Onyshkevych and Nirenburg [91] illustrate how the lexicon can be organized in a KBMT system that uses a rich world model (the ontology) versus syntactically oriented transfer approaches.
In addition, this approach is argued to be more promising than some of the other KBMT approaches that just use a superficial interlingua (e.g., those using Jackendoff's Lexical Conceptual Structure).
The goal in interpreting the lexical entries is to derive a semantic dependency structure, represented as a graph linking conceptual nodes derived from the text. First, structures are created for each word sense (instantiation of concepts from the ontology). Then, these are combined by establishing ties among the conceptual nodes. The ties between concepts represent constraints (mainly selectional restrictions), so the process of combination is viewed as constraint satisfaction. The constraints are checked by using the ontology to detect paths between the concepts. The ontology links are weighted, so the shortest cost path is considered. Word-sense disambiguation is thus performed during the course of the constraint satisfaction. Initially only ISA links are considered when trying to satisfy constraints, which corresponds to just checking for literal interpretations. If this test fails, then other links are considered as well (e.g., for metonymic or metaphoric interpretations). The WSD process is somewhat similar to that of Wiebe et al. [121], where Bayesian network propagation is used instead of a shortest paths algorithm.
WordNet, briefly mentioned above, was one motivated primarily by psycholinguistic principles of meaning representation [82]. However, it became useful for general research in computational linguistics. Princeton's cognitive science group [81] manually created WordNet, a thesaural-style dictionary, in which word senses are defined by a set of a words that are synonymous in a particular context. Several lexical relations were explicitly represented in the database, such as ISA links (hypernyms), their inverses (hyponyms), and part-whole relationships (meronyms). Due partly to this explicitness, WordNet has now become more popular for use in research than LDOCE or any other machine readable dictionary. In addition, there are now projects in progress for producing similar dictionaries for several European languages [119].
A special issue of International Journal of Lexicography was devoted to WordNet [80]. Miller et al. [83] give a concise overview of the organization of WordNet. They also discuss the psycholinguistic motivations behind WordNet and some of the advantages of the approach. For instance, WordNet is based on the differential theory of meaning rather than the constructive theory, which is considered impractical. Most of the discussion is on the use of synonymy to organize words conceptually. Words are grouped into sets of words that are synonymous in particular contexts (synsets). Then explicit relations are in terms of the synsets, except for those relations like antonymy which by nature are among word forms. Miller [79] gives an in-depth discussion of how the nouns are organized in WordNet, motivated by what is lacking in conventional dictionaries. Distinguishing features play an important role in lexical hierarchies, so emphasis is on how WordNet can be used to represent this information, specifically through attribution, part-whole relationships, and functional characteristics. Unfortunately, this has not been fully applied with the result that many synsets don't have distinguishing characteristics (a rough estimate is that 67% don't for the nouns in WordNet version 1.6). Fellbaum et al. [39] discusses the unique way that adjectives are organized in WordNet. This reflects the desirability of semantic relations over lexical ones in WordNet. Basically, descriptive adjectives are organized by a combination of opposition relations and similarity: clusters of similar adjectives are paired by the opposition of salient antonyms. This works well because most descriptive adjectives are associated with bipolar attributes. The other main type of adjective, relational ones, are organized via links to the noun hierarchy.
Fellbaum [38] clarifies the verb organization, which is not as simple as the noun organization, due to the difficulting in categorizing the verbs in taxonomies. Also of interest is the discussion of the role of verb alternations in the organization, specifically in the choice of certain top-level nodes. Beckwith et al. [10] provide a technical description of the WordNet system, including the source files, the database, and the interface. But, there is an interesting discussion of how the polysemy count is used to approximate familiarity, which can be used to filter the display of technical terms not of interest to general terms.