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

Download:

Current browse context:

cs.CL

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

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

Computer Science > Computation and Language

Title: Bipol: Multi-axes Evaluation of Bias with Explainability in Benchmark Datasets

Abstract: We investigate five English NLP benchmark datasets (on the superGLUE leaderboard) and two Swedish datasets for bias, along multiple axes. The datasets are the following: Boolean Question (Boolq), CommitmentBank (CB), Winograd Schema Challenge (WSC), Wino-gender diagnostic (AXg), Recognising Textual Entailment (RTE), Swedish CB, and SWEDN. Bias can be harmful and it is known to be common in data, which ML models learn from. In order to mitigate bias in data, it is crucial to be able to estimate it objectively. We use bipol, a novel multi-axes bias metric with explainability, to estimate and explain how much bias exists in these datasets. Multilingual, multi-axes bias evaluation is not very common. Hence, we also contribute a new, large Swedish bias-labelled dataset (of 2 million samples), translated from the English version and train the SotA mT5 model on it. In addition, we contribute new multi-axes lexica for bias detection in Swedish. We make the codes, model, and new dataset publicly available.
Comments: Accepted at RANLP 2023
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2301.12139 [cs.CL]
  (or arXiv:2301.12139v3 [cs.CL] for this version)

Submission history

From: Tosin Adewumi [view email]
[v1] Sat, 28 Jan 2023 09:28:19 GMT (1658kb,D)
[v2] Mon, 3 Jul 2023 14:00:36 GMT (1803kb,D)
[v3] Sat, 16 Sep 2023 15:56:11 GMT (1803kb,D)

Link back to: arXiv, form interface, contact.