CoChat
Sign Up
Cited by 25Privacy-Preserving Technologies in DataAdversarial Robustness in Machine LearningDomain Adaptation and Few-Shot Learning

KGA: A General Machine Unlearning Framework Based on Knowledge Gap Alignment

Open access
Wang, Lingzhi, Chen, Tong, Yuan, Wei, Zeng, Xingshan, Wong, Kam-Fai, Yin, Hongzhi·2023·Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
DOI 10.18653/v1/2023.acl-long.740
Open this paper in CoChat — free accountOpen PDF

Abstract

Recent legislation of the "right to be forgotten" has led to the interest in machine unlearning, where the learned models are endowed with the function to forget information about specific training instances as if they have never existed in the training set.Previous work mainly focuses on computer vision scenarios and largely ignores the essentials of unlearning in NLP field, where text data contains more explicit and sensitive personal information than images.In this paper, we propose a general unlearning framework called KGA to induce forgetfulness.Different from previous work that tries to recover gradients or forces models to perform close to one specific distribution, KGA maintains distribution differences (i.e., knowledge gap).This relaxes the distribution assumption.Furthermore, we first apply the unlearning method to various NLP tasks (i.e., classification, translation, response generation) and propose several unlearning evaluation metrics with pertinence.Experiments on large-scale datasets show that KGA yields comprehensive improvements over baselines, where extensive analyses further validate the effectiveness of KGA and provide insight into unlearning for NLP tasks 1 .

Integrity signals

Registry

Crossref · 10.18653/v1/2023.acl-long.740

Cites

20 of 20 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Claims the paper makes

We couldn’t retrieve this paper’s full text, so its claims and self-check weren’t run. The registry and citation checks still apply.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • The main advantage of KGA lies in its generic nature, which allows it to be applied to various models and tasks.

    “The main advan -tage of this method lies in its generic nature, which allows it to be applied to various models and tasks.”

    From Digital forgetting in large language models: a survey of unlearning methods (2025) · cites this paper as Wang et al (2023)

  • A few authors use the success of a Membership Inference Attack (MIA) as a metric.

    “Membership Inference Attack (MIA) : A few authors (Wang et al 2023; Chen and Yang 2023; Pawelczyk et al 2023) use the success of a MIA as a metric.”

    From Digital forgetting in large language models: a survey of unlearning methods (2025) · cites this paper as Wang et al 2023

  • Wang et al (2023) measure output similarity with the text to be forgotten using JSD, LPD, and PDLP, while Yao et al (2023) use BLEU.

    “For instance, Wang et al (2023) measure output similarity with the text to be forgotten by using the Jensen-Shannon Divergence (JSD), the Language model Probability Distance (LPD), and the Proportion of instances with Decreased Language model Probability (PDLP), while Yao et al (2023) leverage the BiLingual Evaluation Understudy (BLEU) score.”

    From Digital forgetting in large language models: a survey of unlearning methods (2025) · cites this paper as Wang et al (2023)

  • Eldan and Russinovich (2023), Wang et al (2023), and Jang et al (2022) leverage well-established LLM benchmarks to measure retained utility.

    “General benchmarks . Eldan and Russinovich (2023), Wang et al (2023), and Jang et al (2022) leverage well-established LLM benchmarks, such as those used for comparing LLMs (Sakaguchi et al 2020; Zellers et al 2019; Paperno et al 2016; Bisk et al 2020).”

    From Digital forgetting in large language models: a survey of unlearning methods (2025) · cites this paper as Wang et al (2023)

  • Reported runtime values of unlearning methods range from minutes to hours or even longer than a day.

    “Reported values range from minutes (Wang et al 2023; Yao et al 2023 ; Limisiewicz and Mareček 2022; Jang et al 2022; Yu et al 2023) to hours (Chen and Yang 2023) or even longer than a day (Lu et al 2022 ).”

    From Digital forgetting in large language models: a survey of unlearning methods (2025) · cites this paper as Wang et al 2023

Cites

Flagged first, then by citations

1 of the 21 references couldn’t be matched to a known paper.

Cited by

Sorted by citations

Related

Authors
Wang, L., Chen, T., Yuan, W., Zeng, X., Wong, K.-F., & Yin, H.
Affiliations
  • Chinese University of Hong Kong (HK)
  • The University of Queensland (AU)
ORCID
Venue
Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Publisher
Association for Computational Linguistics
Published
2023
Pages
13264-13276
ISSN
0736-587X (ISSN-L)
DOI
10.18653/v1/2023.acl-long.740
OpenAlex
W4385571042
Cited by
25
References
21
Citations by year
2026: 10 · 2025: 13 · 2024: 2
FWCI
4.43
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • Proceedings of the conference - Association for Computational Linguistics. Meeting (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • Chinese University of Hong Kong
  • Australian Research Council (FT210100624, DE230101033, DP190101985)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
machine unlearning, right to be forgotten, NLP tasks, text classification, machine translation, response generation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review