Extend Adversarial Policy Against Neural Machine Translation via Unknown Token

TL;DR

This paper introduces DexChar, a novel adversarial policy for neural machine translation that uses character perturbations and improved feedback mechanisms to enhance robustness testing, especially where existing methods fail.

Contribution

It proposes DexChar, an adversarial policy incorporating character perturbations and refined self-supervised matching, improving adversarial example generation for NMT robustness.

Findings

Compatible with scenarios where baseline adversaries fail

Generates high-efficiency adversarial examples

Enhances analysis and optimization of NMT systems

Abstract

Generating adversarial examples contributes to mainstream neural machine translation~(NMT) robustness. However, popular adversarial policies are apt for fixed tokenization, hindering its efficacy for common character perturbations involving versatile tokenization. Based on existing adversarial generation via reinforcement learning~(RL), we propose the `DexChar policy' that introduces character perturbations for the existing mainstream adversarial policy based on token substitution. Furthermore, we improve the self-supervised matching that provides feedback in RL to cater to the semantic constraints required during training adversaries. Experiments show that our method is compatible with the scenario where baseline adversaries fail, and can generate high-efficiency adversarial examples for analysis and optimization of the system.