SynBoost: Robust Text Generation Model via Beam Search and Synonym-Driven Boosting

Text generation models are often affected by adversarial examples, such as injecting some adversarial samples into the model and affecting the model’s predictions. Researchers have proposed various adversarial training methods to improve the robustness of text generation models in NLP. However, these methods often focus on improving the model’s success rate against specific attacks while ignoring its overall robustness. Here, we identified specific words that significantly impact model judgments and employed a beam search algorithm to locate these words with high importance. We then adjusted the synonym embeddings for these words. Following targeted tuning of synonyms for specific words, the deep learning model trained using this approach maintains strong robustness against other adversarial attacks. Due to the minor embedding adjustments, the model retains high accuracy when processing regular input.