12X)
e.
The IP sister of V with (
OBJ)
=
f.
All other categories with
=
(Adapted from Bresnan, 2001: 68)
This leads to the following f-structure, where the attribute h is the value of both the TOP and the OBJ attributes:
(27)
(Bresnan,
2001: 68)
Without the functional identification between the TOP and OBJ with their value, the f-structure would be incomplete (because of the empty object) and incoherent (because of the failure to integrate TOP into the f-structure).
Thus, long distance dependencies are implemented through a functional equation which adds an annotation to empty c-structure nodes. The annotation leads to a relationship in f-structure between the function associated (due to Endocentric Mapping Principles) with the gap position and the function of the topicalized element, namely the two attributes (in this case TOP and OBJ) share the same value.
This section provided illustrations on how LFG may be applied to explain different linguistic phenomena. Of course, there are many other examples of analysis which use LFG. The references cited here can lead the reader to such examples.
3.
Similarities and differences between LFG, P&P
and
HPSG
Broadly speaking, LFG, HPSG and P&P have a common goal to provide a theory of language structure, including its universal properties (see Pollard, 1997: 1, Bresnan,2001: 3-5 and Webelhuth, 1995: 3). However, they have different focuses on their goals. Both P&P and HPSG view language as a system of knowledge in the mind of the speaker, without focus on the psycholinguist processing (Pollard, 1997:1). On the contrary, LFG has a large focus on the processing and psychological reality of language (Bresnan, 1982a). Furthermore, linguists working with P&P have placed focus on the language acquisition device which allows the speaker to acquire language (Carnie, 2002: 13-16), while HPSG is focused on the constraints present in the linguistic system of adults, without as much direct focus on acquisition (Pollard, 1997: 3). However, it seems that the major differences between these frameworks is not on their overall goals, but on the way they pursue investigation to achieve it. For example, on the one hand, HPSG requires a high level of mathematical precision in its analysis, at the cost of making slower and more conservative claims than P&P (Pollard, 1996)[17]. On the other hand, P&P focuses on explanatory adequacy, perhaps at the cost of mathematical precision (Horrocks, 1978: 298-299). Like HPSG, LFG also has a focus on mathematical precision, but, as previously mentioned, it seems to be more directly concerned with psychological reality than HPSG.
The three frameworks differ in the assumptions of their model. Some of the assumptions of LFG and HPSG are similar, but differ from P&P. Pollard (1996) points out that LFG may be more similar to a version of HPSG than two different versions of HPSG are to each other. Both LFG and HPSG are unification and constraint based non-derivational approaches, as opposed to P&P which derives one structure from another by applying operations such as move-a (Kim, 2000: 7-8 and Bresnan, 2001: vii). The non-derivational aspect of LFG and HPSG is related to the fact that there is no feel that the phonological and logical contents are derived from, or some type of interpretation of, the syntactic structures in any way (Pollard, 1997: 2). In LFG and HPSG empty categories are avoided (Pollard, 1997: 7 and Austin, 2001: 6-7), whereas they are often used in P&P. LFG and HPSG also have in common the fact that notions such as subject and object are not defined in terms of positions in a tree. Instead, they are formalized in terms of attributes (such as SUBJ and OBJ in LFG and SPR (specifier) and COMP (complement) in the version of HPSG described in this website). Furthermore, P&P functional projections and empty categories make P&P trees much richer in information than the trees in other frameworks (Borsley, 1996: 238). HPSG differs from both P&P and LFG in the number of levels of representations it uses. HPSG has only one level (encoded in a complex sign structure) containing information about phonology, syntax and semantic interpretation of an expression (Cooper, 1996: 192), unlike P&P, which treats these as different levels. It is also in this one level where it is indicated which arguments function as subject and object, differently from LFG which may encode syntactical hierarchical structure in one level (c-structure) and notions of subject and object in another (f-structure) (Bresnan, 2001: 44- 45).
The differences discussed in this section may raise the question as to which framework presents the right approach. Different linguists tend to favor particular frameworks and provide a lot of arguments indicating why their frameworks are the most appropriate. Some points in the different frameworks should lead to different predictions that could help in deciding between them. But it seems that, at this stage, different frameworks are more appropriate to account for different empirical evidence (Carnie, 2002: 371). As much as the differences between these frameworks represent different views on language and may lead to significant empirical differences, it is important to note that all three are formal representations of hypotheses on how language works. As language is part of science, it is tentative, meaning that it provides approximate explanations for phenomena, which are continuously improved (Campbell, 1996:16-17). In other words, the views expressed by none of these frameworks represent the ultimate truth. Although it is possible that none of the three is right (Carnie, 2002: 372), each provides approximations to an explanation on how language works. Linguists working under a particular framework may believe that it provides the best approximation at this stage; however, it is worth taking note of the findings made under other frameworks.
Andrews, Avery D. 1996. Semantic Case-Stacking and Inside-Out Unification. Australian Journal of Linguistics 16(1):1-55.
Austin, P. 2001. Lexical Functional Grammar. In N.J. Smelser and P. Baltes (eds.), International Encyclopedia of the Social and Behavioural Sciences, 8748-8754. Elsevier. <http://www.linguistics.unimelb.edu.au/contact/staff/peter/Elsevier.pdf>
Austin, Peter K. and Joan Bresnan. 1996. Nonconfigurationality in Australian Aboriginal languages. Natural Language and Linguistic Theory 14(2):215-268.
Baker, Mark C. 2001. The Natures of Nonconfigurationality. In Mark Baltin and Chris Collins (eds.), The Handbook of Contemporary Syntactic Theory, 407-438. Oxford: Blackwell Publishing Ltd.
Borsley, Robert D. 1996. Modern Phrase Structure Grammar, Cambridge, MA: Blackwell Publishers Ltd.
Bresnan,
Joan. 1994. Locative Inversion and the Architecture of Universal
Grammar. Language 70:72-131.
Bresnan, Joan. 2001. Lexical-Functional Syntax, Oxford: Blackwell Publishers Ltd.
Campell, Neil A. 1996. Biology. Chapter 1: Themes in the Study of Life. Menlo Park, CA: Benjamin/Cummings Publishing Company
Carnie, Andrew. 2002. Syntax: A Generative Introduction, Oxford: Blackwell Publishing.
Cooper, Richard P. 1996. Head-Driven Phrase Structure Grammar. In K. Brown and J. Miller (ed.), Concise Encyclopedia of Syntactic Theories, 191-196. Oxford: Pergamon.
Dowty,
David R. 1991. Thematic Proto-Roles and Argument Selection. Language
67:547-619.
Haegeman, Liliane. 1994. Introduction to Government and Binding Theory, Oxford: Blackwell Publishing.
Hale, Kenneth. 1983. Warlpiri and the Grammar of Nonconfigurational Languages. Natural Language and Linguistic Theory 1:5-47.
Hornstein, Norbert and Amy Weinberg. 1995. The Empty Category Principle. In Gert Webelhuth (ed.), Government and Binding Theory and the Minimalist Program, 241-296. Cambridge, MA: Blackwell Publishers.
Horrocks, Geoffrey. 1987. Generative Grammar, Essex: Longman.
Kaplan, Ronald M. and Joan Bresnan. 1982a. Introduction: Grammars as Mental Representations of Language. In Joan Bresnan (ed.), The Mental Representation of Grammatical Relations, xvii-lii. Cambridge, MA: MIT Press.
Kaplan, Ronald M. and Joan Bresnan. 1982b. Lexical-Functional Grammar: A Formal System for Grammatical Representation. In Joan Bresnan (ed.), The Mental Representation of Grammatical Relations, 173-281. Cambridge, MA: MIT Press.
Kim,
Jong-Bok. 2000.
The Grammar of Negation: A Constraint-Based Approach.
Chapter 1: Introduction and Theoretical Foundations.
Stanford: CSLI Publications.
Marantz,
Alec. 1995.
The Minimalist Program. In
Gert Webelhuth (ed.), Government and Binding Theory and the Minimalist
Program, 349-382. Cambridge,
MA: Blackwell Publishers.
Mylne, T. 2000. Thematic Role Hierarchies and Role Engagement. In J. Henderson (ed.), Proceedings of the 1999 Conference of the Australian Linguistic Society. <http://www.arts.uwa.edu.au/LingWWW/als99/proceedings>
Nordlinger, Rachel. 1998. Constructive Case: Evidence from Australian Languages, Stanford: CSLI Publications.
Pollard, Carl. 1996. The Nature of Constraint-Based Grammar. Unpublished manuscript: Ohio State University. <http://www-csli.stanford.edu/~sag/L221a/pollard-96.txt>
Pollard,
Carl. 1997.
Lectures on the Foundations of HPSG.
Unpublished manuscript: Ohio State University. <http://www-csli.stanford.edu/~sag/L221a/cp-lec-notes.pdf>
Saeed, John I. 2003. Semantics, Oxford: Blackwell Publishing.
Sells, Peter. 1985. Lectures on Contemporary Syntactic Theories. Chapter 4: Lexical-Functional Grammar. Stanford: CSLI Publications.
Webelhuth, Gert. 1995. Introduction by the Editor. In Gert Webelhuth (ed.), Government and Binding Theory and the Minimalist Program, 1-14. Cambridge, MA: Blackwell Publishers.
[1] In this website, the abbreviation P&P (Principles and Parameters) refers to various versions of Chomskyan approaches, such as GB (e.g., Haegeman, 1994) and Minimalism (e.g. Marantz, 1995). It is not claimed that these are equivalent, but a discussion of the differences between them is outside the scope of this work. This group will be compared to HPSG and LFG as a whole.
[2] Carnie (2002: 336) points out that the Chomskyan approaches in the 1970’s relied much more heavily in transformation than the more recent P&P and Minimalist approaches.
[3] The abstract relations in f-structure are different from the expressions in the language that mark them: the same f-structure relation (as number, case and agreement for example) may be marked morphologically in some languages syntactically in others (Bresnan, 2001: 45).
[4] Bresnan (2001: 47-60) provides a more detailed, but accessible, discussion of f-structures as mathematical functions.
[5] There is also an extended coherence condition which is similar to (3), but applies not only to arguments but to all syntactic functions, ensuring they are appropriately integrated in f-structure (Bresnan,2001: 63).
[6] Notice that this does not apply to syntax as a whole, as f-structure has access to the information contained in words (Bresnan, 2001:93).
[7] Although there is an avoidance of empty nodes, they are used in some LFG analysis (e.g. Bresnan, 2001: 180 – 183).
[8] A number of problems have been identified with the hierarchy in (7). Saeed (2003: 150- 174) provides an accessible introduction to theta-roles and the problems associated with them. Another important reference in this is Dowty (1991). For a discussion of an alternative approach, see Mylne (1999). Bresnan (2001: 308) classifies the roles above into features (based on restrictiveness of syntactic function and potential for being an object) to constrain the mapping of roles, which may address some of the problems with the hierarchy (while weakening the need for it).
[9] It is worth noting that this is a functional head relation (i.e., in f-structure). This is because functional heads are not equivalent to structural c-structure heads (Bresnan, 2001: 53). C-structure heads are heads in terms of X-bar theory, similarly to P&P (e.g. a verb is the structural head of VP in configurational languages). However, in LFG complements of functional categories end up being f-structure co-heads with the c-structure heads (this is related to the fact that the (functional) head relation is concerned the identification of a node’s f-structure), illustrating the imperfect correspondence between function and structure in LFG.
[10] This is added to the other annotations determined in (13) by logical conjunction (see Bresnan, 2001: 109).
[11] For a description of the process of unification, described in the HPSG framework, click here. This description is put in slightly different terms than the one (Sells’) above for LFG, but it expresses the same general intuition of a combination of values and failure by having incompatible features.
[12] Bresnan (2001: 64-65) provides a discussion of the mathematical difference between inside-out and outside-in functional uncertainty. In simplified terms, the difference lies in the way chains of embedded functions (e.g., (((¦a1) a2) a3) = v) are solved. In inside-out functional uncertainty, the internal function (¦a1) is resolved first and the result is then applied to a2, which is then applied to a3. In outside-in functional application, the function evaluation is done by evaluating if the tail of the chain is part of an equation, making the tail progressively smaller. Consult Bresnan for more details.
[13] This is a very simplified description of nonconfigurationality. Free word order is one of other properties that seem to suggest that they lack a hierarchical distinction between subject and object. That is, these languages may not have a VP, having a flat-structure where the subject is not higher than the object. For more on LFG accounts to nonconfigurationality consult Nordlinger (1998) and Austin and Bresnan (1996).
[14] The major innovation of Nordlinger’s account, which is not discussed here, is in implementing the idea that grammatical function is built from in similar ways from either agreement or case. She argues that standard LFG accounts rightly treat agreement morphology as constructing grammatical relations, but fails to formally capture the same constructive property of case morphology (1998:50). In the former case, verbal agreement morphology directly carries the information about argument structures in the clause, as it contains a specification of the grammatical function, whereas the case morphology in languages carries no direct information about grammatical function, carrying only information about case. Grammatical function is assigned indirectly by implicational relationships between case and grammatical function (refer to Nordlinger for more details). Nordlinger argues that case morphology is also directly constructive and uses inside-out functional uncertainty to implement that idea.
[15]
It is ‘unlimited’ in the sense that the distance can be unlimitedly
great. That is not to say that the dependency is not unconstrained by
other factors, as shown buy the sentences below (which have been accounted for
through the use of notions such as the ECP and the that-trace filter in P&P)
(Hornstein and Weinberg, 1995: 245-246):
(a)
*Who do you think that left the present on the table?
(b)
*Who do you recall whether brought a book?
[16] Though Bresnan (2001: 67) points out that there are LFG accounts which proposed ‘outside-in’ functional uncertainty or the use of ‘constituent control’ metavariables instead.
[17] Pollard (1996) provides interesting arguments for the priority of empirical adequacy over explanatory adequacy.