Program Generation Apparatus Using Natural Language Search and Input-Output Modification

The patent implements a feedback mechanism where the generated program is tested against the natural language specification through a tester that checks whether the program satisfies the specification. If the program fails the test, the system automatically modifies the program by adjusting the translation model or adding correction layers, and re-tests until the program passes. This closed-loop feedback ensures program correctness while maintaining automation.

The patent performs preliminary actions by pre-processing the natural language specification to extract structured information, and pre-training translation models on high-quality parallel corpora before generation. The system also pre-defines a set of candidate programs and pre-establishes testing frameworks, so that when generation occurs, the process is guided by pre-prepared resources that ensure syntactic correctness and semantic accuracy.

The patent merges multiple information sources: natural language specifications, input-output examples, and program syntax constraints are combined into a unified generation framework. The translation model processes natural language specifications while simultaneously considering input-output example pairs, and the generated program must satisfy both the natural language meaning and the example constraints. This merging prevents overfitting by requiring the program to generalize from examples rather than merely memorizing them.

The patent adds a new dimension to program generation by introducing a semantic verification layer that operates beyond traditional input-output matching. Instead of only verifying programs against examples (one-dimensional), the system verifies programs against the semantic meaning of natural language specifications (two-dimensional), creating a richer constraint space that prevents overfitting and improves generalization.