The distinction introduced by cruse:86 and described above
between contextual selection of a sense and
contextual modulation of a sense has been reiterated in more
recent work such as copestake_briscoe:95 and
pustejovsky:95a. Contextual selection must occur in the case
of words which have multiple senses represented in the lexicon, such
as those which have two senses related by a regularity which can be
captured by a lexical rule. On the other hand, contextual modulation
occurs in the case of words which have a single sense which can be
contextually specified. For example, the reel in
film reel and fishing reel can be defined as ``a
container artifact with the purpose of (un)winding'' (Copestake and
Briscoe 1995:18), with the modifiers
providing details about what material is (un)wound and thereby making
the sense of reel more specific. As discussed by Copestake
and Briscoe, this modulation involves complex reasoning at the
pragmatic level, but depends on a (default) sense for the word which
is proposed by the lexicon.
The modulation then occurs by specifying
an underspecified representation of the word's meaning or by
overriding default aspects of the proposed sense. For example,
cloud refers to a mass of substance which by default is water
vapour, but it has extended usages in dust cloud and
cloud of smoke which result from overriding the default to
specify that the mass consists of dust and smoke,
respectively, via the modifying phrases. Notice that one would not
want to postulate separate lexical entries for these three senses of
cloud, because the non-default senses can only be triggered
in contexts which explicitly refer to the constituency of the cloud,
and there are unbounded possibilities for this sense broadening.
One could also conceive of a solution to the sense broadening phenomenon for cloud in terms of two lexical entries, one for the standard cloud composed of water vapour, and one which has an underspecified constituency but which requires a complement (adjective or prepositional phrase) which specifies what the cloud is composed of. In the absence of a complement, cloud would always be taken to refer to water vapour. This solution has the problem that the constituency of a cloud can also be specified contextually, such that the word cloud (without a complement) can in context be taken to refer to a cloud made up of something other than water, as in cc9.
A swarm of mosquitoes descended upon Peter's head. The thick cloud obscured his vision and prevented him from sleeping.
If cloud, on the other hand, is taken to be ambiguous between the water vapour and the underspecified senses even without a complement, then (a) there is no account of the preference for the water vapour sense in isolation of a specific context, and (b) the load of determining the specific sense is shifted to pragmatics. The pragmatic component must then for every use of cloud decide which sense is relevant based on the context. This is a much more difficult task for pragmatics to resolve than checking whether a proposed sense (i.e. the default sense) is coherent in the context. Making ``guesses'' on the basis of lexical information about the intended sense should be more efficient than computing the sense from scratch in each case. Furthermore, cloud is often used in contexts in which there is no prior talk of weather or anything which might prime the water vapour sense, yet that sense is (by default) the intended one. The default interpretation of cloud would have to be represented in pragmatics in order to accurately model this, which forces conventional information about the use of individual words, in specific languages, to be brought into the pragmatic component. This is linguistic information, not world knowledge, and specifically lexical semantic information. This fact about the word cloud is unlikely to have exact correspondences in languages other than English, and means that such a pragmatic model would not be reusable cross-linguistically. This is undesirable, for reasons discussed in the introduction of Chapter 2.
Copestake and Briscoe (1995) refer to the relation between different senses of a word which arise due to contextual modulation as one of constructional polysemy -- their claim is that the polysemy arises from the packaging together of several aspects of an entity into a single representation. They argue that acceptable instances of co-predication are evidence that a single sense or lexical structure is available for a word. So the non-zeugmatic acceptability of examples such as cc8 are evidence that the aperture and physical object senses of door and the institution and physical object senses of newspaper are `bundled together' into a single sense (Pustejovsky pustejovsky:95b).
John painted and walked through the door. John used to work for the newspaper that you are reading.
Pustejovsky (1991, pustejovsky:95a) has addressed directly the issue of what the underspecified representation of a word might look like, which I have discussed in various points in this thesis (e.g. Section 2.6), in particular with respect to nouns. As a reminder, the representation he proposes includes specification of various core aspects of the denotation of a noun (form, content, agentive, telic roles). Each of these aspects can be picked out in context, but none of them forms an independent sense of the word. These representations therefore allow the behaviour associated with constructional polysemy to be accounted for directly.
As with lexical rules, the incorporation of structured yet underspecified lexical representations and the use of defaults in NLP systems would greatly enhance the flexibility of the lexicon and would help to achieve context-specific interpretation of words. Acknowledging the ways in which words can interact and providing mechanisms for modeling the interactions will lead to more precise representations of the words in context, and therefore will allow more accurate inferences and responses.
Consider again the example of cloud of mosquitoes. A system which only had the standard water-vapour sense of cloud represented would incorrectly infer from The cloud of mosquitoes descended upon Peter's head that there is water vapour around Peter's head and that he becomes wet, and would probably suffer irresolvable conflict when attempting to establish the role of mosquitoes in the interpretation. Under these circumstances, a text summarisation system might incorrectly inform a user that this text has to do with weather phenomena rather than a mosquito attack (cf. my earlier remarks). This example suggests that NLP systems can benefit from integrating linguistic insights even in cases where limited interpretation is necessary by allowing for interactions between words which shift their meanings.
Furthermore, the structure in lexical representations can help to define the range of shifts which a particular word can undergo and to constrain the ways in which it can interact with other words. Pustejovsky's core aspects of nominals, for example, establishes several dimensions along which nouns can interact with other words (e.g. verbs, adjectives), but excludes others which aren't represented in the lexical entries. Thus, the representations can help not only to increase the generativity of the computational lexicon, but also to constrain the generativity and avoid generation of spurious interpretations.