Mahesh and Nirenburg [75] point out the benefit of a language-neutral ontology for representing the aspects of word meaning common to different languages. The information in the ontology forms the basis for a rich interlingual representation for machine translation. For instance, it includes theme and instrument relations, as well as the more usual is-a and set-membership relations; the ontology also contains selectional restrictions on case roles. The lexicons only need to contain the information specific to particular languages. For a given word, the syntactic information will be provided along with a mapping into the ontology for the underlying concept and a specification of additional constraints as required to model the word's meaning.
This division between the ontology and the lexicon illustrates the distinction between encyclopedic and dictionary knowledge. Strictly speaking, dictionary knowledge only covers the idiosyncracies of particular words, whereas encyclopedic knowledge would cover everything regarding the underlying concepts. In lexicographic practice, this distinction is blurred because dictionaries often incorporate encyclopedic knowledge for sake of understandability and also to bulk up the dictionaries [69].
Jackendoff [60,61] also maintains a division between the language-specific faculties and those used in other conceptual tasks. However, his ontology is fairly shallow and several important details are left unspecified, as in the recourse to Marr's 3D sketch [29] where some perceptual differences are to be represented. The strength of Jackendoff's work lies mainly in the development of the thematic structure, rather in the introduction of new ideas for representing meaning, as will be shown below.
WordNet [83] represents an implicit ontology in which the concepts are lexicalized. It is structured around groups of synonymous words. These groups are called synsets (for synonym sets) and are taken as indistinguishable in particular contexts. WordNet also provides definitions and usage examples; but, more importantly, it provides explicit relationships among the synsets (e.g, is-a and has-a). The main drawback to the WordNet ontology is that it is particular to English. Therefore, separate ontologies would be needed for other languages.2
Talmy [112] presents an interesting sketch on the characteristics of closed-class items, specifically in terms of an organization of the concepts that can be realized as such. For instance, in the spatial domain, close-class items deal mainly with topology rather than magnitude and shape. Comparisons with other cognitive system show that although there are common structural relations (e.g., hierarchical structuring and Gestalts), there are also system-specific relations (e.g., counterfactuals in language).