[I]t is methodologically important to distinguish between linguistic and non-linguistic representation, even though the two have to be interrelated so that linguistic utterances can be interpreted as having some connection with the real world... we wish to provide a testable constrained theory, and a formal representation language, and to avoid problems which arise in knowledge representation which do not have a linguistic dimension.
The view of lexical semantics advocated by Copestake is that lexical
representation must include information about a word which is
necessary to account for its grammatical use, but which is not
entirely predictable from real world knowledge of the entity denoted
by the word. Furthermore, Copestake argues that a certain amount of
real world information must be lexically encoded in order to describe
interactions of lexical semantic variations with syntax (for example,
Copestake extensively discusses the count/mass distinction which has a
syntactic reflex and which in some cases depends on lexical
specification
and in others depends on regular lexical
sense extension properties).
In this thesis I will adopt a rich lexical semantic structure, involving a certain amount of lexical decomposition and hierarchical relations between lexical semantic structures, in order to explain a variety of linguistic phenomena. My point of departure in choosing an appropriate representation is the assumption that the inferences which can be drawn about the use of a particular word in a particular context must play a role in determining lexical semantic structure. That is, there are regularities in the interaction between syntactic structure and word meaning which stem from critical semantic features of the word. Without a lexical semantic representation of such features it is impossible to capture those regularities. This point has been extensively argued by Davis davis:95, who develops a theory of word order based entirely on semantic relations encoded in the lexical entries of verbs (i.e. a linking theory).
A purely pragmatic account of word meaning cannot explain the syntactic reflexes of semantic shifts, unless the pragmatic reasoning is assumed to have access to purely syntactic information. This is the line taken by Hobbs et al (1993). However, on such a view the pragmatic module cannot be language independent since it will be influenced by and would depend on syntactic properties which vary between languages. A pragmatic component which interacts with language dependent syntactic information is undesirable for several reasons.
Firstly, it forces the linguistic system to be modeled nonmodularly since syntactic, semantic and pragmatic information must be considered simultaneously in the construction of an interpretation of a sentence or discourse. This obscures the distinction between general conceptual knowledge and specifically linguistic knowledge, critical for explaining certain aspects of language use. Consider the sentences in rep1.
John met Sue on/*in Tuesday. John met Sue in/*on the morning. John met Sue in/*on March. John met Sue (*in/*on) last night. John met Sue at/*in/*on ten o'clock.
From a world knowledge perspective, phrases like March, Tuesday and the morning pick out particular portions of time, of varying lengths. Periods during the day are subintervals of days of the week, which are subintervals of the weeks of the month, which are in turn subintervals of the months, etc. These interval relations can be conceived of hierarchically, with the days being subtypes of the week, the weeks of the month, and so on. In this knowledge, however, there is nothing to explain the fact evidenced by the data in rep1 that the intervals of time corresponding to days are picked out by a different preposition from other such intervals (on vs. in). It cannot follow simply from differences between the relationship of the described event John met Sue to the specific time interval in the prepositional phrase (PP); in each case that event is interpreted to take place during a sub-interval of the interval picked out by the PP. The choice of preposition is instead purely a matter of linguistic convention (in English). Days and months are typed differently at a linguistic level but not in the world knowledge ontology. The distinction between language-dependent usage facts and facts about the world must be made explicit in the language model.
Secondly, a nonmodular approach leads to overgeneration of word senses within a productive linguistic system. Imagine that speakers of English know that it is possible to refer to the meat of an animal by using the name of the animal, as shown in rep2a. This can be considered a productive generalisation as it can be applied in novel instances rep2b.
We ate chicken/lamb/turkey for dinner last night. We ate kangaroo/aardvark for dinner last night. #We ate cow for dinner last night. We ate beef for dinner last night.
rep2c-d exemplify, however, a process known as lexical blocking (Copestake and Briscoe 1995): the application of a productive rule can be blocked by the presence of a lexeme with the same meaning (but different form) as the lexeme which would be generated by the rule. In this case, the use of cow for the meat of a cow is blocked by the independent existence of the word beef in the lexicon. Lexical blocking must be considered a linguistic phenomenon which occurs in isolation of world knowledge in that it depends purely on lexical information and not on concepts associated with the denotation of a word. A purely pragmatic account of word meaning could therefore not account for such processes which restrict sense generation.
The necessity of representing lexical semantic information is clear: simply treating sentential constituents as syntactic entities cannot result in an analysis which is fine-grained enough to account for the range of grammaticality judgements about sentences. It is necessary to take into account the semantic nature of entities combining into larger constituents in order to adequately capture grammatical sentences. These observations have influenced the definition of constraint-based grammatical theories: HPSG (Pollard and Sag 1994), for example, explicitly allows for the encoding of both semantic and contextual information in the structures representing words and phrases. This information interacts with syntactic specificatons in the construction of grammatical sentences.
An example of a grammaticality difference which cannot be
captured through syntax alone is shown in ji1-ji2. Why is it
possible for one spatial prepositional phrase to modify close and
not a different one, while push does not show this restriction?
There is clearly no surface difference in the structures of ji2b and
ji2c, so the
difference must be due to something semantic.
John pushed the filing cabinet. John pushed the filing cabinet at the office. John pushed the filing cabinet into the wall.
John closed the filing cabinet. John closed the filing cabinet at the office. *John closed the filing cabinet into the wall.
As I will argue below, Davis' (1995) representations cannot be used to explain all linguistic phenomena which depend on lexical semantic features, because it concentrates entirely on the entailments relevant to the lexical semantics-syntax interface. I will therefore extend his representations with further semantic decompositions derived from the work of Jackendoff (e.g. 1983, 1990). This work has provided many insights with respect to the identification of components of lexical semantic structure which influence syntax and the encoding of generalisations about inferences to be drawn from uses of a wide range of verbs and prepositions.
In this chapter, I will outline the approach to the representation of lexical semantic structure, conceptual semantics, which Jackendoff has proposed and the constraint-based version of lexical semantic structure put forth by Davis. I will argue that the framework and analytical methodology Jackendoff develops will give clear structure to the endeavour of capturing linguistically-relevant semantics, and will show how Davis builds on this framework and brings the representations in line with the constraint-based grammar framework of Head-driven Phrase Structure Grammar (HPSG). I assume that conceptual structures are representations of those aspects of the meaning of words and phrases which are relevant to syntax and over which particular, regular inferences can be made, as it is in this sense that these structures can be utilised within a computational framework. I therefore avoid a complete representation of knowledge about entities denoted by a word in that word's representation in the lexicon. The attempt to capture all knowledge associated with a word at the lexical level of representation would place too much pressure on the lexicon, as follows from Copestake's quote above. The representation I settle on will explicitly reflect a linguistic, and not a world knowledge, ontology.