Add adversarial training capability to semi-supervised text classification

AI Impact Summary

Capability introduces adversarial training methods into semi-supervised text classification workflows, enabling models to learn from labeled data and adversarially perturbed unlabeled samples. This improves robustness to input perturbations and helps maintain performance when labeled data is scarce. It will require new data pipelines for perturbation generation, additional hyperparameters, and longer training times, so teams should start with a small pilot to tune robustness gains before broader rollout.