1.4 V LFG: A V ARIANT LFG F ORMALISM
1.4.2 The Phrase Structure Rules
FI-1
]
[ CAT Y 'homonym 2
FS [ FEATURE1 VALUE1 FEATURE3 VALUE3
]
FI-2
] }
1.4.2 The Phrase Structure Rules
A syntactic rule in the vLFG consists of three parts: rule name, pattern matching, and actions. Like LFG, phrase structure rules are augmented with functional expressions.
12. PSR-1: 'rule name X <↓ FEATURE1> =c VALUE1 'pattern Y: GRAMMATICAL FUNCTION
→ 'action
Z
The pattern matching section specifies what kinds of syntactic elements are needed in order to build a higher category. A constituent may assign a certain grammatical function, indicated by either: or ε followed by a function name;
the former usually assigns a subcategorizable function, while the latter assigns an adjunctive function. We will illustrate these two symbols with their LFG equivalents below. The only required part of the action section is the
specification of the higher category that is to be built. When the higher category is built, in the meantime its corresponding f-structure is also built. In the event that the building of the f-structure fails, the higher category will not be built. Other actions, if any, then perform the specified operations on the f-structure thus far built by the rule. Again, if any of the operation of the actions fails, the higher category still will not be built. We will illustrate the differences between the notation of LFG and vLFG below.
13. LFG:
a. S → NP VP (↑ SUBJ)=↓ ↑=↓
b. VP → V (NP) ↑=↓ (↑ OBJ)=↓
c. NP → (DET) (A) N
↑=↓ ↓ε(↑ ADJ) ↑=↓
14. vLFG:
a. PSR-1:
NP: SUBJ VP → S
b. PSR-2:
V
(NP): OBJ →
VP
LEXICALFUNCTIONAL GRAMMARANDAVARIANTFORMALISM 23
c. PSR-3:
(DET)
(A) ε ADJ
N → NP
Given a lexicalist linguistic theory like LFG, we feel that it is more suitable intuitively to express in a phrase structure rule the constituents first: X Y -> Z, rather than Z -> X Y. The former indicates that when two elements, X and Y, are found, Z is built and that between X and Y the one not assigned with any grammatical function is the head. In terms of both parsing and generation, this scheme comports better with the lexicalist view of grammar where it is the actual lexical items that dictate the structure of a string and activate the parser or generator for processing. The pattern section allows optionality and disjunction, and the specification of partial ordering is also facilitated.
15. a. [ X Y ] `X immediately precedes y b. < X Y Z > `X, Y and Z in any order c. { X Y } `either X or Y, but not both
d. < 1-2; X Y Z > `X, Y and Z in any order except that
`X must precede y; this specifies
`partial ordering
e. (X) `X is optional
f. X+ `one or more X
g. X* `zero or more X; equivalent to (X+)
1.4.3 Unification: from C-structure to F-structure
In conventional LFG, phrase structure rules first build the entire c-structure of the sentence. Unification then proceeds according to the specified functional equations on the c-structure to build a corresponding f-structure (Wescoat 1987). In the vLFG however, as we have just seen from the vLFG rule format, when any legal portion of the c-structure of a sentence is built, it must have simultaneously an f-structure built to correspond to it. In other words, when a phrase structure rule applies to build a higher category, the
functional equations specified by that rule are applied at the same time. And only when the f-structure associated with the c-structure is also sound, can the higher category be built. As explicated by Figure V in Wescoat (1987:10), LFG does allow this alternative. We will give a simple example to illustrate some of the points we have covered thus far, assuming the rules in 12 and 13 above, the following lexicon in vLFG, and the lexicon of 2 in LFG, given the sentence: "John loves Mary."
16. a. Mary:
[ CAT N
FS [ FORM 'Mary' PERSON 3rd NUMBER SG ]
]
b. John:
[ CAT N
FS [ FORM 'John'
PERSON 3rd
NUMBER SG
] ]
c. loves:
[ CAT V
FS [ PRED <SUBJ , OBJ>
FORM 'love'
TENSE PRESENT
SUBJ [ PERSON 3rd
NUMBER SG
] ]
]
LEXICALFUNCTIONAL GRAMMARANDAVARIANTFORMALISM 25
17. LFG:
a. NP `no unification `no f-structure ↑=↓
N John
<↑PRED>='John'
b. NP `no unification `no f-structure ↑=↓
N
Mary <↑PRED>='Mary'
c. VP `no unification `no f-structure ↑=↓ <↑OBJ>=↓
V NP
loves ↑=↓
<↑PRED>= N 'love <(SUBJ)(OBJ)>'
Mary <↑PRED>='Mary'
d. S `unification `f-structure <↑SUBJ>=↓ ↑=↓
NP VP
↑=↓ <↑OBJ>=↓
↑=↓ V NP N
loves ↑=↓
<↑PRED>= N John 'love<(SUBJ)(OBJ)>'
<↑PRED>='John' Mary <↑PRED>='Mary' Only when a c-structure that expands the whole word string is reached does unification proceed to build the f-structure according to the functional annotations on the final c-structure:
18.
PRED 'love <(SUBJ)(OBJ)>' SUBJ [ PRED 'John']
OBJ [ PRED 'Mary']
The vLFG on the other hand builds a c-structure and a corresponding f-structure simultaneously in every step:
LEXICALFUNCTIONAL GRAMMARANDAVARIANTFORMALISM 27
19. vLFG: c-structure f-structure
a. NP [FORM 'John' │ NUMBER SG N PERSON 3rd │ ]
John
b. NP [ FORM 'Mary' │ NUMBER SG N PERSON 3rd │ ]
Mary
c. VP [ PRED <SUBJ , OBJ>
FORM ‘love’
TENSE PRESENT V NP SUBJ [ NUMBER SG PERSON 3rd loves N ]
OBJ [PERSON 3rd Mary NUMBER SG FORM 'Mary'
] ]
d. S [ PRED <SUBJ , OBJ>
FORM 'love'
NP VP SUBJ [ FORM 'John' NUMBER SG V NP PERSON 3rd
N ]
loves OBJ [ FORM 'Mary' John N NUMBER SG
PERSON 3rd
Mary ]
TENSE PRESENT ]
The advantage of vLFG in this respect is that violations of grammatical constraints in the f-structure can be detected earlier in the unification process and thus paths with f-structure violations will not be pursued any further. Any constituent formed would have to have a well-formed f-structure as well. LFG, on the other hand, would pursue all paths of c-structures that the PSR's allow and violation of f-structures can only be detected when the final c-structure is reached because only then does unification proceed to build the f-structure.
This difference in the timing of the building of f-structure thus has rather significant implications on computational efficiency as well as the psycholinguistic processing of natural languages. The vLFG approach in this respect resembles a dependency grammar such as Lexicase in that the constituent structure of a word string cannot exist independently from the participating words in that string. The dependency requirements and restrictions are observed in all steps of constructing the constituent structure of a sentence. The conventional LFG approach, however, in constructing the c-structure is only concerned with the requirements and restrictions, in terms of categorial features, specified by the PSR's.
29
CHAPTER 2
GRAMMATICAL FUNCTIONS IN AN LFG GRAMMAR OF MANDARIN CHINESE
In this chapter, we will first discuss the place of grammatical functions in LFG and the relation between thematic roles and grammatical functions. We will present a classification of grammatical functions in a grammar of Chinese, in terms of their subcategorizability and semantic restrictions. More specifically, we will discuss the notion of subcategorization and each grammatical function in the context of LFG as well as the grammar of Mandarin Chinese. Due to the especially confusing status of topic and subject in the literature of Chinese linguistics, we will devote a more lengthy discussion to clarify their use. Furthermore, we will argue for our treatment of TOPIC as non-subcategorizable grammatical function in the grammar of Chinese. We will also examine the semantic restrictions on OBJ2 in the context of ditransitive verbs and its syntactic encodings in Chinese.
2.1 Grammatical Functions and F-structure
Grammatical relations, such as subjects, objects, and modifiers, are known as "grammatical functions" in LFG. For a sentence to be well-formed it must have a well-formed c-structure as well as a well-formed functional structure.
A functional structure of a sentence contains the functional information of a sentence, and thus is the representation of the hierarchies or internal relationships of grammatical functions among all the components of the sentence. To illustrate, wo3 'I' functions grammatically as the subject of the verb hao3 'fine' in the sentence wo3 hao3 'I am fine', which has the following f-structure in LFG.
1. [ PRED 'hao3 <(SUBJ)>' SUBJ [ PRED 'wo3' ] ]
Functional structures can also be viewed as functions from attributes to values in a mathematical sense. For example, the f-structure of the above example sentence wo3 hao3 can be understood as a function and when we apply this function to the attribute subject we get the value wo3. Thus, the
"functional" part of the name of the LFG theory has both grammatical as well as mathematical significance.
2.2 Grammatical Functions and Thematic Roles
In the theory of Lexical Functional Grammar, grammatical relations are lexically encoded by mapping thematic roles in the predicate argument structure (PAS) to grammatical functions. The mapping process between PAS and grammatical functions is one of the foci of recent LFG developments known as the Lexical Mapping Theory, highlighted by Bresnan and Kanerva (1989). Positions in the PAS are identified with individual thematic roles, such as "agent," "theme," "location," etc. The Lexical Mapping Theory associates each thematic role with a subcategorized grammatical function, such as SUBJ, OBJ, SCOMP, etc. The Function-Argument Biuniqueness Principle ensures that a unique function is mapped with each thematic role and a unique thematic role to each function associated with the PAS.
However, the relation between thematic roles and grammatical functions may not always be one-to-one because of the possibility of non-thematic grammatical functions, e.g., the raised subject in "He seems to be sick" or the object of the idiom string "kick the bucket." In addition, lexical rules, such as Passivization and Locative Inversion, in Lexical Mapping Theory may change the thematic roles of grammatical functions. Most recently, Lexical Mapping Theory has been preliminarily and rather sketchily applied to Mandarin Chinese (Huang 1989). It is certainly an area worth much further exploration.
Once again, the notion of correspondence is important here: Lexical Mapping Theory assigns the thematic structure to the predicate structure of grammatical functions, or the structure of subcategorization. The theory of LFG, with this recent addition of Lexical Mapping Theory, therefore can be
GRAMMATICALFUNCTIONSANDANLFGGRAMMAROFMANDARINCHINESE 31
Lexical Mapping Theory subcategorization structure of grammatical functions
further extended to model the mapping relations from conceptual structures to linguistic representations (Huang 1989), as shown below.
Figure 2.1 Correspondence between Conceptual Structure and Linguistic Structures
In this study, we are primarily interested in the possible patterns of grammatical functions allowable in the predicate structure of Mandarin verbs.
We will have little to say about their semantic/thematic counterparts and the mapping process between them. Since grammatical functions have little intrinsic semantic significance, our focus is on the surface syntactic structure required by the verbs. Even so, due to the intimate relation between thematic roles and grammatical functions, often semantic criteria are used to test whether a certain element is a subcategorized argument of the verb, as we shall see in later sections.
2.3 Syntactic Encoding of Grammatical Functions
In LFG, grammatical functions are considered universal and they constitute a small, finite set. However, these universal functions are encoded in the syntactic categories in various manners in different languages. For instance, in
Extra-linguistic process
functionally augmented PSR's unification
a so-called configurational language like English or Chinese, grammatical functions are assigned to c-structure positions, while in a non-configurational language like Japanese or Malayalam, grammatical functions are assigned to case features. In other words, the correspondence between grammatical functions and syntactic categories is very language-dependent and needs to be specified in the grammar of any particular language. The correspondence between syntactic categories and grammatical functions again need not to be one-to-one. More specifically, the same grammatical function may be assigned by different c-structure positions, and in turn the same c-structure position may assign different grammatical functions. However, the Direct Syntactic Encoding Principle in the universal grammar requires that all non-lexical rules preserve function assignment; that is, syntactic rules can never delete any function or replace one function name with another. This principle ensures that the syntactic encoding of grammatical functions applies directly to surface structures without the mediation of any syntactic or functional derivation. The following figure illustrates the mapping from PAS to the surface syntactic structure, given the sentence Li3si4 da3 Zhang1san1 'Lisi hit Zhangsan'.
da3 ( agent , patient ) predicate argument structure
↓ ↓ assignment of functions
SUBJ OBJ via the Lexical Mapping Theory [ PRED < SUBJ , OBJ > subcategorization of da3 FORM 'da3'
]
GRAMMATICALFUNCTIONSANDANLFGGRAMMAROFMANDARINCHINESE 33
S surface phrase structure
surface grammatical functions
NP(SUBJ) VP
N V NP(OBJ)
N Li3si4 da3 Zhang1san1
Figure 2.2 Lexical and Syntactic Encoding of Grammatical Functions
2.4 Subcategorization in LFG
In the literature of linguistics, it is well-accepted that the term
"subcategorization" refers to the syntactic nature of the elements which a lexical item is required to combine with (Pollard and Sag 1987:113). Thus, subcategorization is taken to be a syntactic notion. Various members of the syntactic category of verb are most commonly classified into different subcategories according to the number and nature of their required dependent, or "subcategorized," elements, although subcategorization restrictions are not limited to verbs. In LFG, the dependency relationships between verbs and their subcategorized elements are characterized in terms of grammatical functions, such as SUBJ, OBJ, XCOMP, etc., not in terms of categories, such as NP, VP, S, etc. The subcategorization of a verb is specified as part of PRED's value in its lexical entry. For instance, in the following entries, da3 'hit', a verb, subcategorizes for SUBJ and OBJ, while wo3 'I', a noun, does not subcategorize for any grammatical function.
2. LFG:
a. da3 V, `hit (↑ PRED) = 'da3 <(SUBJ) (OBJ)>'
b. wo3 N, `I
(↑ PRED) = 'wo3'
Thus, in LFG, PRED contains a special value which is composed of the semantic form of the lexical item and also contains its subcategorization, if there is any. It is important to point out that the PRED feature in LFG is divided into two features in the vLFG: FORM, for the semantic form of the word, and PRED, for the predicate structure, the subcategorization. PRED therefore takes a list of grammatical functions as its value, since in LFG subcategorization is specified in terms of grammatical functions rather than categories. The PRED feature in our vLFG is therefore rather similar to the SUBCAT of syntax in the Head-driven Phrase Structure Grammar (HPSG) and the vLFG's FORM value is similar to HPSG's RELN (semantic relation) (Pollard and Sag 1987). We illustrate with an example:
3. vLFG:
a. da3: 'hit
[ CAT V
FS [ PRED < SUBJ , OBJ >
FORM 'da3' ]
]
b. wo3: 'I
[ CAT N
FS [ FORM 'wo3' ]
]
Therefore, while in LFG a lexical entry with a PRED value is known as a
"semantic form" and a semantic form that subcategorizes for grammatical functions is known as a "lexical form," in vLFG a semantic form is a lexical
GRAMMATICALFUNCTIONSANDANLFGGRAMMAROFMANDARINCHINESE 35
entry with the FORM value, and a lexical form is one with a FORM as well as PRED value.
The concept of subcategorization for grammatical functions, rather than syntactic categories, allows certain category-independence in grammatical processes and generalizations. Grimshaw (1982) and Sells (1985:157) provide several arguments for the functional basis of subcategorization. Lexicase takes a similar position, where subcategorization is stipulated in terms of required and implied case roles, not surface syntactic categories such as NP, VP, and S.
It should also be pointed out that our use of the terms "subcategorize" and
"subcategorization" is actually theory-dependent, in that in LFG their use has a somewhat different meaning from its use in the GB theory. In LFG verb subcategorization is stipulated in terms of grammatical functions and all required functions, including non-thematic ones, are taken to be
"subcategorized" (Bresnan 1982:288-292).
In GB theory, "subcategorize" is used in a different sense. GB recognizes only "strict subcategorization," whose features must be "strictly local" in the deep structure, a constituent structure. Thus, subject, which is not a constituent dominated by VP, according to the PSR, S → NP VP, is regarded as "outer argument" and thus non-subcategorizable; only "inner arguments"
can be subcategorized. As Starosta (1986) has pointed out in greater detail, this GB division of two types of grammatical arguments, which is a direct result of the recognition of a VP category and the stipulation of an INFL and ARG category, is an unnecessary complication in an otherwise straightforward situation and has missed several otherwise statable generalizations of grammatical arguments.
Similar to the position taken in the theory of HPSG where subjects are treated as subcategorized-for, in LFG, although VP is still a recognized category, the strict locality of subcategorized functions is required in the f-structure, not the c-structure, of the verb that subcategorizes for them. Since the verb is the lexical head of the f-structure of its clause, which, unlike the deep structure VP of GB, does contain the function SUBJ locally, subject is subcategorizable (Ford, Bresnan and Kaplan 1982:773). Actually, in LFG, SUBJ is the function that is required in every verb subcategorization, according to the universal subject constraint in Lexical Mapping Theory (Bresnan and Kanerva 1989). A subcategorization without the SUBJ function is considered ill-formed. In this study of Mandarin verb subcategorization, we
also adopt this position, assuming that the absence of surface subjects in some Chinese sentences is due to discourse factors and thus they are discoursally recoverable.
2.5 Classification of Grammatical Functions
Grammatical functions fall into two different classes: subcategorizable functions and non-subcategorizable functions, and subcategorizable functions can be divided into two types: semantically restricted and semantically unrestricted. Note that in the LFG literature, "subcategorizable" and
"non-subcategorizable" are also referred to as "governable" and
"non-governable" respectively.
2.5.1 Subcategorizability
A grammatical function is subcategorizable if there is at least one lexical item that subcategorizes for it. A grammatical function is considered subcategorized by a certain lexical item when it appears in the PRED's value of that lexical item. In the following example of the verb "seem" in sentences like "He seems to be sick," two functions are subcategorized by it, SUBJ and XCOMP.
4. seem V, `LFG
(↑ PRED) = 'seem <(↑XCOMP)> (↑SUBJ)'
Note that here SUBJ is also an example of a subcategorized, and yet non-thematic, argument indicated by the notation that it appears outside of the angle-brackets. Non-thematic arguments are subcategorized grammatical functions that are not linked with thematic roles in the PAS. The verb "seem"
in this particular construction requires only one thematic role; however, in terms of grammatical functions, it requires two. The following notation in our vLFG formalism illustrates the same fact. A function in the PRED's value followed by an @ sign is a non-thematic argument.
GRAMMATICALFUNCTIONSANDANLFGGRAMMAROFMANDARINCHINESE 37
5. seem:: `vLFG
[ CAT V
FS [ FORM 'seem'
PRED < XCOMP , SUBJ @ >
] ]
Thus, to put it simply, a function is subcategorizable if it appears within the PRED's value of at least one lexical item. In other words, only subcategorizable functions may appear within the PRED's value of a lexical form, a form with a PRED attribute, and may receive assignment form thematic roles in the PAS. A non-subcategorizable function thus may never appear within the PRED's value and may never receive a thematic assignment.
For Chinese verbs, we identify SUBJ, OBJ, OBJ2, OBLΘ (oblique function which includes subtypes OBLTHME (theme), OBLGOAL (goal), OBLBNFC (beneficiary), OBLLOCT (location)), and COMP (complement function that includes subtypes XCOMP, SCOMP, and NCOMP) as subcategorizable grammatical functions. TOPIC, ADJUNCTS (adjunctive function that has two subtypes ADJ and XADJ), and POSS are non-subcategorizable.
Sells (1985:155-156) suggests that there is some evidence in English for subcategorizable POSS (for instance, "the professor's knowledge" but not
Sells (1985:155-156) suggests that there is some evidence in English for subcategorizable POSS (for instance, "the professor's knowledge" but not