Although generative grammar does incorporate the notions of case and thematic roles, it is restricted to describing how the roles are assigned by particular governing categories. In other approaches, these notions are fully developed and are central to the theory. Two examples are Fillmore's [40,41] work on case frames and Jackendoff's [61] semantic structures which incorporate thematic relations into his framework for general cognition [60].
Fillmore [40] holds that deep structure defines the relevant case relations and that surface case relations are often insignificant in that the mapping from deep structure to surface structure is not one to one. Fillmore identifies a half dozen or so cases that are any case system would include, although more would be needed (e.g., benefactive). See Table 1. A significant aspect of Fillmore's work is the emphasis on semantics. Although the particulars of the transformations he used for illustration might not be in line with current syntactic theory, and although ``syntactic deep structure'' is perhaps more widely viewed as the base structure, the idea that surface case is an imperfect reflection of semantic case is an important ``constraint'' against syntactically-oriented approaches.
| ||||||||||||||||
Jackendoff [60] presents a unified framework for representing conceptual knowledge for all aspects of cognition, not just linguistic. Much of the representation is an outgrowth of earlier work in artificial intelligence, such as by Schank and Wilks,
but there are a few innovations that are noteworthy. One is that all ontological categories are treated uniformly: in particular, manners and directions have the same status as things and events. Another is the emphasis on thematic relations, showing how prepositions play an integral part in the analysis of several different semantic fields. Moreover, for each semantic field, the thematic relations (determined from the prepositions) are all based on extensions to the thematic relations defined for the spatial field.
Later work [61] builds upon this framework to present a detailed analysis of various types of expressions in natural language. The semantic structures are related to syntax mostly via correspondence rules, which apply at construction time. Occasionally, lexical rule variants are given, which can be seen as corresponding to the LFG approach [15]. Correspondence rules are generally preferred over lexical ones when productivity becomes an issue, but the latter type have some advantages when dealing with idioms. Thematic roles play a central role in the representation. These correspond to structural configurations rather than case markers, as in other approaches.
Jackendoff's conceptual representations of words, called Lexical Conceptual Structures (LCS), tend to be at a course level. For instance distinctions in meaning of perceptual objects and motion verbs are to be captured elsewhere in an extension to the 3D sketch of Marr [29]. Two separate levels of conceptual roles are used: one level describes the interaction among the participants, such as the beneficiary; the other encodes other thematic information, such as the theme and destination. The benefit is that it facilitates the mapping into syntax and that high-level differences or similarities become more apparent. An important aspect of his work is that the interpretation of adjuncts is given full treatment, a requirement for sentential interpretations. Often interpretation is limited to predicate argument structure [15,96]. Some of these adjuncts fill in or replace information already present in the conceptual structure, such as the specialization of a default value; others add subordinate conceptual structures such as in describing the manner of an action. These are handled by general correspondence rules so that the lexical entries can be minimized (compared to a lexical approach) and that the relation to syntax is more apparent (compared to syntactic transformations).
Although it is natural to use a 3D representation for encoding fine
distinctions, it seems that some abstract conceptual information is
not being modeled. In this respect, Amsler [3]
provides a more standard approach for capturing the distinctions, as
illustrated with motion verbs. This uses a decomposition with binary
features such as [+/- velocity], [+/- steady], [+/- continuity], as
well as more complex features [medium=
air
water...
]. Jackendoff
and others criticize the featural approach in general as being too
unconstrained as to what can serve as features. Medin and Smith
[111] acknowledge such problems but feel that
they are not insurmountable. Cruse [31] takes a middle
ground on the view by distancing himself by not using the term
semantic feature but still believing in their benefit (p. 22):
Representing complex meanings in terms of simpler ones is as problem-ridden in theory as it is indispensable in practice.Furthermore, as pointed out by Fellbaum et al. [39], the effect of decomposition can be achieved through relational analysis, for which there appears to be less theoretical objections.
Although Higginbotham [55] discusses thematic relations within the constrained framework of generative grammar, he provides some useful insights to their use, specifically on how they are filled by various constructs. He incorporates Davidson's [32] notion of event variable and treats them as a type of theta role. For instance, verbs have an open event position that normally get bound by tense; but, the position can also be bound by gerunds. Theta positions also come into play with adjectives. To account for attributive versus predicative usages, separate roles are used. One gets filled by co-referencing a position in the adjective with that of a noun; the other gets filled by association a position with the noun structure itself, which can be viewed as the noun's intension [43].
Raskin and Nirenburg [98] provide a critical review of several linguistic theories that specifically address the semantics of adjectives. An important consideration in their assessment was the suitability of the theory to the practical needs of lexical acquisition. One general comment they make is that much of the work emphasizes rare or interesting cases at the expense of commonplace adjectives, the semantics of which turn out to be nontrivial. They developed a practical methodology for applying diverse theories, as discussed below.