Small [110] popularized the idea of ``word experts'' that are autonomous agents that encapsulate all sorts of knowledge regarding a word (or stem, affix, etc.). In his model, the control mechanism is modeled after the Unix-style processes and demons. Specifically, the word experts become active only for as long as they can perform useful work, such as refining a concept based on long-term memory (e.g., world-knowledge). When they can do no more productive work they post type-specific interrupts or signals and then suspend themselves. Note that some aspects of this work were not implemented; in particular, human input is required for certain forms of pattern matching and incorporation of world knowledge. The state of the experts at intermediate stages are similar to that of Hirst's Polaroid words (see below). Also, the operating system control paradigm is similar to Slator's [108] PREMO parser. But, in both cases, although the operating system metaphor clarifies the general processing, it is stretched past the limit and becomes awkward. The following quote gives an idea of the complexity of these word experts [110, p. 146]:
Hirst [58] developed a declarative system for representing lexical knowledge, using a conventional frame-based representational language. This work was influenced by psychological research into negative priming. To model priming, spreading activation is implemented via marker passing among nodes in the knowledge base. To avoid extraneous connections, a few constraints are imposed during marker passing, for example, limiting maximum length of paths, restricting traversals of certain types of links in reverse, and blocking paths from nodes with many connections.
Hirst proposed the notion of self-developing objects, called Polaroid Words, for modeling the incremental development of lexical knowledge during comprehension. Communication among the objects is syntactically governed, with the exception that communication with prepositions is unrestricted to facilitate case satisfaction checks.
If the objects are not fully resolved after the sentence is processed, then several fallback (procedural) mechanisms are applied, such as selecting preferred senses, and relaxing the marker passing constraints. Note Hirst and Small both rely on word-specific ``agents'' to encapsulate lexical knowledge and world knowledge. However, Small's approach is heavily procedural, whereas Hirst's is very declarative.
The collection edited by Viegas [118] provides further information on semantic lexicons. Several of the papers deal with lexical acquisition and also lexical rules for expanding the lexicon's coverage. For instance, Chang and Chen [27] describe an algorithm for augmenting LDOCE with information from Longman's Lexicon of Contemporary English (LLOCE), a thesaurus organized using 14 subjects and 129 topics. These topic identifiers are used as a coarse form of sense division. The matching algorithm works by computing a similarity score for the degree of overlap in the list of words for each LDOCE sense compared to the list of words from the LLOCE topics that contain the headword (expanded to include cross-references). Briscoe and Copestake [16] show how lexical rules can be controlled through the use of probabilities. To account for unattested rule applications for particular words, this uses frequency smoothing based on the relative productivity of each rule, as estimated from a corpus. Obtaining the required probability estimates will not be straightforward. However, a rough approximation can be made for cases when distinct syntactic realizations are produced for the same lexeme by using a part-of-speech tagger and tabulating the times the lexeme of the appropriate (taxonomic) type occurs with different parts-of-speech. Note that this approach handles blocking by assigning very low probabilities to cases unattested.