Listening to Affected Communities to Define Extreme Speech: Dataset and Experiments

Maronikolakis, Antonis; Wisiorek, Axel; Nann, Leah; Haris, Jabbar,; Udupa, Sahana; Schuetze, Hinrich

doi:10.18653/v1/2022.findings-acl.87

Cited by 4 publications

(3 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…After the completion of this process, 50% of the annotated passages were cross-annotated by another factchecker to check the interannotator agreement scores. Cohen's kappa (κ, McHugh, 2012), Krippendorff 's alpha (α, Krippendorff, 2011), intraclass correlation coefficient (two-way mixed, average score ICC(3, k) for k = 2; Cicchetti, 1994) and accuracy were measured, which is the percentage of passages where both annotators agreed (Maronikolakis et al, 2022). For the three labels of derogatory, exclusionary and dangerous speech, we obtained the values of κ = 0.23, α = 0.24 and ICC(3, k) = 0.41, which is considered "fair" (Cicchetti, 1994;Maronikolakis et al, 2022).…”

Section: Methods and Datamentioning

confidence: 99%

Online anti-immigrant discourse in Germany: ethnographically backed analysis of user comments

Nann,

Udupa,

Wisiorek

2024

Front. Commun.

View full text Add to dashboard Cite

This article investigates discourse- and language-specific features of online anti-immigrant extreme speech in Germany. We analyze a context rich dataset collected and annotated through a collaborative effort involving fact-checkers, ethnographers and natural language processing (NLP) researchers. Using a bottom-up annotation scheme, we capture the nuances of the discourse and develop a typology of lexical innovations. The analysis combines thematic and critical discourse analysis with a linguistic perspective, revealing that direct forms of racism intertwine with argumentative forms of antagonism and playful word games within anti-immigrant discourses, in ways that center around narratives of victimhood and perceived threats from migrants. We further show how the specific conditions in Germany, including strict legal regulations of speech, have shaped the emergence of a non-standard language variety in and through anti-immigrant discourse, which helps the discourse community maintain group identity. The ethnographically backed analysis provides a granular understanding of the phenomenon of online extreme speech in the German context, contributing to the broader field of discourse studies and offering insights into the varieties within anti-immigrant discourse.

show abstract

Section: Methods and Datamentioning

confidence: 99%

Online anti-immigrant discourse in Germany: ethnographically backed analysis of user comments

Nann,

Udupa,

Wisiorek

2024

Front. Commun.

View full text Add to dashboard Cite

show abstract

“…Noise audit While limited literature exists on investigating the generalizability of offensive speech detection systems across datasets (Arango et al, 2019), political discourse (Grimminger and Klinger, 2021;Maronikolakis et al, 2022), vulnerability to adversarial attacks (Gröndahl et al, 2018), unseen use cases , and geographic biases (Ghosh et al, 2021), to the best of our knowledge, no work exists on a comprehensive, in-the-wild evaluation of offensive speech filtering outcomes on large-scale, real-world political discussions. One key impediment to performing in-the-wild analysis of content moderation systems is a lack of ground truth.…”

Section: Definitionsmentioning

confidence: 99%

Vicarious Offense and Noise Audit of Offensive Speech Classifiers: Unifying Human and Machine Disagreement on What is Offensive

Weerasooriya,

Dutta,

Ranasinghe

et al. 2023

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

View full text Add to dashboard Cite

This paper discusses and contains content that is offensive or disturbing. Offensive speech detection is a key component of content moderation. However, what is offensive can be highly subjective. This paper investigates how machine and human moderators disagree on what is offensive when it comes to real-world social web political discourse. We show that (1) there is extensive disagreement among the moderators (humans and machines); and (2) human and large-language-model classifiers are unable to predict how other human raters will respond, based on their political leanings. For (1), we conduct a noise audit at an unprecedented scale that combines both machine and human responses. For (2), we introduce a firstof-its-kind dataset 1 of vicarious offense. Our noise audit reveals that moderation outcomes vary wildly across different machine moderators. Our experiments with human moderators suggest that political leanings combined with sensitive issues affect both first-person and vicarious offense.

show abstract

“…XtremeSpeech English [46]: The complete dataset contains 5,180 texts collected from Facebook, Twitter and WhatsApp. The dataset is not yet public, but the authors have kindly shared with us a subset of 2,639 texts written in English that focuses on Kenya as a geographic location.…”

Section: Datamentioning

confidence: 99%

Knowledge-Grounded Target Group Language Recognition in Hate Speech

Reyero Lobo,

Daga,

Alani

et al. 2023

Knowledge Graphs: Semantics, Machine Learning, and Languages

View full text Add to dashboard Cite

Hate speech comes in different forms depending on the communities targeted, often based on factors like gender, sexuality, race, or religion. Detecting it online is challenging because existing systems are not accounting for the diversity of hate based on the identity of the target and may be biased towards certain groups, leading to inaccurate results. Current language models perform well in identifying target communities, but only provide a probability that a hate speech text contains references to a particular group. This lack of transparency is problematic because these models learn biases from data annotated by individuals who may not be familiar with the target group. To improve hate speech detection, particularly target group identification, we propose a new hybrid approach that incorporates explicit knowledge about the language used by specific identity groups. We leverage a Knowledge Graph (KG) and adapt it, considering an appropriate level of abstraction, to recognise hate speech-language related to gender and sexual orientation. A thorough quantitative and qualitative evaluation demonstrates that our approach is as effective as state-of-the-art language models while adjusting better to domain and data changes. By grounding the task in explicit knowledge, we can better contextualise the results generated by our proposed approach with the language of the groups most frequently impacted by these technologies. Semantic enrichment helps us examine model outcomes and the training data used for hate speech detection systems, and handle ambiguous cases in human annotations more effectively. Overall, infusing semantic knowledge in hate speech detection is crucial for enhancing understanding of model behaviors and addressing biases derived from training data.

show abstract

Listening to Affected Communities to Define Extreme Speech: Dataset and Experiments

Cited by 4 publications

References 24 publications

Online anti-immigrant discourse in Germany: ethnographically backed analysis of user comments

Online anti-immigrant discourse in Germany: ethnographically backed analysis of user comments

Vicarious Offense and Noise Audit of Offensive Speech Classifiers: Unifying Human and Machine Disagreement on What is Offensive

Knowledge-Grounded Target Group Language Recognition in Hate Speech

Contact Info

Product

Resources

About