Skip to content

We demonstrate that near SOTA summarization models are fairly robust to different types of transformations and goal-directed perturbations.

Notifications You must be signed in to change notification settings

a1noack/summarization_robustness

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Inverstigating the robustness of summarization models

With this repository, we investigate the robustness of summarization models (specifically PEGASUS and BART) when their test set samples are transformed with various attacks (directed perturbations) and transformations (undirected perturbations).

We use two commonly referenced seq2seq adversarial attacks, Seq2Sick and MORPHEUS. Because both of these attacks were originally intended for use with translation models, we propose two new attacks more suited to the summarization task, Conceal and ROUGE Attack.

The goal of the "Conceal" attacker is to remove the word in the original generated summary with the lowest TF-IDF score by changing the input document without removing the target word from the input. The goal of the ROUGE Attacker is to minimize the ROUGE score between the original generated summary and the new summary without changing more than two words in the input.

In the table below, we show how these various attacks affect the ROUGE score (each cell in the table holds the ROUGE-2 score of the model on the perturbed test set samples). We then compare these directed attacks to various undirected textual transformations.

In the table, PEG=PEGASUS, CNN=CNN-Dailymail, Gig=Gigaword.

Transformation name Transformation type Transformations allowed PEG/CNN PEG/Gig PEG/X-Sum BART/CNN BART/Gig BART/X-Sum
No Attack NA NA 21.43 20.52 24.53 21.07 18.58 22.36
Seq2Sick Directed 1+ 18.2 11.66 18.45 18.54 12.41 16.33
MORPHEUS Directed 1+ 13.61 15.93 14.86 13.26 14.8 13.4
Conceal Directed 1+ 16.56 11 17.61 17.36 11.6 16.76
ROUGE Attack Directed 1-2 14.08 8.1 11.69 15.81 7.6 12.68
Adj. char. swap Undirected 1 20.83 18.59 24.39 20.66 17.28 22.14
Word synonym swap Undirected 1 20.62 18.96 24.48 20.24 17.51 22.19
Add punctuation Undirected 1 20.61 20.3 24.43 19.4 18.68 22.28
Remove punctuation Undirected 1 20.61 - 24.56 20.72 - 22.4
Swap sentence order Undirected 1 20.24 - 24.44 20.85 - 22.33
Append irrelevant Undirected 1 20.54 13.99 24.51 18.77 29.15 22.29
Adj. char. swap Undirected 2 - 17.19 23.61 20.13 16.17 21.85
Word synonym swap Undirected 2 - 18.21 23.86 19.98 17.32 21.56
Add punctuation Undirected 2 20 20.74 23.63 20.02 18.55 21.6
Remove punctuation Undirected 2 20.13 - 22.13 20.79 - 19.98
Swap sentence order Undirected 2 20.38 - 21.12 19.94 - 19.06
Append irrelevant Undirected 2 20.91 19.67 23.67 20.31 18.73 22.16

About

We demonstrate that near SOTA summarization models are fairly robust to different types of transformations and goal-directed perturbations.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published