Limited context semantic translation from a single knowledge-base for a natural language and structuring metarules

Abstract

The article introduces an experimental system which produces multilingual semantic translations from relatively short texts from a given context.

The system was conceived as an investigation instrument whose characteristics are the following:

—the elaboration of a single analyzer and generator able to receive, in the form of data, specific information concerning national language, within the limits of a given area of application;

—the use of exclusively semantic internal representation, whose formation is derived from “frames” (an object is defined as a list of “attribute-value” couples, permitting recursion);

—a single knowledge-base is used for each natural language as initial data (the grammar transmitted was of a semantic-syntacticatn type);

—the structure of grammar used in computer language processing being rarely as well adapted to analysis as to generation, the system itself is provided with the possibility of transforming and reorganizing the information for more efficient use at one stage of processing (thus, theatns are used directly in analysis, whereas for the purposes of generation they are “explored” and their information reorganized);

—to produce a text, the use of general structuring principles (independent of language) are experimented with. These principles are given in the form of metarules. The application of these metarules to the restructured grammar of a natural language produces specific structuration rules, peculiar to this language.

Although the system was conceived for any conceptual area or language, the present knowledge-base of the system (the experimental support) is based on a collection of elementary exercises in three-dimensional geometry written in Rumanian and in French.