16.3.3 Subcategorization
Recall that subcategorization is the notion that verbs can be picky about the patterns of arguments they will allow themselves to appear with. In Ch. 12, to prevent the generation of ungrammatical sentences with verbs and verb phrases that do not match, we were forced to split the category of verb into multiple sub-categories. These more specific verb categories were then used in the definition of the specific verb phrases that they were allowed to occur with, as in the following.
$$ Verb-with-S-comp\to think $$
$$ VP\ \to\ \operatorname{Verb-with-S-comp}\ S $$
Clearly, this approach introduces exactly the same undesirable proliferation of categories that we saw with the similar approach to solving the number problem. The proper way to avoid this proliferation is to introduce feature structures to distinguish among the various members of the verb category. This goal can be accomplished by associating an atomic feature called SUBCAT, with an appropriate value, with each of the verbs in the lexicon. For example, the transitive version of serves could be assigned the following feature structure in the lexicon.
$$ Verb\ \to\ s e r v e s $$
$$ \langle V e r b\;H E A D\;A G R E E M E N T\;N U M B E R\rangle=s g $$
$$ \langle V e r b\;H E A D\;S U B C A T\rangle=t r a n s $$
The SUBCAT feature is a signal to the rest of the grammar that this verb should only appear in verb phrases with a single noun phrase argument. This constraint is enforced by adding corresponding constraints to all the verb phrase rules in the grammar, as in the following.
$$ VP\to Verb $$
$$ \langle V P\mathrm{H E A D}\rangle=\langle V e r b\mathrm{H E A D}\rangle $$
$$ \langle V P\mathrm{~H E A D~S U B C A T}\rangle=i n t r a n s $$
$$ VP\to Verb NP $$
$$ \langle V P\mathrm{H E A D}\rangle=\langle V e r b\mathrm{H E A D}\rangle $$
$$ \left\langle V P\ \mathrm{H E A D}\ \mathrm{S U B C A T}\right\rangle=\mathit{t r a n s} $$
$$ VP\ \to\ \textit{Verb NP NP} $$
$$ \langle V P\mathrm{H E A D}\rangle=\langle V e r b\mathrm{H E A D}\rangle $$
$$ \langle V P{\mathrm{~H E A D~S U B C A T}}\rangle=d i t r a n s $$
The first unification constraint in these rules states that the verb phrase receives its HEAD features from its verb constituent, while the second constraint specifies what the value of that SUBCAT feature must be. Any attempt to use a verb with an inappropriate verb phrase will fail since the value of the SUBCAT feature of the VP will fail to unify with the atomic symbol given in the second constraint. Note that this approach requires unique symbols for each of the 50–100 verb phrase frames in English.
This is a somewhat clumsy approach since these unanalyzable SUBCAT symbols do not directly encode either the number or type of the arguments that the verb expects to take. To see this, note that one can not simply examine a verb's entry in the lexicon and know what its subcategorization frame is. Rather, you must use the value of the SUBCAT feature indirectly as a pointer to those verb phrase rules in the grammar that can accept the verb in question.
A more elegant solution, which makes better use of the expressive power of feature structures, allows the verb entries to directly specify the order and category type of the arguments they require. The following entry for serves is an example of one such approach, in which the verb's subcategory feature expresses a list of its objects and complements.
$$ Verb\ \to\ s e r v e s $$
$$ \langle V e r b\;H E A D\;A G R E E M E N T\;N U M B E R\rangle=s g $$
$$ \langle V e r b\;H E A D\;S U B C A T\;F I R S T\;C A T\rangle=N P $$
$$ \langle V e r b\;H E A D\;S U B C A T\;S E C O N D\rangle=e n d $$
This entry uses the FIRST feature to state that the first post-verbal argument must be an NP; the value of the SECOND feature indicates that this verb expects only one argument. A verb like leave Boston in the morning, with two arguments, would have the following kind of entry.
$$ V e r b\to l e a v e s $$
$$ \langle V e r b\;H E A D\;A G R E E M E N T\;N U M B E R\rangle=s g $$
$$ \langle V e r b\;H E A D\;S U B C A T\;F I R S T\;C A T\rangle=N P $$
$$ \langle V e r b\;H E A D\;S U B C A T\;S E C O N D\;C A T\rangle=P P $$
$$ \langle V e r b\;H E A D\;S U B C A T\;T H I R D\rangle=e n d $$
This scheme is, of course, a rather baroque way of encoding a list; it is also possible to use the idea of types defined in Sec. 16.6 to define a list type more cleanly.
The individual verb phrase rules must now check for the presence of exactly the elements specified by their verb, as in the following transitive rule.
$$ VP\to Verb NP $$
$$ \langle V P\mathrm{H E A D}\rangle=\langle V e r b\mathrm{H E A D}\rangle $$
$$ \langle V P~H E A D~S U B C A T~F I R S T~C A T~\rangle=\langle N P~C A T~\rangle $$
$$ \langle V P{\mathrm{~H E A D~S U B C A T~S E C O N D}}\rangle=e n d $$
The second constraint in this rule's constraints states that the category of the first element of the verb's SUBCAT list must match the category of the constituent immediately following the verb. The third constraint goes on to state that this verb phrase rule expects only a single argument.
Our previous examples have shown rather simple subcategorization structures for verbs. In fact, verbs can subcategorize for quite complex subcategorization frames, (e.g., NP PP, NP NP, or NP S) and these frames can be composed of many different phrasal types. In order to come up with a list of possible subcategorization frames for English verbs, we first need to have a list of possible phrase types that can make up these frames. Fig. 16.3 shows one short list of possible phrase types for making up subcategorization frames for verbs; this list is modified from one used to create verb subcategorization frames in the FrameNet project (Johnson, 1999; Baker et al., 1998), and includes phrase types for special subjects of verbs like there and it, as well as for objects and complements.
To use the phrase types in Fig. 16.3 in a unification grammar, each phrase type could be described using features. For example the form VPto, which is subcategorized for by want might be expressed as:
$$ Verb\to want $$
$$ \langle V e r b\;H E A D\;S U B C A T\;F I R S T\;C A T\rangle=V P $$
$$ \langle V e r b\;H E A D\;S U B C A T\;F I R S T\;F O R M\rangle=i n f i n i t i v e $$
Each of the 50 to 100 possible verb subcategorization frames in English would be described as a list drawn from these phrase types. For example, here is an example of the two-complement want. We can use this example to demonstrate two different notational possibilities. First, lists can be represented via an angle brackets notation $ \langle $ and $ \rangle $. Second, instead of using a rewrite-rule annotated with path equations, we can represent the lexical entry as a single feature structure:
$$ \begin{aligned}&\begin{bmatrix}ORTH&want\\CAT&Verb\end{bmatrix}\\ &\text{HEAD}\quad\begin{bmatrix}\text{SUBCAT}&\langle\begin{bmatrix}\text{CAT}NP\end{bmatrix},\begin{bmatrix}\text{CAT}VP\\ \text{HEAD}\begin{bmatrix}\text{VFORM}infinitival\end{bmatrix}\end{bmatrix}\rangle\end{bmatrix}\\ \end{aligned} $$
| Noun Phrase Types | ||
| There It NP | nonreferential there nonreferential it noun phrase | There is still much to learn\nIt was evident that my ideas\nAs he was relating his story |
| Preposition Phrase Types | ||
| PP PPing PPpart | preposition phrase gerundive PP particle | couch their message in terms censured him for not having intervened turn it off |
| Verb Phrase Types | ||
| VPbrst VPto VPwh VPing | bare stem VP to-marked infin. VP wh-VP gerundive VP | she could discuss it\nWhy do you want to know?\nit is worth considering how to write\nI would consider using it |
| Complement Clause types | ||
| Finite Clause | ||
| Sfin Swh Sif | finite clause wh-clause whether/if clause | maintain that the situation was unsatisfactory it tells us where we are ask whether Aristophanes is depicting a |
| Nonfinite Clause | ||
| Sing Sto Sforto Sbrst | gerundive clause to-marked clause for-to clause bare stem clause | see some attention being given know themselves to be relatively unhealthy She was waiting for him to make some reply commanded that his sermons be published |
| Other Types | ||
| AjP Quo | adjective phrase quotes | thought it possible asked “What was it like?” |
Combining even a limited set of phrase types results in a very large set of possible subcategorization frames. Furthermore, each verb allows many different subcategorization frames. For example, here are just some of the subcategorization patterns for the verb ask, with examples from the BNC:
| Subcat | Example |
| --- | --- |
| Quo | asked [ $ Quo $ “What was it like?”] |
| NP | asking [NP a question] |
| Swh | asked [ $ Swh $ what trades you’re interested in] |
| Sto | ask [ $ Sto $ him to tell you] |
| PP | that means asking [PP at home] |
| Vto | asked [ $ Vto $ to see a girl called Evelyn] |
| NP Sif | asked [NP him] [ $ Sif $ whether he could make] |
| NP NP | asked [NP myself] [NP a question] |
| NP Swh | asked [NP him] [ $ Swh $ why he took time off] |
ALTERNATIONS
A number of comprehensive subcategorization-frame tagsets exist, such as the COMLEX set (Macleod et al., 1998), which includes subcategorization frames for verbs, adjectives, and nouns, and the ACQUILEX tagset of verb subcategorization frames (Sanfilippo, 1993). Many subcategorization-frame tagsets add other information about the complements, such as specifying the identity of the implicit subject in a lower verb phrase that has no overt subject; this is called control information. For example, Tommy promised Ruth to go (at least in some dialects) implies that Tommy will do the going, while Tommy persuaded Ruth to go implies that Ruth will do the going. Some of the multiple possible subcategorization frames for a verb can be partially predicted by the semantics of the verb; for example many verbs of transfer (like give, send, carry) predictably take the two subcategorization frames NP NP and NP PP:
NP NP sent FAA Administrator James Busey a letter
NP PP sent a letter to the chairman of the Armed Services Committee
These relationships between subcategorization frames across classes of verbs are called argument-structure alternations, and will be discussed in Ch. 19 when we discuss the semantics of verbal argument structure. Ch. 14 will introduce probabilities for modeling the fact that verbs generally have preferences even among the different subcategorization frames they allow.
Subcategorization in Other Parts of Speech
VALENCE
Although the notion of subcategorization, or valence as it is often called, was originally designed for verbs, more recent work has focused on the fact that many other kinds of words exhibit forms of valence-like behavior. Consider the following contrasting uses of the prepositions while and during.
(16.15) Keep your seatbelt fastened while we are taking off.
(16.16) *Keep your seatbelt fastened while takeoff.
(16.17) Keep your seatbelt fastened during takeoff.
(16.18) *Keep your seatbelt fastened during we are taking off.
Despite the apparent similarities between these words, they make quite different demands on their arguments. Representing these differences is left as Exercise 16.5 for the reader.
Many adjectives and nouns also have subcategorization frames. Here are some examples using the adjectives apparent, aware, and unimportant and the nouns assumption and question:
It was apparent [Sfin that the kitchen was the only room...]
It was apparent [pp from the way she rested her hand over his]
aware [Sfin he may have caused offense]
it is unimportant [Swheth whether only a little bit is accepted]
the assumption [Sfin that wasteful methods have been employed]
the question [Swheth whether the authorities might have decided]
See Macleod et al. (1998) and Johnson (1999) for descriptions of subcategorization frames for nouns and adjectives.
Verbs express subcategorization constraints on their subjects as well as their complements. For example, we need to represent the lexical fact that the verb seem can take an Sfin as its subject (That she was affected seems obvious), while the verb paint cannot. The SUBJECT feature can be used to express these constraints.