Zum Inhalt springen

Nur-Lesen-Fassung: Das Forschungsinformationssystem (FIS) ist aus Sicherheitsgründen vorübergehend durch diese Nur-Lesen-Fassung ersetzt. Alle öffentlichen Inhalte und Volltexte sind weiterhin abrufbar; Anmeldung, Bearbeitung, Einreichung und Statistiken stehen nicht zur Verfügung. Bei Fragen: fis@uni-bamberg.de

Start › Publikationen › Hierarchical Adversarial Correction to Mitigate Identity Term Bias in Toxicity Detection

Publikation

Hierarchical Adversarial Correction to Mitigate Identity Term Bias in Toxicity Detection

Schäfer, Johannes; Heid, Ulrich; Klinger, Roman (2024): Hierarchical Adversarial Correction to Mitigate Identity Term Bias in Toxicity Detection, in: Orphee De Clercq, Valentin Barriere, Jeremy Barnes, u. a. (Hrsg.), Proceedings of the 14th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, Bangkok, Thailand: Association for Computational Linguistics, S. 35–51, doi: 10.18653/v1/2024.wassa-1.4.

Details

Fakultät/Lehrstuhl
Autorin(nen)/Autor(en)
Schäfer, Johannes
Heid, Ulrich
Titel des Sammelwerks
Proceedings of the 14th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis
Herausgeberin(nen)/Herausgeber des Sammelwerks
Clercq, Orphee De
Barriere, Valentin
Barnes, Jeremy
Sedoc, João
Tafreshi, Shabnam
Konferenz
14th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, Bangkok, Thailand
Verlagsangabe
Association for Computational Linguistics : Bangkok, Thailand
Erscheinungsjahr
2024
Sprache(n)
Englisch
DOI
URL
Abstract

Corpora that are the fundament for toxicity detection contain such expressions typically directed against a target individual or group, e.g., people of a specific gender or ethnicity. Prior work has shown that the target identity mention can constitute a confounding variable. As an example, a model might learn that Christians are always mentioned in the context of hate speech. This misguided focus can lead to a limited generalization to newly emerging targets that are not found in the training data. In this paper, we hypothesize and subsequently show that this issue can be mitigated by considering targets on different levels of specificity. We distinguish levels of (1) the existence of a target, (2) a class (e.g., that the target is a religious group), or (3) a specific target group (e.g., Christians or Muslims). We define a target label hierarchy based on these three levels and then exploit this hierarchy in an adversarial correction for the lowest level (i.e. (3)) while maintaining some basic target features. This approach does not lower the toxicity detection performance but increases the generalization to targets not being available at training time.

Freie Schlagwort(e)
Hierarchical Adversarial Correction
Peer Reviewed
ja
Internationale Verbreitung
ja
Open-Access-Zeitschrift
ja
Dokumenttyp
Konferenzveröffentlichung
URI
Datum der Freischaltung
20.09.2024

Permalink: https://fis.uni-bamberg.de/handle/uniba/98197 · Daten als JSON · Zitieren: BibTeX · RIS

Datenstand: 16.11.2026

Eingeschränkter Lesezugriff: Anmeldung, Bearbeitung und Einreichung sind vorübergehend deaktiviert.