Challenges for Toxic Comment Classification: An In-Depth Error Analysis

Aken, Betty van; Risch, Julian; Krestel, Ralf; Löser, Alexander

doi:10.18653/v1/w18-5105

Cited by 167 publications

(109 citation statements)

References 33 publications

Supporting

Mentioning

107

Contrasting

Order By: Relevance

“…Humor, irony and sarcasm. Supposedly humorous, ironic or sarcastic abusive content is often viewed as a source of classification error (Nobata, Thomas, Mehdad, Chang, & Tetreault, 2016;van Aken, Risch, Krestel, & Löser, 2018). However, drawing on critical studies of prejudice and hate, we propose that such content is still abusive (Weaver, 2010).…”

Section: Recognizing Abusive Contentmentioning

confidence: 91%

“…Annotation is a notoriously difficult task, reflected in the low levels of inter-annotator agreement reported by most publications, particularly on more complex multi-class tasks (Sanguinetti, Poletto, Bosco, Patti, & Stranisci, 2018). Noticeably, van Aken suggests that Davidson et al's widely used hate and offensive language dataset has up to 10% of its data mislabeled (van Aken et al, 2018). Few publications provide details of their annotation process or annotation guidelines.…”

Section: Creating and Sharing Datasetsmentioning

confidence: 99%

See 1 more Smart Citation

Challenges and frontiers in abusive content detection

Vidgen¹,

Harris²,

Nguyen³

et al. 2019

Proceedings of the Third Workshop on Abusive Language Online

168

210

View full text Add to dashboard Cite

Online abusive content detection is an inherently difficult task. It has received considerable attention from academia, particularly within the computational linguistics community, and performance appears to have improved as the field has matured. However, considerable challenges and unaddressed frontiers remain, spanning technical, social and ethical dimensions. These issues constrain the performance, efficiency and generalizability of abusive content detection systems. In this article we delineate and clarify the main challenges and frontiers in the field, critically evaluate their implications and discuss solutions. We also highlight ways in which social scientific insights can advance research.

show abstract

Section: Recognizing Abusive Contentmentioning

confidence: 91%

Section: Creating and Sharing Datasetsmentioning

confidence: 99%

Challenges and frontiers in abusive content detection

Vidgen¹,

Harris²,

Nguyen³

et al. 2019

Proceedings of the Third Workshop on Abusive Language Online

168

210

View full text Add to dashboard Cite

show abstract

“…Despite a public effort to recognize and reduce-if not eliminate-their occurrence (Kim, 2013;Neff, 2015), there has been no computational work to detect and analyze MAS at scale. Instead, much of the recent work has focused on explicitly toxic language (e.g., Waseem et al, 2017), with surveys of the area also overlooking this important and challenging task of recognizing this subtle toxicity (van Aken et al, 2018;Salminen et al, 2018;Fortuna and Nunes, 2018). Indeed, as Figure 1 suggests, current popular tools for toxic language detection do not recognize the toxicity of MAS and further, sentiment tools can label these comments as being positive.…”

Section: Introductionmentioning

confidence: 99%

Finding Microaggressions in the Wild: A Case for Locating Elusive Phenomena in Social Media Posts

Breitfeller¹,

Ahn²,

Jurgens

et al. 2019

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conferen

104

View full text Add to dashboard Cite

Microaggressions are subtle, often veiled, manifestations of human biases. These uncivil interactions can have a powerful negative impact on people by marginalizing minorities and disadvantaged groups. The linguistic subtlety of microaggressions in communication has made it difficult for researchers to analyze their exact nature, and to quantify and extract microaggressions automatically. Specifically, the lack of a corpus of real-world microaggressions and well-defined criteria for annotating them have prevented researchers from addressing these problems at scale. In this paper, we devise a general but nuanced, computationally operationalizable typology of microaggressions based on a small subset of microaggression data that we have. We then create two datasets: one with examples of diverse types of microaggressions recollected by their targets, and another with gender-based microaggressions in public conversations on social media. We introduce a new, more objective criterion for annotation and an activelearning based procedure that increases the likelihood of surfacing posts containing microaggressions. Finally, we analyze the trends that emerge from these new datasets.

show abstract

“…Sample size (Seo et al, 2016) 50 (Kundu and Ng, 2018) 50 (Hu et al, 2018) 50 (Min et al, 2018) 50 (Weissenborn et al, 2017) 55 (Chen et al, 2016) 100 (Min et al, 2017) 100 (Wadhwa et al, 2018) 100 (Fader et al, 2013) 100 (van Aken et al, 2018) 200 Average 85.5 Table 3: Surveyed papers and their error sample sizes.…”

Section: Papermentioning

confidence: 99%

Errudite: Scalable, Reproducible, and Testable Error Analysis

Wu¹,

Ribeiro²,

Heer³

et al. 2019

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics

104

View full text Add to dashboard Cite

Though error analysis is crucial to understanding and improving NLP models, the common practice of manual, subjective categorization of a small sample of errors can yield biased and incomplete conclusions. This paper codifies model and task agnostic principles for informative error analysis, and presents Errudite, an interactive tool for better supporting this process. First, error groups should be precisely defined for reproducibility; Errudite supports this with an expressive domainspecific language. Second, to avoid spurious conclusions, a large set of instances should be analyzed, including both positive and negative examples; Errudite enables systematic grouping of relevant instances with filtering queries. Third, hypotheses about the cause of errors should be explicitly tested; Errudite supports this via automated counterfactual rewriting. We validate our approach with a user study, finding that Errudite (1) enables users to perform high quality and reproducible error analyses with less effort, (2) reveals substantial ambiguities in prior published error analyses practices, and (3) enhances the error analysis experience by allowing users to test and revise prior beliefs.

show abstract

Challenges for Toxic Comment Classification: An In-Depth Error Analysis

Cited by 167 publications

References 33 publications

Challenges and frontiers in abusive content detection

Challenges and frontiers in abusive content detection

Finding Microaggressions in the Wild: A Case for Locating Elusive Phenomena in Social Media Posts

Errudite: Scalable, Reproducible, and Testable Error Analysis

Contact Info

Product

Resources

About