Assessing Annotator Identity Sensitivity via Item Response Theory: A Case Study in a Hate Speech Corpus

Sachdeva, Pratik S.; Barreto, Renata Rezende; Vacano, Claudia von; Kennedy, Chris

doi:10.1145/3531146.3533216

Cited by 8 publications

(5 citation statements)

References 34 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Details on the development of the hate speech measure are documented [ 25 ] in previous research. The training dataset, consisting of 50,000 social media comments sourced from YouTube, Twitter, and Reddit, was labeled by 10,000 US-based Amazon Mechanical Turk [ 26 ] workers on those components of hate speech (the dataset is available at ) (accessed on 21 December 2022) [ 27 ]. Amazon Mechanical Turk [ 26 ] is a crowdsourcing marketplace where tasks can be performed.…”

Section: Methodsmentioning

confidence: 99%

Examining Exposure to Messaging, Content, and Hate Speech from Partisan News Social Media Posts on Racial and Ethnic Health Disparities

Nguyen

Merchant

et al. 2023

IJERPH

View full text Add to dashboard Cite

We investigated the content of liberal and conservative news media Facebook posts on race and ethnic health disparities. A total of 3,327,360 liberal and conservative news Facebook posts from the United States (US) from January 2015 to May 2022 were collected from the Crowd Tangle platform and filtered for race and health-related keywords. Qualitative content analysis was conducted on a random sample of 1750 liberal and 1750 conservative posts. Posts were analyzed for a continuum of hate speech using a newly developed method combining faceted Rasch item response theory with deep learning. Across posts referencing Asian, Black, Latinx, Middle Eastern, and immigrants/refugees, liberal news posts had lower hate scores compared to conservative posts. Liberal news posts were more likely to acknowledge and detail the existence of racial/ethnic health disparities, while conservative news posts were more likely to highlight the negative consequences of protests, immigration, and the disenfranchisement of Whites. Facebook posts from liberal and conservative news focus on different themes with fewer discussions of racial inequities in conservative news. Investigating the discourse on race and health in social media news posts may inform our understanding of the public’s exposure to and knowledge of racial health disparities, and policy-level support for ameliorating these disparities.

show abstract

Section: Methodsmentioning

confidence: 99%

Examining Exposure to Messaging, Content, and Hate Speech from Partisan News Social Media Posts on Racial and Ethnic Health Disparities

Nguyen

Merchant

et al. 2023

IJERPH

View full text Add to dashboard Cite

show abstract

“…Similarly, disagreement convolution (Gordon et al 2021) incorporates disagreement fully into the evaluation pipeline, showing that on natural language tasks related to toxicity and misinformation the performance of traditional ML models is often overstated. Furthermore, Basile (2021), Rizos and Schuller (2020) and Sachdeva et al (2022b) showed an additional advantage of strong perspectivism in supervised learning, namely its potential impact on the interpretability and fairness of the models. In experiments on real data ( annotated corpora of hate speech), Basile (2021) showed how individual labels can be used to cluster the raters by affinity, leading to the emergence of patterns that helps identifying sociodemographic aspects of the raters themselves, which are in principle opaque, especially in a crowdsourcing scenario, while Sachdeva et al (2022b) showed how the same information could be useful to assess annotator identity sensitivity and thus identify biases in annotation patterns; also Rizos and Schuller (2020) described how a similar approach could be used to detect biases in the data and labels provided by raters.…”

Section: Review Of Perspectivist Approaches In Aimentioning

confidence: 99%

“…Furthermore, Basile (2021), Rizos and Schuller (2020) and Sachdeva et al (2022b) showed an additional advantage of strong perspectivism in supervised learning, namely its potential impact on the interpretability and fairness of the models. In experiments on real data ( annotated corpora of hate speech), Basile (2021) showed how individual labels can be used to cluster the raters by affinity, leading to the emergence of patterns that helps identifying sociodemographic aspects of the raters themselves, which are in principle opaque, especially in a crowdsourcing scenario, while Sachdeva et al (2022b) showed how the same information could be useful to assess annotator identity sensitivity and thus identify biases in annotation patterns; also Rizos and Schuller (2020) described how a similar approach could be used to detect biases in the data and labels provided by raters. Far from being an exhausted topic, the discussion over perspectivism has recently been fostered, among other venues, at international workshops such as Investigating and Mitigating Biases in Crowdsourced Data 5 (ACM CSCW 2021) and the 1st Workshop on Perspectivist Approaches to NLP 6 .…”

Section: Review Of Perspectivist Approaches In Aimentioning

confidence: 99%

Toward a Perspectivist Turn in Ground Truthing for Predictive Computing

Basile¹,

Cabitza²,

Campagner³

et al. 2023

AAAI

View full text Add to dashboard Cite

Most current Artificial Intelligence applications are based on supervised Machine Learning (ML), which ultimately grounds on data annotated by small teams of experts or large ensemble of volunteers. The annotation process is often performed in terms of a majority vote, however this has been proved to be often problematic by recent evaluation studies. In this article, we describe and advocate for a different paradigm, which we call perspectivism: this counters the removal of disagreement and, consequently, the assumption of correctness of traditionally aggregated gold-standard datasets, and proposes the adoption of methods that preserve divergence of opinions and integrate multiple perspectives in the ground truthing process of ML development. Drawing on previous works which inspired it, mainly from the crowdsourcing and multi-rater labeling settings, we survey the state-of-the-art and describe the potential of our proposal for not only the more subjective tasks (e.g. those related to human language) but also those tasks commonly understood as objective (e.g. medical decision making). We present the main benefits of adopting a perspectivist stance in ML, as well as possible disadvantages, and various ways in which such a stance can be implemented in practice. Finally, we share a set of recommendations and outline a research agenda to advance the perspectivist stance in ML.

show abstract

“…Biased annotators could inadvertently induce their implicit prejudices into the labels (Sap et al, 2019;Davidson et al, 2019;Sachdeva et al, 2022). If machine learning models are trained on such biased data, they potentially amplify these biases and generate harm by inference (Shah et al, 2020).…”

Section: Cognitive Reflection Testmentioning

confidence: 99%

“…(3) Biased and noisy labels: Manual annotation is a predominant approach for constructing labeled data. However, the labeling process could reflect inherent biases of human annotators, and it is hard to control the label quality (Sap et al, 2019;Sachdeva et al, 2022). Training a model on such a noisy and biased dataset can amplify biases by inference (Shah et al, 2020).…”

Section: Introductionmentioning

confidence: 99%

Acceptance and Metamorphosis of New Culture at Soongsil School in Pyongyang during the Modern Transition Period - Introduction and Spread of Modern Western Music and Sports -

Park¹,

부교수²

2020

The Historical Association for Soong-Sil

View full text Add to dashboard Cite

Numerous datasets have been proposed to combat the spread of online hate. Despite these efforts, a majority of these resources are Englishcentric, primarily focusing on overt forms of hate. This research gap calls for developing high-quality corpora in diverse languages that also encapsulate more subtle hate expressions. This study introduces K-HATERS, a new corpus for hate speech detection in Korean, comprising approximately 192K news comments with target-specific offensiveness ratings. This resource is the largest offensive language corpus in Korean and is the first to offer targetspecific ratings on a three-point Likert scale, enabling the detection of hate expressions in Korean across varying degrees of offensiveness. We conduct experiments showing the effectiveness of the proposed corpus, including a comparison with existing datasets. Additionally, to address potential noise and bias in human annotations, we explore a novel idea of adopting the Cognitive Reflection Test, which is widely used in social science for assessing an individual's cognitive ability, as a proxy of labeling quality. Findings indicate that annotations from individuals with the lowest test scores tend to yield detection models that make biased predictions toward specific target groups and are less accurate. This study contributes to the NLP research on hate speech detection and resource construction. The code and dataset can be accessed at https://github. com/ssu-humane/K-HATERS.

show abstract

Assessing Annotator Identity Sensitivity via Item Response Theory: A Case Study in a Hate Speech Corpus

Cited by 8 publications

References 34 publications

Examining Exposure to Messaging, Content, and Hate Speech from Partisan News Social Media Posts on Racial and Ethnic Health Disparities

Examining Exposure to Messaging, Content, and Hate Speech from Partisan News Social Media Posts on Racial and Ethnic Health Disparities

Toward a Perspectivist Turn in Ground Truthing for Predictive Computing

Acceptance and Metamorphosis of New Culture at Soongsil School in Pyongyang during the Modern Transition Period - Introduction and Spread of Modern Western Music and Sports -

Contact Info

Product

Resources

About