Summarization Attack via Paraphrasing (Student Abstract)

Many natural language processing models are perceived to be fragile on adversarial attacks. Recent work attack has demonstrated a high success rate sentiment analysis as well classification models. However, attacks summarization have not been studied. Summarization tasks rarely influenced by word substitution, since advanced abstractive summary utilize sentence level information. In this paper, we propose paraphrasing-based method We first rank the sentences in document according their impacts summarization. Then, apply paraphrasing procedure generate samples. Finally, test our algorithm benchmarks datasets against others methods. Our approach achieved highest and lowest substitution rate. addition, samples semantic similarity with original sentences.