Quillian's [97] work on semantic memory is significant for several reasons. The main contribution is the introduction of semantic networks for knowledge representation. Also, it was one of the first computational attempts to emphasize semantics over syntax. In relation to the proposed work, it is quite relevant because his work centered around encoding entire dictionary definitions, not just the superordinate information, thus addressing word sense distinctions. Each word sense is represented by a graph where the nodes are the defining words and the links are the relations between the words, based on the definition. Compared to current typical lexicons, this scheme lacks syntactic information and only has limited support for subcategorizations (via the parenthetical parts of dictionary definitions). However, it does have a strong emphasis on the conceptual encoding.
Schank [107] popularized the notion of semantic-based analysis in the use of conceptual dependencies to represent meaning. The motivations for the conceptual dependency representation are to account readily for paraphrases, to facilitate inference, and to model features of memory. To this end, a small set of semantic primitives was developed, with which all expressions were encoded. For example, ATRANS encodes abstract transfer. Conceptual categories serve as the basic unit of the representation. Relations among these conceptual categories are called dependencies. This approach has an advantage over Quillian's in facilitating inferences over the encoded meaning representation. However, one important drawback is that subtle distinctions in meaning might be lost in the conversion process.
At a high level, Wilks' [123,124] work was similar to Schank's, but he emphasized the resolution of lexical ambiguity, which makes it more relevant for the proposed research.
Furthermore, his representation clearly distinguishes criterial aspects of word meaning from optional (or preferred) aspects: one element defines the general category (as with genus terms) and the other elements provide optional characteristics (as with differentia).
In his basic mode of analysis, interpretation is performed by finding the set of word-sense formulas maximizing the density of satisfied preferences, which mainly cover selectional restrictions. His extended mode of analysis is more similar to Schank's approach, with commonsense knowledge represented by rules operating over the semantic primitive encoding. Wilks's rules were declarative in nature, whereas Schank's were procedural. However, procedural knowledge was used in the heuristics for the selection of competing interpretations. Later extensions [124] organized the vocabulary through a thesaural hierarchy.
This hierarchy was also used to organize a much broader base of common-sense knowledge, encoded in a frame-based representation similar to that used for the text. This common-sense knowledge was used to handle more complicated forms of preference breaking, requiring extrapolations such as substituting a general action derived from a superordinate's frame (e.g., [car uses gasoline]) in place of a specific action that violates selectional restrictions (e.g., [car drinks gasoline]).
Bruce [19] provides a survey of early uses of cases systems in natural language processing, as well as providing background on surface cases vs. deep cases. Several criteria for selecting deep cases are discussed, such as the need for distinguishing word senses, for specifying events uniquely, and for modeling relevant domain aspects [19, p. 336]:
A case is a relation which is ``important'' for an event in the context in which it is described.Sinc the notion importance is pragmatically determined, so this makes the choice of cases open-ended.