TITLE>Berkeley Construction Grammar
These notes are to bring students up-to-date who missed (or forgot) the decisions made in the 1995 version of the text regarding the feature representations of the basic constructions. The following is an adaptation of the "Embarrassing Digression" which began chapter 5.
In the earliest proposed version of the Determination construction, we gave no particular analysis to the notion "determiner": we simply called the word that in the phrase that person a Det. The most important fact about determiners at that point was that they can combine with something that is itself incapable of being a self-standing nominal argument to form a maximal NP. Our diagram looked like this:
Figure 20, chapter 2
We knew that this wasn't going to be adequate, since the first position of this construction can be filled by articles, demonstratives, possessive NPs, as well as several other sorts of things. The idea of being a determiner, we argued, was more that of a "role" than that of a kind of thing. So in the improved version of the construction in that chapter, we said that the left sister had "role Det", intending to express in that way the insight just mentioned. We assigned no role name to the right sister.
Figure 25, Chapter 2
But we also used the word role in that chapter to name the attribute whose value consisted of a complex consisting of the gf and the theta features. It was awkward, in that discussion, to be using the same word, role, for each of these notions, and we puzzled about what to do.
Then in Chapter 3 we introduced a slightly different way of talking, where (i) the word function (abbreviated fcn) replaced "role" in contexts like the one in Figure 2, (ii) it was suggested that a "function" label should be given to each member of a phrasal construction, not just one, and (iii) the word role was limited to use as the higher attribute for the the gf and q features of valence elements. The first constituent of the Determination construction (Chapter 3, Figure 15) was provided with the feature [fcn det], and the second constituent had [fcn head] (as shown in boldface in Figure 3). But this time, instead of having these fcn features at the highest level in the AVM for the constituent, we (unwisely) subordinated them to the attribute syn.
Figure 15, Chapter 3
We are making several changes in the feature-structure apparatus now, and we would like to think that the new notation, and the re-conceptualization that it represents, will seem more orderly.
First change: we wish to be consistent in giving names to the roles that constituents have within the phrases that contain them, and this role feature will be introduced at the highest level in each daughter AVM. For this we will use the term phrasal role when we need to be careful, but since it will be the only place in the notation itself that uses the word, we will simply write role.
Second change: we wish to make a clear distinction between relational features and inherent features. This is a distinction between what is true about a particular word or phrase, on its own, and what properties it has by virtue of its relation to a predicator whose requirements it satisfies. To present the relational features of arguments, we will now use the word relation (rel) rather than role and will intend the values of this attribute to be feature structures containing gf features and theta features. For grouping the inherent features, we will introduce an attribute name that combines the syntactic and semantic properties, called "syntax/semantics" or "synsem" but abbreviated as ss.
Figure 4 gives a general schema for a two-part phrasal construction. In the Determination construction, the a and the b will be replaced by spec (see below) and head respectively. In the MN construction the role names will be mod and head. If the mother of a construct that satisfies such a construction is itself a constituent of a higher phrase, then it too will have a role feature corresponding to its role in that higher phrase. The point is that it is phrasal constructions that supply phrasal roles.
Schema for roles in
phrasal constructions with 2 daughters
Arguments will have ss features and rel features. That is, constituents which are arguments will carry syntactic information revealing their grammatical type, semantic information, showing their meaning, and relational information indicating the grammatical (gf) and semantic (theta) relations they bear to the predicators whose valence requirements they satisfy. A general schema for arguments is given by Figure 5. If an argument finds itself as a member of a phrasal construction, it will also have a role feature.
Schema for arguments
Predicates are valence-bearing constituents. They will bear syntactic information, indicating their category and semantic information giving their meaning, plus a set of valence requirements. A schema for predicates is given by Figure 6. In this schema, the raised 'plus' sign indicates that the entity which precedes it can occur in one or more repetitions: the valence is presented as a set of one or more arguments. A simple intransitive verb will have just one valence member; a simple transitive verb will have two; a ditransitive will have three; and there are various means of augmenting a valence that will be discussed later on.
Schema for predicates
Predicates have val features; but since predicates can in fact be arguments, they may contain rel features as well; and since either predicates or arguments may be the constituents of phrases, either may contain a role feature too. Consider what there is to say about the phrase read this in sentence (1).
(1) You should read this.
Since the features of ss, role, val and rel are all mutually compatible, they can all - as has just been shown with read this in sentence (1) - appear in a single feature structure.
The newly introduced role feature, then, will introduce the names of the roles that constituents have within the phrases that they take part in. We can distinguish two main types of phrasal constructions, headed and non-headed. For a headed construction, there is one constituent that participates more centrally in determining the character of the whole, and that is the head of the construction. For non-headed constructions (such as the one that gives us the coordinate conjunction John and Mary) no single constituent has such a function.
The Determination construction, as we have seen, has a head and (to its left) what we will now call a specifier. This role will appear in other constructions as well, as you will see later on. Thus, the role features of the two constituents of the phrase this person are [role specifier] and [role head] respectively.
For the MN construction the role names will be modifier and head; the role features of the two constituents of the phrase warm milk will be [role modifier] and [role head] respectively.
For certain phrases we will find that one of the constituents has a simple "marking" function, without any independent accompanying semantic function. An example is the Marked Infinitive construction, which is simply the word to followed by a bare-stem verb phrase. We will say that the two parts of to sleep soundly have the features [role marker] (to) and [role head] (sleep soundly).
Most other uses of [role head] will be to mark the heads of complementation structures, where the head constituent has a valence and the other constituents of the phrase are its complements. Since we have other uses for the word complement, we need a separate term when speaking of the non-head role of a phrasal member whose head has an unsatisfied valence: for this we will use "phrasal complement", abbreviated pcomp.
The head of a preposition phrase is the preposition and its object is its pcomp. The two parts of into the cave, then, are [role head] and [role phrasal-complement]. The head of a verb phrase is the verb, and all of that verb's sisters have the role pcomp. Thus, for a phrase like demonstrated the proof to the class, the verb demonstrated has the feature [role head] and each of the constituents the proof and to the class will have the feature [role pcomp].
There is a third new development in our AVM representations, and that has to do with syn features. Syn features can be divided into head features (the features that are identical between a head daughter and its mother) and what we will call level features. The head features that we will speak of first are category (cat) and lexical head (lexh). For various reasons we sometimes need to be able to refer to headed phrases by their lexical heads, so the lexh of the noun boy will be boy and the lexh of the NP the boy will also be boy. Since a NP is a phrase headed by a noun, and a PP is a phrase headed by a preposition, etc., the cat feature will also belong to the AVM that serves as the value of head.
The level features are those that indicate degrees of "completion" of a phrase of a particular category. The feature lex indicates whether the constituent is a lexical item ([lex +]) or a phrase ([lex -]); the feature max indicates whether the constituent is ([max +]) or is not ([max -]) capable of standing as an argument. In the case of predicating constituents, we will add srs, standing for "subject requirement satisfied". [srs +] will indicate that the clause's subject valence is either locally present or locally interpreted; [srs -] will mean that the subject argument still needs to be provided.
In short, ss features will fit the following schema:
The syn attribute and its values
With all of these changes we have now introduced a new possible source of confusion, in our recycling of the word head. In one use it is the atomic value of an attribute, namely role; but as a part of a syn AVM, it is an attribute which takes as its value an AVM characterizing those properties of a head constituent which are shared by the mother constituent. Using the same word for both purposes is deliberate, of course, but only as a mnemonic: the features that are called head features are the ones that are shared by a phrase and its head daughter, the daughter which bears the notation '[role head]'.
One last change has to be mentioned. Now that we recognize lexh as a "head" feature within syn, we will let that stand for what we earlier called "lexeme", and we will use, at the outermost element of a lexical constituent, an attribute phonology (abbreviated phon) that will stand for the actual occurring form of the word.
(The phonology values, in general, will be determined by reference to the "underlying phonological form" of each lexical item and other relevant information found, for example, in the specification of the inflection of an inflected word. Here we assume this feature only for lexical items, but it seems clear that the theory of phonology that works best with construction grammar will be one which assigns a phonology value to every phrasal constituent, not just lexical ones. Phonological processes - which may or may not be unificational in nature - will relate the phonology of phrases to the phonology of their constituents. Since we will seldom have anything to say about phonological form in this book, we will represent this part of a lexical entries with representations in standard spelling.)