In HPSG, the repository of information about the subcategorisation of signs is the local syntactic feature SUBCAT. The feature SUBCAT is used to encode the various dependencies that hold between a lexical head and its complements (including subjects). SUBCAT takes a list of (partially specified) synsems as its value. Given that synsems bear local values for CATEGORY and CONTENT, lexical signs can exert restrictions on category selection and government as well as case and role assignment.
In HPSG the flow of subcategorisation information up projection paths is handled by a principle of universal grammar -- the Subcategorisation Principle -- expressed in the following terms:
In a headed phrase, the list value of
DAUGHTERS HEAD-DAUGHTER SYNSEM LOCAL CATEGORY SUBCAT
is the concatenation of the list value of
SYNSEM LOCAL CATEGORY SUBCAT
with the list consisting of the SYNSEM values (in order) of the elements of the list value of
Broadly speaking, the Subcategorisation Principle establishes that the SUBCAT value of a phrase is the SUBCAT value of the lexical head minus those specifications already satisfied by some constituent in the phrase. The Subcategorisation Principle, then, works much the same way as cancellation in categorial grammar.
A more recent version of HPSG distinguishes between two features SUBJ and COMPLS, which were conflated in the former feature SUBCAT.
One essential function of the SUBCAT feature is to set up the correspondence between grammatical relations (subject, object, etc.) and the roles that are present in the situation described by the verb. In HPSG, then, role assignment is the connection between the constituents of a utterance and the constituents of the thing the utterance is about. Thus, the lexical entry for a ditransitive verb such as give assigns semantic roles to its subcategorised-for dependents
The variables associated with the elements of the SUBCAT list unify with the corresponding variables of the roles in the CONTENT description. Thus the subject variable (first element of the SUBCAT list) unifies with the variable filling the giver role, the first object variable (second element of the SUBCAT list) unifies with the variable corresponding to the giver rol, and, finally, the second object variable (third element in the SUBCAT list) unifies with the variable corresponding to the given role in the verb's described situation.
The theory of subcategorisation in HPSG makes essential use of a hierarchical conception of grammatical relations. Leaving aside the subject which has its own list feature, other syntactic functions (such as first object, second object,...) are defined in terms of the order of the corresponding elements on the head's COMPLS list. That is, order on this list corresponds to the traditional grammatical notion of obliqueness of grammatical relations, with more oblique elements occurring further to the left.
Four different classes of linguistic generalisations provide motivation for the hierarchical theory of grammatical relations:
In HPSG, lexical dependencies crucially involve category selection. It is well known that there are subcategorisation restrictions which involve differences of syntactic category which cannot be reduced to semantic or functional distinctions. Examples (24) to (27) below show that verbs, which are semantically very close, impose specific restrictions on the syntactic category of their arguments.
|(24)||Sandy trusts Kim.|
|*||Sandy trusts on Kim.|
|(25)||Sandy relies on Kim.|
|*||Sandy relies Kim.|
|(26)||Terry is a complete madman.|
|Terry is out of his mind.|
|(27)||Terry became a complete madman.|
|*||Terry became out of his mind.|
Verbs of trusting such as trust and rely in (24) and (25) may have similar argument structures but display different category selection. In HPSG, category selection is achieved in the SUBCAT list specifications. Thus, in (28), the SUBCAT description for trust specifies that the category of its second complement is N:
|(28)||trust: SUBCAT ..., SYNSEM|LOC|CAT|MAJ NP|
Relay not only specifies the category of its complement as prepositional but also demands the preposition be on, as in (29):
|(29)||rely: SUBCAT ..., SYNSEM|LOC|CAT [MAJ P, PFORM on]|
Similar variation is exhibited for finer syntactic selections. For instance, English verbs (including auxiliaries) differ with respect to the inflectional form required on the verb they subcategorise for, as in (30) and (31):
|(30)||Sandy made Kim throw up.|
|*||Sandy made Kim to throw up.|
|(31)||Sandy forced Kim to throw up.|
|*||Sandy forced Kim throw up.|
In order to account for (30) and (31), the corresponding lexical signs must specify different values for the feature VFORM on the xcomp:
Likewise, in languages which inflect for case, semantically close
verbs may require objects in different cases. Since CASE is a head
feature, it follows (in virtue of the Head Feature Principle, which
percolates the head features from daughter to mother node) that
whenever a lexical form selects a phrasal complement specified as
SYNLOCHEADCASE ACC or
the lexical head of that complement is also specified in this way.
For languages which lack case inflection (like English or Spanish) an analogous situation obtains with respect to government of the particular preposition that heads a PP complement, as in (32):
|(32)||Kim relies/depends on Sandy.|
Semantically related verbs may assign closely corresponding roles to different PP complements, as in (33):
|(33)||The authorities blamed Greenpeace for/*with/*of the bombing.|
|The authorities accused Greenpeace *for/*with/of the bombing.|
|The authorities charged Greenpeace *for/with/*of the bombing.|
These examples show that government of particular prepositions is not semantically predictable, instead, different verbs which take PP complements require different values for the head feature PFORM on that complement.
All equi and raising verbs take a predicative complement the subject of which is identified with a co-argument of the verb. They differ in that all subcategorised dependents of equi verbs are assigned semantic role, while one dependent of raising verbs is not asigned a role. In HPSG, this distinction is collapsed in lexical entries.
An equi verb such as try subcategorises a subject NP and a complement VP. Try assigns the tryer role to the referential index corresponding to its subject and the CONTENT value of its VP complement to the psoa argument. In addition, the index of the subject is in structure-sharing with the subject of the VP complement in the SUBCAT list:
Given that the subject of the infinitive VP complement is also assigned a role in the infinitive CONTENT description, a sentence such as John tries to run would be as follows:
In the case of raising verbs such as tend below, subjects are not assigned a role in the matrix psoa. Here, subject identification is not made by structure-sharing of indexes but by structure-sharing of the whole synsem of the subjects. Thus, the SUBCAT list tend specifies that the synsem of its subject is structure-shared with the synsem of its subcategorised VP complement:
The HPSG treatment of subcategorisation dependencies suggests a uniform treatment of functional, formal and semantic selections, for the values of the feature SUBCAT (or SUBJ and COMPLS) contain all three kinds of information (they are lists of synsems): functional information (e.g. the order of the elements on the COMPLS list); formal information (specifications for values of the attribute CATEG); and semantic information (specifications for values of the attribute CONTENT).