Seq2seq RepairGAN for Automated Software Bug Correction

Instead of training on paired (incorrect code, correct code) examples as in conventional seq2seq approaches, this patent inverts the training approach by using only incorrect code examples as input and training the model to generate correct code without requiring the correct code to be explicitly provided in the training data. The model learns to correct code through adversarial training where the discriminator tries to distinguish between real correct code and generated correct code, and the generator (seq2seq model) tries to fool the discriminator.

The discriminator acts as an intermediary component that mediates between the generator (seq2seq model) and the training process. The discriminator receives both real correct code examples and generated correct code examples, and its ability to distinguish between them provides feedback that guides the generator's learning. This intermediary mechanism enables the model to learn code correction without requiring explicit paired training data, as the discriminator's feedback substitutes for the need for labeled correct-code examples.

The system performs self-service by automatically generating correct code examples from incorrect code examples through the generator's correction capability. Instead of requiring human annotators to manually create paired correct-code examples, the model itself generates these examples during training. The discriminator then validates these generated examples, and through iterative training, the model improves its accuracy without human intervention in the data preparation process.

The patent replaces the mechanical process of manual human annotation with an automated machine learning-based system. The generator (seq2seq model) automatically transforms incorrect code into correct code examples, substituting the mechanical task of hand-annotating code pairs with an automated neural network process. This substitution dramatically reduces time consumption while maintaining or improving model accuracy through the adversarial training mechanism.

The discriminator provides continuous feedback during training by evaluating the quality of generated correct code examples. This feedback mechanism allows the generator to learn from its mistakes and improve its correction capability. The adversarial setup creates a feedback loop where the discriminator's ability to distinguish real from generated code pushes the generator to produce higher-quality corrections, thereby improving reliability while maintaining high productivity through automation.

The system employs beforehand cushioning by training the generator to produce code that passes the discriminator's validation before the corrected code is actually used. The adversarial training process pre-validates the correctness of generated corrections through the discriminator's filtering mechanism, cushioning against the introduction of new bugs by ensuring that only high-quality corrections make it into the output. This prior validation layer protects against reliability issues while enabling automated high-speed correction.