- Sense enumeration: One could explain this data by
assuming that each of the eventive verbs is associated with a list
of senses which corresponds to all the possible meanings that verb
can take on in the logical metonymy construction. So begin
might have senses begin reading, begin eating, begin smoking,
etc. licensed lexically. This approach, however, has several flaws:
- There is nothing in this analysis which explains the
complement-dependent nature of the meaning variation. Thus
beg91a could just as easily be assigned the interpretation
John began smoking/eating/drinking the book as the more
standard John began reading the book. This analysis would
lead to overgeneration of interpretations for individual instances
of the logical metonymy construction. For example, it could result in
the incorrect prediction that the meanings of all the sentences in
beg90 involve a reading event on the basis of the begin
reading sense associated with begin.
-
- John began the coffee. (began drinking)
- John and Mary began Scrabble. (began playing)
- John began the computer program. (began writing/running)
This problem could be avoided to a certain degree through the addition
of selectional restrictions specifying the required semantic type of
the NP complement of begin on each of its senses, but this misses
an important generalisation about the relationship between the sense
of begin and the semantics of the complement NP and why a
particular sense of begin can combine with a particular kind of
complement. For example, coffee is normally drunk, which is why it
combines with the begin drinking sense of begin. This
relationship would not be recorded in such an account.
- It is very inefficient to attempt to list in advance all of the
possible meanings the eventive verbs can take on in every possible
context and in combination with every possible complement NP, and
probably impossible given the creativity of language use. Thus
there would likely be undergeneration of the possible
interpretations theoretically available for the construction.
- This analysis is straightforward at the lexical level, but
increases the load on the pragmatic component of interpretation
dramatically. The problem of identifying the particular sense of
the eventive verb intended in context is not aided by restrictions
at the lexical semantic level, beyond any selectional restrictions
which might be introduced. Selectional restrictions are very rarely
hard and fast (we will see below for example that the begin
eating sense of begin can combine with book under
certain circumstances, which would not be allowed if this sense were
restricted to combining with food-type objects) and unlikely
to be restrictive enough (selectional restrictions will likely only
be restrictions on very general semantic types, so for example the
begin drinking sense might require a complement of type
liquid, but then this would license a sentence like
*John/The car began the petrol or John began the river),
and therefore of questionable use for accurately modeling this
phenomenon. In the absence of any restrictions, a logical metonymy
will be associated with a very long list of possible
interpretations, one of which must be chosen through pragmatic
reasoning (or some other mechanism for word sense disambiguation --
see Section 6.4). The difficulty of this task is proportional
to the number of choices which are available, in that each
alternative must be evaluated and the most context-compatible option
identified, and is thus arduous under the circumstances deriving
from this analysis.
- Meaning postulates: An alternative approach is to
specify meaning postulates which explicitly encode the
interpretation which is to be given to particular logical
metonymies. For example, there might be meaning postulates like
those in beg92.
-
- begin the book
begin reading the
book.
- enjoy the beer
enjoy drinking the
book.
These give precise conversions of the underspecified interpretation of
the logical metonymies into fully specified interpretations. Meaning
postulates, however, suffer from many of the same problems as the
sense enumeration technique.
- Advance specification of meaning postulates means that no
contextual influence on the interpretation of logical
metonymies can be accommodated under this analysis, and that novel
instances of logical metonymies cannot be given any interpretation.
- Some mechanism would again be needed for identifying the precise
interpretation of a logical metonymy in a case where multiple
``translations'' are defined by meaning postulates. The sense
disambiguation problem is much more constrained in this case as
compared to the sense enumeration analysis, as there should only be
a few meaning postulates for any given logical metonymy, but it must
still be addressed.
- In the absence of the use of types within the meaning postulates, there
would be much redundancy, in that the interpretations given to
sets of logical metonymies varying in the eventive verb but not in
the noun phrase complement should involve the same event (e.g.\
begin/finish/enjoy the book = begin/finish/enjoy reading
the book).
- This analysis generally fails to make any generalisations about
how the missing or underspecified event in the logical metonymy
might derive from knowledge of the complement noun phrase.
- Pragmatic extension: It is possible to imagine a
treatment of this data which relies solely on pragmatic reasoning
for resolving the meaning underspecification, along the lines of
Hobbs (1993), which uses inferences
drawn from world knowledge or context to elaborate (make more
explicit) the logical form associated with a particular syntactic
structure. On such an account, logical metonymies would be
interpreted on the basis of knowledge of the object referred to by
the complement NP of the eventive verb. The search for this
knowledge would be triggered by the semantic incompatibility between
the eventive argument the eventive verbs seek and the object denoted
by the NP. The pragmatic component would identify some event which
is highly associated with the object, and suggest that event as the
``missing'' event. For example, reading is a prototypical
event in which a book participates, so through pragmatic
reasoning John began a book is likely to mean John began
reading a book.
The main problem with such a treatment is that not every event which
is commonly associated with an object leads to a felicitous
interpretation of a logical metonymy involving that object. There
seem to be various linguistic constraints governing the phenomenon
which cannot be adequately captured under this analysis. So the
suggested interpretations of the following metonymies do not seem to
be possible, despite the apparent relevance of the event to the
object.
-
- *John began a chair. (began sitting in/on)
- *John began a keyboard. (began typing on)
- *John began the tunnel. (began driving through)
- *John began the trees. (began growing, planting,
watering)
I will discuss the pragmatic solution further in
Section 5.5.4. There we will see that a purely
pragmatic approach would fail to constrain the possible
interpretations of logical metonymies and does not allow for the
incorporation of lexically specified restrictions. It therefore would
fail to account for the full range of data. Consider the data in
beg28a and the discourse in beg30, for example.
-
- John enjoyed the book.
- John enjoyed reading the book.
- John enjoyed writing the book.
- John enjoyed eating the book.
-
- John will be audited by the tax
service, so he has been destroying things which might incriminate
him. He has destroyed the files and the computer disks.
- *He will begin the books tomorrow. (destroying)
- He will begin on the books tomorrow. (destroying)
If we consider the sentence beg28aa independent of a context
we are likely to interpret it as beg28ab or possibly
beg28ac. If we then insert the prior context John is my
pet goat; he loves eating things we suddenly prefer the
interpretation in beg28ad. In this case context is able to
influence the interpretation of the metonymic construction.
In beg30, in contrast, we have two cases in which a metonymic
construction needs to be interpreted in a way which is at
variance with the default interpretation, due to contextual
requirements. The sentence He will begin the books tomorrow
would by default mean He will begin reading the books tomorrow,
while He will begin on the books tomorrow does not have a strong
default interpretation. The context clearly cues the interpretation
He will begin destroying the books tomorrow for both sentences.
The sentence beg30b, however, is infelicitous in the context
while beg30c is felicitous in the discourse and has the
expected interpretation. These contrasts suggest a difference in the
behaviour of begin and that of begin on in metonymic
constructions, and unexpected differences in the way particular
instances of metonymy interact with the discourse context. There is
no way for these differences to be explained on the basis of
pragmatics alone, as the difference between the phrases is largely
syntactic. Later in this chapter I will present a treatment of
lexical differences between begin and begin on which is
able to account for these contrasts.
- Co-composition: Similar to the pragmatic account,
analyses of logical metonymy have been proposed (e.g. Pustejvosky
1991, pustejovsky:95a) which rely on lexical encoding
of ``core aspects'' of the meaning of words, and which locate the
meaning underspecification of logical metonymies at the level of
combination of the eventive verb with the complement noun rather
than at the level of the verb itself. So the individual words in
the logical metonymy do not have underpecified meanings; it is their
combination which requires semantic resolution. The difference
between these accounts and the pragmatic account is that they assume
lexical specification of certain knowledge, which is then available
for interaction with syntactic and/or semantic constraints on the
combination of particular words into particular structures. We will
see that this fact enables a more thorough analysis of the range of
grammatical logical metonymy data.