Boris Galitsky, Daniel Usikov
We attempt programming spatial algorithms in natural lan-guage. The input of the proposed system is a natural language description of a spatial processing algorithm, and the output is the object-oriented program code to be compiled and executed. Two approaches are proposed and evaluated: the first one is based on textual pattern matching: the best fit pattern is selected for each sentence, and objects and methods are instantiated according to this textual pattern. The second one converts text into logic forms subject to a number of transformations to derive resultant code. A number of heu-ristic rules are built to perform these transformations at every step. Additional pass is then required to process sentences’ coreferences to find identical objects, methods and variables between code statements. The result of this preliminary research suggest that the using the system for programming in natural language in the interactive mode, where a user manually edits the generated code, is a way to noticeably increase coding performance. However, at this time, an accuracy of a fully automated (i.e. non-interactive) code generation mode is still too low to be usable.
Subjects: 13. Natural Language Processing; 3.2 Geometric Or Spatial Reasoning
Submitted: May 7, 2008