What emerges very clearly from the recent work on the interface between lexical and non-lexical semantic information is that polysemy is not a single, monolithic phenomenon. Rather, it is the result of both compositional operations in the semantics [...] and of contextual effects, such as the structure of rhetorical relations in discourse and pragmatic constraints on co-reference. What is necessary is for research to tackle the difficult question of how other components in the natural language interpretation process interact with the lexicon to disambiguate and fully determine the semantics of the words in context. This work, emerging from very different traditions, illustrates how lexical semantics can be made sensitive to sentence level compositional processes as well as discourse level inference mechanisms, reacting to the diverse and multiple causes of lexical ambiguity.
(Pustejovsky pustejovsky:95a:236)
Traditionally, representations of meaning have been used for the purpose of modeling inferences which can be made on the basis of the meaning of a phrase or sentence. Logical representations of meaning are rigorously defined and conclusions drawn from them can be proven within the formal framework established for the logic. As a linguistic representation, however, purely logical representations (i.e. representations devoid of syntactic information) are inadequate. Traditional approaches to the problem of how the meaning of a sentence can be built up from the meanings of the components of the sentence (e.g. Montague Grammar) are flawed in their dependency on strict rules of combination and their ignorance of the distinction between syntax and meaning. So, for example, a verb such as drink is semantically bivalent -- it has two semantic arguments, the drinker and what is drunk -- yet can be used in both the transitive (John drinks milk) or intransitive (John drinks) syntactic frames. A semantic grammar which assumes a strict semantics-syntax correspondence cannot account both for the two different syntactic forms in which this verb can appear and for the constancy of its semantic valency across these forms (the intransitive form still suggests that something is drunk). Furthermore, the inference mechanisms associated with classical logical representations are hampered by the fact that inference is not always a matter of drawing logically valid conclusions; it might depend on a particular person's model of the world, the context in which that inference takes place, or on his knowledge. The interpretation of the intransitive John drinks, for example, can depend on the discourse or situational context in which it is uttered, or might be interpreted as John drinks alcohol without any information to the contrary (by default).
On the other hand, a theory of syntax which says nothing about the meaning associated with the syntactic structures entirely misses the point of language -- what are language and its varied structures for if not to convey meaning? Furthermore, such a theory will contribute nothing to the problem of explaining how language is naturally used, and cannot account for the distinction between ungrammatical sentences (e.g. *John ice cream likes not), infelicitous sentences (e.g. the famous #Curious green ideas sleep furiously), and fully grammatical and felicitous ones (e.g. The baby sleeps quietly).
To address such issues, recent work in linguistics and computational linguistics has attempted to construct representations which accommodate both syntactic and semantic information about a word. The representations of the lexicalist (word-driven) grammar framework Head-Driven Phrase Structure Grammar (HPSG) (Pollard and Sag 1994) provide an example of such a representation. These representations reflect an important step in the development of linguistic theory towards investigations of phenomena which lie at the intersection of syntax and semantics. The next step, as suggested in Pustejovsky's quote above and as will be argued in this thesis, is to consider how word-level representations can be used to capture word usage and how these structures might interact with pragmatic reasoning.
In this thesis I will advocate a view of linguistic research which challenges modular investigations of language -- i.e. research concentrating on some aspect of language use (syntax, semantics, discourse structure) in isolation of the others. I will show, through the example of several linguistic phenomena, that a model of language use must take the interactivity of distinct (linguistic and non-linguistic) knowledge sources into account in order to achieve a full explanation of the range of usage. Syntactic structure does not exist in isolation of meaning, while meaning is influenced through context. I will argue that a distinction does exist between syntax, semantics, and pragmatics, and that attempts to explain language understanding in a non-modular framework that conflates linguistic and non-linguistic knowledge (e.g. Hobbs 1993) will not be able to account for conventionalised language use, but that the structure and functions of these modules must be considered in relation to each of the others. We will see in Chapter 5, for example, that the representation assumed for a verb depends on both the syntactic structure in which it is used, the representational structure of its complement, and its behaviour in the discourse context. Thus only through knowledge of a word's use in context -- both the linguistic context (syntactic frame) and the discourse context -- can an appropriate representational structure for it be established.
I am certainly not the first to make arguments for interactivity in linguistic process. A body of research exists which focuses on various aspects of linguistic interactivity: the theory of Conceptual Structures (Jackendoff jack:83,jack:90) investigates regularities in the relationship between meaning and syntax; Construction Grammar (e.g. Fillmore fillmore:88; Goldberg goldberg:95) argues that not all meaning is built up compositionally but that certain syntactic forms are associated with particular interpretations and that these interpretations merge in consistent ways with the meanings of the constituents; the Generative Lexicon (Pustejovsky pustejovsky:91,pustejovsky:95a) is a theory based on the idea that the words are generated through interactions between rich word-level representations; Persistent Default Unification (Lascarides lascarides_etal:96) and related work (e.g. Lascarides and Copestake lasc_copestake:95; Lascarides, Copestake, and Briscoe lasc_cope_brisc:96) looks at the interaction of default interpretations of words with world knowledge. This thesis therefore is a reflection of a more general shift in focus which has recently occurred in linguistics. The work presented here in fact builds on the research cited above, integrating ideas from each theory in order to create a research framework which allows investigation of the structure of word meaning required to explain language use in context.
I will examine phenomena where only an interaction of semantic, syntactic, and pragmatic information can fully explain the range of felicity of the phenomena, and I will argue along the way that structured information about word meaning plays a key role in this interaction. The examination of each phenomenon will result in proposals focusing on different aspects of word-level representation, but each proposal fits into a single framework, outlined in Chapter 2. The proposals are data-driven; reflecting as far as possible the range of usage of the constructions examined. To that end, the thesis contains many example sentences and even some corpus analysis (Chapter 5).