Chapter 2 discusses issues of lexical semantic representation, arguing for a certain amount of decomposition in lexical representation in order to efficiently capture generalisations about the behaviour of semantically and/or syntactically related words. In this chapter I argue for a distinction between linguistic knowledge and world knowledge, and suggest that the ontology underlying lexical representation should reflect word-specific linguistic knowledge, including information about conventional usages and default interpretations.
I propose a lexical representation for which the level of decomposition stems from insights by Jackendoff jack:83,jack:90 and Pustejovsky pustejovsky:91,pustejovsky:95a, and for which the formal details derive from proposals by davis:95. The representation integrates with the framework of HPSG, overriding the standard representation of semantic relations in that theory with one which captures relationships between word meanings in a multiple-inheritance hierarchy.
Chapter 3 considers prepositional phrases (PPs) and variations in their contribution to the meaning of the sentences they appear in. This chapter discusses the status of various PPs, focusing on dative PPs such as those in intro5 in contrast to adjunct PPs such as those in intro6, and investigates how to account for the different ways a particular PP can interact with the meaning expressed in the modified main clause. An example of this variation is the contrast in the interpretation of the PP for Mary in intro5b and intro6c. In the former example the PP specifies someone who benefits from an object (a cake), and in the latter it specifies the beneficiary of the full event specified in the main clause (the jogging a mile event).
John gave a book to Mary. John baked a cake for Mary.
John jogged in the park yesterday. John jogged for twenty minutes twice a day. John jogged a mile for Mary.
I review proposals by Kasper kasper:93 and van Noord and Bouma vannoord_bouma:94 for adjunct incorporation in HPSG, and present an analysis of the contribution of the PPs which depends on the lexical semantic structure of both the modified verbs and the modifying PPs. The analysis is justified in terms of the syntactic and semantic properties of the PP data. The proposals here will also provide the basis for accounting for the sense variation of manner of motion verbs in the following chapter (compare intro6c with intro7).
John jogged a mile to the store.
Chapter 4 investigates manner of motion verbs which can acquire a directed motion interpretation, shown in intro4, comparing and contrasting them with constructions known as resultatives, as exemplified by the sentences in intro3.
The horse jumped. The horse jumped over the fence.
John hammered the metal flat. The river froze solid. John sneezed the tissue off the table. John laughed himself silly.
These constructions are associated with a resultative interpretation which is not inherent in the meaning of the main verb (so intro3a, for example, means John hammered the metal and as a result the metal became flat), and often violate normal subcategorisation requirements of the main verb (contrast intro3c and intro3d with *John sneezed the tissue and *John laughed himself).
The two constructions exemplified in intro4-intro3 have in the past been modeled with a single analysis (e.g. Levin and Rappaport Hovav levin_rh:95). I will argue on the basis of semantic intuitions and contextualised data that these constructions are different, and that their behaviour must be explained in distinct ways. I will show how the manner of motion verbs acquire their extended sense due to semantic properties encoded in their lexical representations, and how the resultative construction can be explained in terms of a combination of construction-specific semantics, conventionalisation, and contextual constraints.
Chapter 5 considers the phenomenon of logical metonymy, constructions in which a word (verbs in the following examples) requiring an eventive complement appears with an object complement. In these cases the event which fills the relevant semantic argument position of the word is not explicit in the sentence and must be inferred. For the sentences intro1-intro2, for example, the (a) sentences can be interpreted as having a meaning identical to that explicitly expressed in the (b) sentences.
John began the book. John began reading/to read the book.
John enjoyed the book. John enjoyed reading the book.
The chapter discusses how that interpretation process might proceed, and argues that this phenomenon is governed by lexically specified usage conventions. The content of individual lexical entries is shown to depend on linguistic knowledge, including discourse-level knowledge, particular to the words they represent.
Chapter 6 discusses the implications of the contextual dependency of semantic interpretation for lexicon design and computational processing in Natural Language Processing (NLP) systems. It focuses on the problem of polysemy, and the issues of lexical representation and lexical acquisition for NLP systems which stem from the sense variation which words can exhibit. I will discuss automatic lexicon acquisition techniques from machine-readable dictionaries and corpora, and will argue that this acquisition must necessarily be guided by a theoretical linguistic framework.
Chapter 7 brings the thesis to a close, highlighting the main conclusions of the thesis and suggesting a few areas of future investigation. The broad result which can be drawn from the analyses in this thesis is that syntactic and semantic phenomena cannot be considered in isolation of one another, or in isolation of the context in which they occur. An account of linguistic phenomena which aims to accurately model the full range of usage -- from ungrammatical and infelicitous constructions to grammatical and fully felicitous constructions -- must consider the interaction of the distinct knowledge sources contributing to the determination of the meaning of a sentence in context.