Sense variation

The core chapters of this thesis (Chapters 3-5) examine particular linguistic phenomena which involve interactions of lexical semantic structure. Each phenomenon examined involves subtle shifts in the meaning conveyed by the main verb in a sentence, due to the influence of complements or adjuncts. In each case, meaning is inferred which does not appear to be explicitly present in the sentence. The main question which the chapters address for each distinct phenomenon is, where does this extended meaning come from? In each case, the solution will involve lexical specification of semantic structure. Furthermore, that lexical semantic structure will be shown to interact with constraints deriving from discourse coherence and pragmatic reasoning.

This thesis essentially investigates sense variation of verbs and how that sense variation can be modeled through use of the lexicon and the information formalised therein. Words are often polysemous -- they often have several different but related meanings -- and it has been recognised that the variations in meaning which a particular word shows often can be generalised to entire classes of words (see e.g. Levin levin:93 for a thorough overview of verbal syntactic alternations and their corresponding semantic alternations). Recent lexical research has focused on characterising those classes and modeling the alternations in terms of lexical semantic structure (Pinker pinker:89, Jackendoff jack:90, marks_sadler:95, Sanfilippo sanfilippo:95, Pustejovsky pustejovsky:95a, inter alia) but little of the insights from this work has influenced lexicon design for computational systems or constraint-based grammars.

There are various ways in which sense variation has been captured:

• Enumeration: Different senses of the same word can be listed as distinct lexical entries. This could also include listing of all the different subcategorisation frames in which a word can appear, with each frame associated with its own specific semantics, which might ignore the independent contribution of the subcategorised complements. This approach might be combined with constraints specified in the lexical entries about the kind of arguments a word expects, which help to resolve the sense of a word intended on a particular use.
• Lexical rules: A rule can be added to the lexical component which captures a regular meaning shift of a certain class of words. The rule has an input form which specifies the kind of word which undergoes the shift, and an output form which is the shifted version of the word. The input and output forms are distinct lexical entries for the same word, with different semantic and possibly syntactic information. Lexical rules capture generalisations and can be utilised productively for words which match the required input structure, to produce unestablished or novel word senses.
• Type coercion: Pustejovsky pustejovsky:91,pustejovsky:95a suggests that some sense variation can be captured using operators which shift the semantic type of an entity, turning for example an object-referring noun into an event-referring noun by a well-defined generative operation. This operation might be triggered by lexical constraints (e.g. a conflict in the actual type of a verbal argument and the type that the verb expects it to be). This approach differs from the lexical rule approach in that the operators are generally available (not specific to particular word classes) and will only be applied if necessary to resolve a constraint violation.

The enumeration approach fails to record any regularities in sense variation which exist and is a weak lexical model in that it is not productive. The lexical rule approach does allow the capture of these regularities, but as it is restricted to manipulation solely of lexical entries for individual words it cannot be applied in cases in which sense variation stems from properties of non-subcategorised elements in a sentence (e.g. adjuncts). The type coercive approach suffers from a lack of constraints -- it is not always desirable to attempt to shift the type of a word when there is a conflict in types, as the word may simply be infelicitous in that context. These approaches are, however, complementary rather than competing, since each handles some aspect of the creativity of language use.

The approach taken in the thesis will build on the lexical rule and type coercive approaches, in order to achieve maximum generality in the representation. I will utilise lexical rules which can operate at both the word and phrasal levels, by replacing the standard HPSG mechanisms for adjunct addition with lexical rules which add adjuncts to the subcat lists (see Section 3.5). I will identify specific phenomena in which type coercion does seem to occur, and will discuss how the coercion can be represented in the lexicon, and what constrains the coercion. In addition, the use of a hierarchical typed lexicon allows generalisations over related groups of words to be stated at one node in the lexicon and passed down to lower nodes through inheritance.

Furthermore, the implementation of a lexical semantic representation which builds in elements which have been identified in the previous research to be important for modeling alternations and fits directly into the grammatical framework of HPSG (subject to the modification mentioned above) opens up the possibility of capturing generalisations about at least some sense variations in computational systems. As I will argue in Chapter 6, this is important for development of robust systems which can adequately handle the interpretation of natural language, given the flexibility of language use. The proposals I make are for productive processes generating extended senses for verbs, which are constrained by syntactic, lexical semantic, or discourse factors. These constraints must be in place in order to prevent overgeneration of meanings, that is, generation of interpretations for a sentence which are infelicitous or unattested in the language.