Perturbations
With their success on text classification, we examine the robustness of summarization models against adversarial perturbations, which can be in different levels – character, word, sentence, and document. The space of possible modifications at every level is huge. We show how an attacker, leveraging the biases in summarization models, can implement sentence exclusion attack which can also result in quality degradation.
Character Level
The main motivation behind applying these perturbations is that they can be used to simulate common typo errors and input noise that can occur in real-world scenarios.
Charater swap
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation word_deleteCharater Delete
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation character_deleteCharater Replace
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation character_replaceCharater Insert
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation character_insertCharater Repeat
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation character_repeatWord Level
Just like the character level, it simulates common typo errors and input noise that can occur in real-world scenarios.
Word Delete
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation word_deleteWord Synonym
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation word_synonymReplacement with Homoglyphs
Homoglyphs are visually similar characters/ words that are less noticeable to human readers and can be used for deceptive purposes. To assess the models’ performance, one character or word at a time was replaced with its homoglyph counterpart.
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation word_homographSentence/Document Reordering
In natural language, the order of sentences and paragraphs is important as they account for understanding the context. This can be disrupted due to formatting issues or intentional manipulation. We evaluate the models’ robustness against such changes in structure by moving one of the sentences in a document from the top to the bottom, and in the case of documents, by placing the top document at the bottom
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation document_reorderSentence Paraphrasing
Paraphrasing is a common phenomenon in natural language, and models should be able to handle paraphrased expressions while capturing the core meaning. With this perturbation, we test the models’ ability to summarize effectively without any change, while replacing the original sentence with its paraphrased version.
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation sentence_paraphraseReplacement of original sentences
Once these perturbations are executed at the character, word, and sentence level, we replace the original sentences with the sentences containing them. In case of document perturbations, we just rearrange the order of documents and observe the model’s capability to identify the document again.
summarization_robustness --model t5-small --dataset yaolu/multi_x_science_sum --split validation --size 50 --perturbation sentence_reorder