NLP Training Samples with Inflectional Perturbations

The patent applies preliminary action by generating adversarial training samples with inflectional perturbations before the actual training process. These perturbed samples, which include variations in verb conjugations, noun plurals, and other morphological changes, are created in advance to expose the model to diverse linguistic patterns during training, thereby improving its adaptability to non-standard English while maintaining standard English accuracy

The patent implements parameter changes by systematically modifying linguistic parameters in the training data, specifically inflectional morphology parameters such as verb tenses, noun plurals, and adjective forms. By varying these parameters across training samples while preserving semantic meaning, the model learns to recognize patterns across different inflectional forms, resolving the contradiction between accuracy on standard English and adaptability to inflectional variations

The patent converts the harmful effect of linguistic discrimination into a benefit by using adversarial perturbation techniques. Instead of directly addressing bias through complex debiasing methods, the system introduces controlled inflectional variations that mimic the errors made by non-native speakers. This transforms the training process to actively combat linguistic discrimination by making the model robust to such variations, thereby eliminating bias while maintaining training efficiency

The patent applies parameter changes by modifying the linguistic parameters of training samples to include inflectional variations. By systematically varying morphological parameters such as verb conjugations and noun forms while keeping the core semantic content unchanged, the training process remains relatively simple yet effectively reduces linguistic discrimination by exposing the model to diverse language patterns