A linguist, a computational linguist, and an epistemological engineer propose a theory of knowledge structures using the Prolog system OSKAR, translated from German as “Object Schematic for the Conceptual Analysis of Spatial Properties of Objects.” In their quest for a key to humanity’s complex perception of space, after rejecting alternative ontologies (such as Jackendoff’s conceptual semantics, Herskovits’s experiential approach, and the prototype semantics of Vandeloise, Hottenroth, and Lakoff), the authors propose their solution to problems of spatial representation. Their object-centered and axis-based approach draws on a predetermined inventory of object concepts and categorization grids that define constitutive spatial properties, formalized as primary perceptual space.
By analogy with phylogeny and ontogeny and assuming a universal grammar with transitory internal states, the Prolog program establishes a modular grid: (a) the dimensional designation of objects according to assigned and finite spatial dimensions, such as length, width, height, depth, and thickness; (b) a categorization of spatial objects according to their dimensional gestalt (which comprises position properties and correlates with mobility); and (c) positional specification of objects (such as lying down, standing upright, or upside down). From the top-down schema emerges a universal conceptual structure, Cspace. The convergence of syntactically combined lexical items forms the heart of the authors’ linguistic approach to spatial knowledge (p. 16).
A spatial grid of six-dimensional semantic parameters and their assigned values illustrates “acceptable” combinatorial restrictions and compatibility restrictions. Twenty interpretable combinations of dimensional adjectives yield 78 arithmetically possible variations, of which only 40 are interpretable. For example, a semantically linked adjective such as “high” may have a fixed (as in “river”), canonical (as in “desk”), inherent (as in “book”), or unspecified (as in “pole”) orientation. The distinction between semantic form and concept structure marks one of the major differences between the present approach to semantics and previous analyses (p. 24). Contextually induced perspective is linked with observer orientation (so-called “encounter situations,” in which the observer faces the object from a distance, are distinguished from “coincidence situations,” in which the observer instantiates a bounding effect). From the contextually induced specifications, including deictics, the authors’ modular approach to cognition reduces “seemingly complicated facts about dimensional designation of objects in space [to] clear-cut results” (p. 68).
The remainder of the book deals with implementation of the Prolog program OSKAR and its integration with an IBM Germany LILOG system. Input consists of any combination of noun object and one or more of the designated dimensional adjectives (such as “high tower,” “long and thick pole,” or the predicate “A thick pole is long”). Natural language adjectives are transformed into dimension assignment parameters and nouns into object schemata according to the three rules of the knowledge representation system: identification of dimension value, gestalt specification of value, and context specification (Lang’s principle of one-to-one assignment precludes multiple evaluations). Unfortunately, ambiguous cases outside of the inventories and hierarchies--does the object stand or lie?--produce a diagnostic error message. The authors concede that “the last word on rules for positional properties has yet to be worked out” (p.91). OSKAR will infer, for example, that the semantic group “long tower” is acceptable because the object may be tilted on its side, yet “long hill” contains an unacceptable entailment “because hills are immobile” (p. 100). Despite such restrictions, the authors envisage the eventual application of OSKAR’s object-centered representations in tutorial systems for language learning, naming tasks, and recognition tests.
Through the LILOG system, OSKAR bridges the gap between natural language processing and knowledge representation. Using a modularized object ontology in an AI-style hierarchical taxonomy, the LILOG system performs syntactic and semantic analyses of spatial and temporal knowledge. LILOG’s inference engine remains domain independent because it treats only variations in the dimensional meaning of objects and their contexts. A descendant of the KL-one family of languages and order-sorted predicate logic, it provides set-theoretic operators, roles, and feature structures [1].
Again, referential objects (RefOs) assigned to arguments or functional values defined by the sorts as “illogical” or “incoherent” must be excluded. LILOG combines formalized defaults, a form of non-monotonic reasoning, with the structure of the sort hierarchy, an integration of Lang’s theory of dimensional designation and positional variation. According to the authors, OSKAR has demonstrated the context dependency of dimensional expressions by modeling position properties of objects (standing, lying down, and so on), but without representing context explicitly. LILOG articulates different position properties and different object schemata (OSs) over time.
In the LILOG ontology, constitutive properties are again encoded as unalterable primary entries. To avoid confusion of dimension parameters and dimension values, a knowledge engineer must monitor the sort hierarchy. If canonical properties assumed by default are contradicted by subsequent evidence, then a new object schema is assigned to the RefO in question. Modification of OSs for positional variation is combined with treatment of temporal intervals needed for tense and aspect. Underlying Lang’s theory of dimensional expressions is a theory of gradation, compatible with context changes by positional manipulation.
The book emphasizes a logic of spatial cognition and the design of input and output filters. The limited spatial universe does attempt to simulate changes in convergence among lexical elements admitted to the set. No demonstrations of extensive discourse output are provided, however. As an exercise in objectivist perceptions and entailments, AI researchers may find the examples of interest. Yet the stratifications, while logically compatible with a linear (point-even) time frame, do not fully encompass experiential representations of space along the still fuzzy boundaries between pragmatics and semantics.
A modest bibliography, but no index, is provided.
