Anthony Davis introduces a linking theory in his thesis (Davis 1995) which utilises to a limited extent insights made by Jackendoff (1983, 1990) about lexical decomposition. He formalises this decomposition in terms of feature structures, capturing verb meanings in a multiple-inheritance lexical semantics compatible with HPSG. I will take Davis' representation as a starting point for the development of the representation to be used in subsequent chapters of this thesis. I will augment it with further decomposition, in order to address the phenomena relevant to this thesis. In this section I will introduce his representation and in the following I will discuss my extensions to it. I will not go into the details of Davis' proposals for the mapping between semantics and syntax, for although the issues raised by that mapping drive his choice of the level of lexical semantic decomposition in his representations, there is nothing in his proposals which precludes further detail in the representation.
Davis begins his discussion of lexical semantic representation with a critique of the `traditional' conception of thematic roles as introduced by gruber:65 and fillmore:68. He argues, drawing from dowty:91, that it is difficult to identify a (relatively small) set of thematic roles which can be used to classify each argument of every predicate, and furthermore that argument mapping cannot easily be explained in terms of such a set. He points out that the traditionally assumed set of thematic roles runs into difficulty in accounting for the primacy of causation relations: the causer of a causative verb is realised as the subject of a sentence despite the presence of other agentive arguments. Dowty's dowty:91 solution to this problem is to eliminate an explicit representation of thematic roles, instead explaining the mapping in terms of a numerical comparison of the lexical entailments associated with the arguments of a predicate which leads to identification of proto-agents which map to subject position and proto-patients which map to object position. Davis argues that this explanation is too dependent on surface transitivity and identifies weaknesses with the numerical comparison approach. He goes on to show how explicit representation of certain lexical entailments can better model linking.
The lexical entailments which Davis represents are derived from suggestions by dowty:91, wechsler:91, and Jackendoff (1983, 1990). He defines a multiple-inheritance hierarchy of lexical semantic relations, each of which specifies the proto-role properties which hold of the relation's arguments. The relations characterise those properties that are relevant for linking and argument selection. The sort hierarchy of lexical semantic relations and the list of proto-roles are presented in Figures 2.2 and 2.3, respectively. In the sort hierarchy, I specify the features which are defined for each type. The values on some features are left unspecified -- this indicates that Davis has not specified a typal restriction on the values. For others, a specific type is specified (e.g. the value of grnd for a mot-rel must be a path); this is indicated by a non-parenthetical italicised type. Values given in parentheses are intended to provide only an intuitive indication of the entailment associated with that feature in a particular relation; they do not correspond to specific types in the hierarchy. So for the cause-rel relation, the value of the act feature should be interpreted as a ``causer'', while the value of the und feature should be interpreted as ``caused''.
Figure 2.2: Sort hierarchy of proto-roles in Davis (1995)
Figure 2.3: Summary of proto-roles in Davis (1995)
Every verb in the lexicon will correspond to at least one sort in the sort hierarchy. This sort aims to capture basic entailments associated with the verb, minimally identifying the participant roles, designated in terms of the proto-roles in Figure 2.3, of the kind of event that is picked out by the verb. There is in addition a constraint, the entailment-to-attribute-condition (Davis 1995, ch. 3:27), which requires that participants in an event ``of which a particular proto-role entailment holds (in virtue of playing a particular participant-role in an event) must be denoted by the value of a particular proto-role (act, und, or some other attribute) in lexical semantic structure''. This condition ensures that the values of features in lexical semantic structure will correspond to the appropriate participants in an event.
Participant roles in different events are related to one another by virtue of the inheritance relationships between the sorts. These relationships provide the basis for Davis' generalisations about the mapping between semantic and syntactic roles.