We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: Normal vs. Adversarial: Salience-based Analysis of Adversarial Samples for Relation Extraction

Abstract: Recent neural-based relation extraction approaches, though achieving promising improvement on benchmark datasets, have reported their vulnerability towards adversarial attacks. Thus far, efforts mostly focused on generating adversarial samples or defending adversarial attacks, but little is known about the difference between normal and adversarial samples. In this work, we take the first step to leverage the salience-based method to analyze those adversarial samples. We observe that salience tokens have a direct correlation with adversarial perturbations. We further find the adversarial perturbations are either those tokens not existing in the training set or superficial cues associated with relation labels. To some extent, our approach unveils the characters against adversarial samples. We release an open-source testbed, "DiagnoseAdv" in this https URL
Comments: IJCKG 2021
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2104.00312 [cs.CL]
  (or arXiv:2104.00312v4 [cs.CL] for this version)

Submission history

From: Ningyu Zhang [view email]
[v1] Thu, 1 Apr 2021 07:36:04 GMT (2277kb,D)
[v2] Sun, 11 Apr 2021 07:39:20 GMT (2278kb,D)
[v3] Sun, 22 Aug 2021 16:38:47 GMT (2312kb,D)
[v4] Thu, 25 Nov 2021 09:04:41 GMT (2715kb,D)

Link back to: arXiv, form interface, contact.