Adversarial machine learning algorithms for English translation quality estimation

Reliable evaluation of the quality of machine translation is essential to ensure a reliable automatic translation system.However, adversarial attacks can reduce evaluation performance by subtly disturbing sentences and endanger the security of key applications.This paper proposes a comprehensive confrontational robustness enhancement framework specially designed for translation quality evaluation, an adversarial robustness enhancement framework.The framework integrates a multi-grained confrontation sample generator, a dynamic confrontation training mechanism based on the relaxation of master and apprentice labels, and online defence module.The experiment was verified on the machine translation and multilingual quality evaluation seminar and post-editing task data set: this method increased the robustness of the model by 34.2%, reduced the average prediction error from 18.7% to 12.3% in the attack state.The framework shows stable performance in multiple fields providing an effective solution for building safe and reliable actual scene translation quality evaluation system.