Lexical Functional Grammar

Lexical Functional Grammar
Carol Neidle, Boston University
The term Lexical Functional Grammar (LFG) first appeared in print in the 1982 volume edited by
Joan Bresnan: The Mental Representation of Grammatical Relations, the culmination of many years
of research. LFG differs from both transformational grammar and relational grammar in assuming
a single level of syntactic structure. LFG rejects syntactic movement of constituents as the
mechanism by which the surface syntactic realization of arguments is determined, and it disallows
alteration of grammatical relations within the syntax. A unique constituent structure,
corresponding to the superficial phrase structure tree, is postulated. This is made possible by an
enriched lexical component that accounts for regularities in the possible mappings of arguments
into syntactic structures. For example, the alternation in the syntactic position in which the logical
object (theme argument) appears in corresponding active and passive sentences has been viewed by
many linguists as fundamentally syntactic in nature, resulting, within transformational grammar,
from syntactic movement of constituents. However, LFG eliminates the need for a multi-level
syntactic representation by allowing for such alternations to be accomplished through regular,
universally constrained, productive lexical processes that determine multiple sets of associations of
arguments (such as agent, theme) with grammatical functions (such as SUBJECT, OBJECT)—
considered within this framework to be primitives—which then map directly into the syntax.
This dissociation of syntactic structure from predicate argument structure
(a rejection of Chomsky’s Projection Principle, in essence) is crucial to the LFG framework. The
single level of syntactic representation, constituent structure
(c-structure), exists simultaneously with a functional structure (f-structure) representation that
integrates the information from c-structure and from the lexicon. While c-structure varies
somewhat across languages, the f-structure representation, which contains all necessary
information for the semantic interpretation of an utterance, is claimed to be universal.

---

Phenomena that had been explained by the interaction of transformations are accounted for in
LFG by the regular interaction of lexical processes. Bresnan shows that some of the classic
arguments for syntactic transformations do not, in fact, distinguish between a transformational and
a lexical account of the regularities. In Bresnan 1982b, she argues that the lexical account of
passivization is superior to the transformational approach, e.g., in explaining why passivized
forms can undergo further lexical rules, such as Adjective Conversion and compounding (giving
rise to such forms as “snow-covered”).
Bresnan and other contributors to Bresnan 1982a offer evidence and arguments in support of
the formulation of such alternations in terms of alternative assignments of grammatical functions to
arguments rather than syntactic movement. They suggest that the model has psychological validity,
and is consistent with evidence about grammatical processing and acquisition. It also captures
cross-linguistic generalizations about languages that have comparable alternations in the realization
of arguments as grammatical functions despite the use of very different syntactic means for
expressing functions like subject and object. Bresnan suggests that the “illusion” of NP-movement
in the English active/passive alternations is just an artifact of the structural encoding of object and
subject through word order in English. This is in contrast with languages like Malayalam (see
Mohanan 1982), in which word order is much freer; accordingly, passivization in Malayalam
involves an apparent change in morphological case. In LFG, the different realizations of active
and passive sentences in Malayalam and English follow directly from independent principles that
determine how subject, object, and oblique phrases are expressed syntactically in those languages.
Recent work in Lexical Mapping Theory, an outgrowth and extension of Lexical Functional
Grammar, has been refining the principles for association of arguments with grammatical
functions, so some of the earlier work could now be recast accordingly to simplify the lexical
component. Thanks to a more general set of mapping principles, the associations of grammatical
functions with arguments no longer need to be stipulated, and many lexical redundancy rules can
be eliminated or greatly simplified since many characteristics of the associations are now

---

predictable. Sections 1-5 below summarize the 1982 framework; Lexical Mapping Theory will be
discussed in section 6.
1. Levels of Representation
Lexical Functional Grammar postulates three distinct but interrelated levels of representation:
lexical structure, functional structure, and constituent structure, which are present simultaneously.
See Kaplan and Bresnan 1982 for details of the LFG formalism, which is briefly summarized
below.
1.1 Lexical Structure
The lexical entry (or semantic form) includes information about the meaning of the lexical item, its
argument structure, and the grammatical functions (e.g., subject, object, etc.) that are associated
with those arguments. The verb ‘hit’, for example, has a predicate argument structure that consists
of an agentive argument associated with SUBJECT and a patient or theme argument associated with
the OBJECT function.

(SUB) (OBJ) <-—- lexical assignment of grammatical functions
| |
‘hit ( agent, theme )’
<——- predicate argument structure

Grammatical functions are universal primitives within this framework, and since they are
associated both with lexical items and with syntactic positions—by means of annotated phrase
structure rules—they mediate between lexical and constituent structure representations.
Grammatical functions play an essential role in Lexical Functional Grammar; however, they have
no intrinsic significance and are situated at the interface between the lexicon and the syntax. LFG
imposes the restriction of Direct Syntactic Encoding, which prevents any syntactic process from
altering the initial assignment of grammatical function.
Each lexical entry consists of a pairing of arguments and grammatical functions. The
principle of Function-Argument Biuniqueness requires that each argument be associated with a

---

unique grammatical function (even if that assignment is Ø, which entails that the argument will be
interpreted as a bound variable, as in ‘John ate’ where it is implied that there is something John
ate), and conversely that no grammatical function may occur more than once within a predicate
argument structure. An actual lexical entry for the verb ‘hit’, then, might look something like this:
hit, Verb
( ↑ PRED) = ‘{meaning of hit} < SUB, OBJ >’
where the PRED feature has as its value some representation of the “meaning” of ‘hit’, which in
this case is a two-place predicate. The variable ‘↑’ in this representation refers to the lexical item
under which this entry is found, here ‘hit’.
A grammatical function may, however, be directly associated with no logical argument of
the predicate with which it occurs. This is the situation for the object of ‘consider’ in the sentence
John considered her to be a fine candidate.
where ‘her’ is the logical subject of the infinitival complement. This is indicated in the lexical entry
for ‘consider’ by placement of the function OBJ outside of the angled brackets containing the
arguments of the verb. So, the verb ‘consider’ would be represented as:
consider, Verb
( ↑ PRED) = ‘{meaning of consider} < SUB, XCOMP > (OBJ)’
( ↑ OBJ ) = ( ↑ XCOMP SUB)
where the XCOMP is an open complement, i.e., a complement whose subject is controlled
grammatically; the control equation is added (by default).
Any other grammatical information associated with a lexical item will also be encoded in the
semantic form. The name ‘Mary’, for example, comes with the grammatical information about

---

gender and number features (expressed here using ‘–’ to indicate the unmarked value of the feature
in the Jakobsonian sense; cf. Neidle 1988), which may also be expressed by equations:
Mary, Noun
( ↑ PRED) = ‘{meaning of ‘Mary’}’
( ↑ NUM ) = –PL
( ↑ GEND) = +FEM
These equations are referred to as constituting equations because the information contained in them
will be incorporated into any f-structure that contains this semantic form. It is also possible have
constraint equations in a lexical entry; in such a case the f-structure would only be well-formed if
the equation holds, but the information expressed by the equation would not be added to the
functional structure. Verb agreement in English may be accomplished in this way, by associating a
constraint equation with a form like ‘speaks’ (contributed by a redundancy rule added to all forms
that have the same inflectional ending):
speaks, Verb
( ↑ PRED) = ‘{meaning of ‘speak’} <SUB>’
( ↑ SUB NUM ) =c -PL
A sentence like ‘They speaks’ would be ill-formed since the constraint equation is not satisfied.
Lexical redundancy rules relate alternate pairings of arguments to grammatical functions. So,
for example, passivization may involve suppression of the first argument (associated with SUB in
the active form) and realization of the second argument (the OBJ in the active) as SUB; the
morphological form associated with this operation is the participial form of the verb.
(i) (SUB)
—


>
Ø

(ii) (OBJ) —>
(SUB)
Notice that Function-Argument Biuniqueness ensures that part (ii) of the passivization rule is
contingent upon part (i); there can be only one subject.

---

The output of that lexical redundancy rule on the previous lexical form given for ‘hit’ would be:

Ø (SUB)
| |

‘ hit ( agent, theme ) ‘

However, the rule applies quite generally to lexical items having the appropriate grammatical
functions. Notice that the rule would apply as well to grammatical functions contained within
control equations, so the passivized version of ‘consider’ would be:
consider, Verb
( ↑ PRED) = ‘{meaning of consider} <
Ø
, XCOMP > (
S

UB
) ‘
( ↑
S

UB
) = ( ↑ XCOMP SUB)
Some support for formulation of such rules in terms of grammatical functions rather than
structural configuration comes from the contrast illustrated by the following two sentences,
distinguished by the fact that ‘a doctor’ is an OBJECT in the first but is a NOMINAL COMPLEMENT
in the second. Only the first may be subject to passivization.
John hit a doctor.
A doctor was hit.
John became a doctor.
*A doctor was become.
While the configuration of the post-verbal NP in both cases may be the same, it is the difference in
grammatical function that accounts for the contrast with respect to passivization.
1.2 Constituent Structure
Constituent structure encodes linear order, hierarchical groupings, and syntactic categories of
constituents, and is the input to the phonological component of the grammar. Language-specific
annotations of phrase structure rules identify the grammatical functions that may occur in specific
syntactic positions. Examples of phrase structure rules for English:

---

S ——> NP VP
VP ——> V NP
(↑ SUB) =↓ ↑=↓
↑=↓ (↑ OBJ) =↓
The arrows are variables; ‘↑’ is to be instantiated by the node immediately dominating the
constituent under which the arrow is placed, and ‘↓’ by that node itself. So, the first equation for
the rule on the left states that the NP under which the equation is written is the SUB of the S that
dominates it. The ‘↑=↓’ equation beneath VP indicates that the features of that node are shared
with the higher node. This is the default assignment to phrasal heads, which share information
with the dominating phrasal node. These equations are used to construct the f-structure
representations described in 1.3.
It should be noted that the equations illustrated here are in the form in which LFG phrase
structure rules were written in 1982. Similar associations of grammatical functions could be made
with phrase structure rules conforming to current versions of X’-theory.
The terminal nodes of the tree are lexical items. The Lexical Integrity Hypothesis requires
that fully formed lexical items are inserted into the syntax. A rule like Affix-hopping would be
disallowed. Syntactic rules are prohibited from moving any element into or out of lexical
categories.
1.3 Functional Structure
Structural and lexical information is integrated and unified within functional structure (f-structure),
which consists of hierarchically organized attribute-value matrices. A straightforward algorithm
for transferring information from c-structure to f-structure is presented in Kaplan and Bresnan
1982. When the lexical items that occupy the terminal nodes of the tree are inserted into f-
structure, the information contained in the lexical entry (including relevant equations) is retrieved
and included in the f-structure. It is in this way that lexical information is combined with the
structural information available from the c-structure tree.

---

So, the f-structure corresponding to the sentence ‘John hit Bill’—constructed from the c-
structure representation generated by the phrase structure rules illustrated in 1.2 and the lexical
information from the entry for ‘hit’ discussed in 1.1—would include the following information:
SUB
[ PRED ‘John’ ]

PRED
‘hit < SUB , OBJ >’
OBJ
[ PRED ‘Bill’ ]
The validity of the f-structure representation is ensured by a number of well-formedness
conditions.
2. Well-formedness Conditions on Functional Structure
The following basic well-formedness conditions, which have counterparts in other frameworks,
apply to f-structures.
2.1 Coherence
Coherence requires that every meaningful semantic form be a grammatical function mentioned in
the predicate argument structure (or in a constituting equation) of a predicate in its clause. This
prevents extraneous material from appearing.
2.2 Completeness
An f-structure is ill-formed if it does not contain values for the grammatical functions that are
subcategorized by the predicate. The following sentence, for example, lacks a value for the SUB,
and is therefore incomplete:
* Speaks.

---

2.3 Consistency
Consistency, also known as functional uniqueness, requires that each attribute in the matrix
have a unique value. So, for example, if an f-structure contained a matrix with the following :
GEND + FEM
GEND – FEM
the f-structure would be inconsistent.
Notice that this comonsense principle can also be used to guarantee the complementary
distribution of elements that may fulfill a single grammatical function. This kind of complementary
distribution has motivated many syntactic movement analyses, such as clitic-movement. In
French, for example, both a full NP object appearing post-verbally and a pre-verbal direct object
clitic may be associated with the OBJECT function. Thus, both of the following are grammatical:
Jean le voit.
John him sees
‘John sees him’
Jean voit l’homme.
John sees the man
‘John sees the man’
However, the following is ungrammatical without a pause before the final NP:
*
Jean le voit
l’homme.
John him sees
the man
While this distribution could be accounted for by a movement analysis (such as Kayne’s 1975
proposal that both clitics and full NP’s are generated in post-verbal position, and that a rule of
clitic-placement applies) there is a straightforward account of these facts without movement. As
discussed in Grimshaw 1982, the object function is associated with the pre-verbal clitic and with
the post-verbal NP, both of which are optionally included in the phrase structure expansion of VP.

---

If either a direct object clitic or full NP occurs with a verb that does not subcategorize for an object,
then the f-structure is incoherent. If a lexical item like ‘voit’ requires an object, then either the clitic
or the full NP must be present; otherwise the sentence will be incomplete. However, if both are
present with ‘voit’, then the f-structure is inconsistent because the value of the OBJECT’s PRED
would not be unique. Many phenomena for which arguments of functional equivalence and
distributional complementarity have been used to argue for syntactic movement (not only NP-
movement but also V-movement, for example) could be analyzed in similar fashion.
2.4 Semantic Coherence
All semantic forms that are not semantically empty (i.e., that are not dummy elements) must
be linked to the logical argument of another lexical form—in order to be coherently interpreted.
3. Control and Complementation
Among the universal set of grammatical functions are complements and adjuncts.
Complements are an essential part of the argument structure (part of the subcategorization frame),
while adjuncts provide additional information and are interpreted by association with some other
subcategorized argument. Adjuncts are not required for grammaticality, while omission of a
complement results in an ill-formed sentence. Adjuncts have greater mobility than complements
and are often set off by pauses. The following contrasts illustrate this:
Complement
John didn’t sound ashamed of himself.
*John didn’t sound.
*John, ashamed of himself, didn’t sound.
Adjunct
John looked down, ashamed of himself.
John looked down.
John, ashamed of himself, looked down.
Complements and adjuncts may either be closed, i.e., semantically complete, containing within
them all the elements required for logical interpretation of the predicate, or open, lacking a subject

---