In this thesis I have investigated how lexical semantic structure can be utilised in a contextually-dependent interpretation framework, to handle linguistic phenomena which involve inference of meaning which does not appear to be explicitly present. I have looked at how the meaning of certain prepositions and verbs can vary in particular contexts, and how information represented in the lexicon can lead to models of this variation. I have argued that such models are important for the development of flexible natural language understanding systems, and I have hinted at how that development can be aided by lexical semantic research. To conclude, I would like to highlight the core themes developed in this thesis and provide an overall picture of the resulting framework.
I have proposed a lexical semantic representation which is typed and utilises multiple inheritance, incorporates limited semantic decomposition derived from linguistic research at the syntax/semantics interface (e.g. Jackendoff jack:83,jack:90; Davis davis:95), makes use of non-monotonicity to a limited degree to allow for generalisations to be stated despite the existence of exceptions, and utilises a consistent and constrained structure defined by the lexical hierarchy in which the features are associated with particular entailments. The use of types and multiple inheritance allows the efficient representation of information in common to related words, by minimising redundancy. I have concentrated on verb semantics, but have also adopted the general framework for nominal representation as found in Pustejovsky pustejovsky:91,pustejovsky:95a, to the extent that the information encoded there can help to explain the use of certain nouns in particular phenomena. The representation fits in with the spirit of the constraint-based grammatical framework HPSG (Pollard and Sag pollard_sag:94), although some of the details of that framework have been altered to accommodate more interaction between syntax and semantics at the sentence level. It could fairly straightforwardly be implemented in a feature-structure formalism such as ALE, the attribute logic engine (Carpenter carpenter:93), and integrated with the parsing mechanisms provided there, although the default feature structures may pose a challenge to that implementation. The result would be a system which would generate representations reflecting the interpretation of a sentence as results from syntactic and lexical semantic processing. These representations might include some default elements, or some underspecification, which would need to be resolved with respect to the discourse context.
The thesis is essentially an extended argument in favor of the interaction of distinct knowledge sources in language interpretation, and a view of the lexicon as repository of specifically linguistic knowledge about a word which can help to account for any idiosyncrasies in its use, and the way that word is used in context. I have provided many arguments for maintaining a distinction between world knowledge and linguistic knowledge, but have suggested ways in which the result of linguistic processing might interact with pragmatic processing.
The use of a rich lexical semantic structure in the development of accounts of several phenomena has been shown to be critical for explaining the range of the data and why particular combinations of words or particular interpretations of a certain combination are impossible or infelicitous in context. We saw in Chapter 5, for example, that certain linguistic phenomena like logical metonymy are governed by conventionality, even though the process of interpreting logical metonymy boils down to inferring content that's not explicitly stated and which appears to stem from world knowledge. That is to say that the interpretation of certain sentences and the ability of certain words to appear in particular constructions cannot be predicted from world knowledge or even contextual influences, but depends on specifically linguistic knowledge about a word which must be lexically captured.
In Chapter 3, we saw that the type system could be constructed in such a way as to allow potentially ambiguous prepositions to only be represented once in the lexicon. The ambiguity therefore did not derive from lexical ambiguity, but rather from lexical underspecification of the kind of entity the preposition could modify and the fact that the semantic contribution of that preposition could be made in several different ways. Such techniques for lexicon construction capture generalisations, since several prepositions might be ambiguous in precisely the same ways. The generalisations are represented implicitly in the inheritance relations between types, and explicitly in the existence of different rules which govern the interaction of elements of a particular type with other elements in a sentence. The combination of underspecified types with which words are associated and rules which effectively specify (or precisify, in Pinkal's pinkal:95 terminology) the type of those words in a particular linguistic context and control their semantic contribution, allows regularities to be efficiently captured and increases the flexibility of the linguistic processor.
Chapter 4 teased apart two phenomena, the use of manner of motion verbs on a directed motion interpretation and resultatives, that have traditionally been lumped together into a single analysis due to their syntactic similarity. Through consideration of the semantics of sample data for each phenomenon and of their interpretation in context, I showed that they are distinct. Differences between the two phenomena which had posed challenges to the uniform analyses were shown to stem from underlying lexical differences in the verbs participating in the different forms, and from the distinct analyses which the phenomena require. This chapter showed how those analyses could be formalised given the lexical semantic structures and interactivity mechanisms developed in other parts of the thesis. As in Chapter 3, I showed the use of rules governing the interaction of elements at the sentential level to be critical in capturing regularities and explaining interpretations which do not seem to derive directly from the literal meaning of the words in the sentence. The proposal of these rules is in line with research in the framework of Construction Grammar (Fillmore fillmore:88; Goldberg goldberg:95) which suggests that certain syntactic constructions are paired with specific, non-compositional interpretations, but it goes one step further in claiming that the lexical semantic structure of the words in the constructions also play a role in determining the applicability of the rules.
This thesis has contained much linguistic discussion and has examined a wide range of data. This work differs from much linguistic research in that in addition to considering sentences in isolation, I looked at minimal variants of sentences for comparative purposes and at sentences in context. I believe that the results have shown that a complete account of a linguistic phenomenon must involve contextual analysis, in addition to consideration of subtle semantic differences between similar sentences, in order to identify those aspects of interpretation which stem from conventionalisation and word-specific idiosyncrasies, those aspects which stem from regular, productive processes, and those aspects which stem from an interaction of conventionality and regular processes. The information which is to be encoded in the computational lexicon must reflect the results of such detailed analysis.