Research to tackle hate speech plaguing online media has made strides in providing solutions, analyzing bias and curating data. A challenging problem is ambiguity between hate speech and offensive language, causing low performance both overall and specifically for the hate speech class. It can be argued that misclassifying actual hate speech content as merely offensive can lead to further harm against targeted groups. In our work, we mitigate this potentially harmful phenomenon by proposing an adversarial debiasing method to separate the two classes. We show that our method works for English, Arabic German and Hindi, plus in a multilingual setting, improving performance over baselines.

With the widespread use of social media, online hate is increasing, and microaggressions are receiving attention. We explore the potential for using pretrained language models to automatically generate messages that combat the associated offensive texts. Specifically, we focus on using prompting to steer model generation as it requires less data and computation than fine-tuning. We also propose a human evaluation perspective; offensiveness, stance, and informativeness. After obtaining 306 counterspeech and 42 microintervention messages generated by GPT-2, 3, Neo, we conducted a human evaluation using Amazon Mechanical Turk. The results indicate the potential of using prompting in the proposed generation task. All the generated texts along with the annotation are published to encourage future research on countering hate and microaggressions online.

Digital harms can manifest across any interface. Key problems in addressing these harms include the high individuality of harms and the fast-changing nature of digital systems. We put forth GreaseVision, a collaborative human-in-the-loop learning framework that enables end-users to analyze their screenomes to annotate harms as well as render overlay interventions. We evaluate HITL intervention development with a set of completed tasks in a cognitive walkthrough, and test scalability with one-shot element removal and fine-tuning hate speech classification models. The contribution of the framework and tool allow individual end-users to study their usage history and create personalized interventions. Our contribution also enables researchers to study the distribution of multi-modal harms and interventions at scale.

Despite recent advances in machine learning based hate speech detection, classifiers still struggle with generalizing knowledge to out-of-domain data samples. In this paper, we investigate the generalization capabilities of deep learning models to different target groups of hate speech under clean experimental settings. Furthermore, we assess the efficacy of three different strategies of unsupervised domain adaptation to improve these capabilities. Given the diversity of hate and its rapid dynamics in the online world (e.g. the evolution of new target groups like virologists during the COVID-19 pandemic), robustly detecting hate aimed at newly identified target groups is a highly relevant research question. We show that naively trained models suffer from a target group specific bias, which can be reduced via domain adaptation. We were able to achieve a relative improvement of the F1-score between 5.8% and 10.7% for out-of-domain target groups of hate speech compared to baseline approaches by utilizing domain adaptation.

This paper presents a comprehensive corpus for the study of socially unacceptable language in Dutch. The corpus extends and revise an existing resource with more data and introduces a new annotation dimension for offensive language, making it a unique resource in the Dutch language panorama. Each language phenomenon (abusive and offensive language) in the corpus has been annotated with a multi-layer annotation scheme modelling the explicitness and the target(s) of the message. We have conducted a new set of experiments with different classification algorithms on all annotation dimensions. Monolingual Pre-Trained Language Models prove as the best systems, obtaining a macro-average F1 of 0.828 for binary classification of offensive language, and 0.579 for the targets of offensive messages. Furthermore, the best system obtains a macro-average F1 of 0.667 for distinguishing between abusive and offensive messages.

This work describes the process of creating a corpus of Twitter conversations annotated for the presence of counterspeech in response to toxic speech related to axes of discrimination linked to sexism, racism and homophobia. The main novelty is an annotated dataset comprising relevant tweets in their context of occurrence. The corpus is made up of tweets and responses captured by different profiles replying to discriminatory content or objectionably couched news. An annotation scheme was created to make explicit the knowledge on the dimensions of toxic speech and counterspeech.An analysis of the collected and annotated data and of the IAA that emerged during the annotation process is included. Moreover, we report about preliminary experiments on automatic counterspeech detection, based on supervised automatic learning models trained on the new dataset. The results highlight the fundamental role played by the context in this detection task, confirming our intuitions about the importance to collect tweets in their context of occurrence.

Analyzing ethnic or religious bias is important for improving fairness, accountability, and transparency of natural language processing models. However, many techniques rely on human-compiled lists of bias terms, which are expensive to create and are limited in coverage. In this study, we present a fully data-driven pipeline for generating a knowledge graph (KG) of cultural knowledge and stereotypes. Our resulting KG covers 5 religious groups and 5 nationalities and can easily be extended to more entities. Our human evaluation shows that the majority (59.2%) of non-singleton entries are coherent and complete stereotypes. We further show that performing intermediate masked language model training on the verbalized KG leads to a higher level of cultural awareness in the model and has the potential to increase classification performance on knowledge-crucial samples on a related task, i.e., hate speech detection.

Toxic language can take many forms, from explicit hate speech to more subtle microaggressions. Within this space, models identifying transphobic language have largely focused on overt forms. However, a more pernicious and subtle source of transphobic comments comes in the form of statements made by Trans-exclusionary Radical Feminists (TERFs); these statements often appear seemingly-positive and promote women’s causes and issues, while simultaneously denying the inclusion of transgender women as women. Here, we introduce two models to mitigate this antisocial behavior. The first model identifies TERF users in social media, recognizing that these users are a main source of transphobic material that enters mainstream discussion and whom other users may not desire to engage with in good faith. The second model tackles the harder task of recognizing the masked rhetoric of TERF messages and introduces a new dataset to support this task. Finally, we discuss the ethics of deploying these models to mitigate the harm of this language, arguing for a balanced approach that allows for restorative interactions.

In an era of increasingly large pre-trained language models, knowledge distillation is a powerful tool for transferring information from a large model to a smaller one. In particular, distillation is of tremendous benefit when it comes to real-world constraints such as serving latency or serving at scale. However, a loss of robustness in language understanding may be hidden in the process and not immediately revealed when looking at high-level evaluation metrics. In this work, we investigate the hidden costs: what is “lost in distillation”, especially in regards to identity-based model bias using the case study of toxicity modeling. With reproducible models using open source training sets, we investigate models distilled from a BERT teacher baseline. Using both open source and proprietary big data models, we investigate these hidden performance costs.

We present a cleansed version of the multilingual lexicon HURTLEX-(EL) comprising 737 offensive words of Modern Greek. We worked bottom-up in two annotation rounds and developed detailed guidelines by cross-classifying words on three dimensions: context, reference, and thematic domain. Our classification reveals a wider spectrum of thematic domains concerning the study of offensive language than previously thought Efthymiou et al. (2014) and reveals social and cultural aspects that are not included in the HURTLEX categories.

Social platforms such as Gab and Parler, branded as ‘free-speech’ networks, have seen a significant growth of their user base in recent years. This popularity is mainly attributed to the stricter moderation enforced by mainstream platforms such as Twitter, Facebook, and Reddit.In this work we provide the first large scale analysis of hate-speech on Parler. We experiment with an array of algorithms for hate-speech detection, demonstrating limitations of transfer learning in that domain, given the illusive and ever changing nature of the ways hate-speech is delivered. In order to improve classification accuracy we annotated 10K Parler posts, which we use to fine-tune a BERT classifier. Classification of individual posts is then leveraged for the classification of millions of users via label propagation over the social network. Classifying users by their propensity to disseminate hate, we find that hate mongers make 16.1% of Parler active users, and that they have distinct characteristics comparing to other user groups. We further complement our analysis by comparing the trends observed in Parler to those found in Gab. To the best of our knowledge, this is among the first works to analyze hate speech in Parler in a quantitative manner and on the user level.

Most of the published approaches and resources for hate speech detection are tailored for the English language. In consequence, cross-lingual and cross-cultural perspectives lack some essential resources. The lack of diversity of the datasets in Spanish is notable. Variations throughout Spanish-speaking countries make existing datasets not enough to encompass the task in the different Spanish variants. We annotated 9834 tweets from Chile to enrich the existing Spanish resources with different words and new targets of hate that have not been considered in previous studies. We conducted several cross-dataset evaluation experiments of the models published in the literature using our Chilean dataset and two others in English and Spanish. We propose a comparative framework for quickly conducting comparative experiments using different previously published models. In addition, we set up a Codalab competition for further comparison of new models in a standard scenario, that is, data partitions and evaluation metrics. All resources can be accessed trough a centralized repository for researchers to get a complete picture of the progress on the multilingual hate speech and offensive language detection task.

Uses of pejorative expressions can be benign or actively empowering. When models for abuse detection misclassify these expressions as derogatory, they inadvertently censor productive conversations held by marginalized groups. One way to engage with non-dominant perspectives is to add context around conversations. Previous research has leveraged user- and thread-level features, but it often neglects the spaces within which productive conversations take place. Our paper highlights how community context can improve classification outcomes in abusive language detection. We make two main contributions to this end. First, we demonstrate that online communities cluster by the nature of their support towards victims of abuse. Second, we establish how community context improves accuracy and reduces the false positive rates of state-of-the-art abusive language classifiers. These findings suggest a promising direction for context-aware models in abusive language research.

In this work, we present a new publicly available offensive language dataset of 10.278 German social media comments collected in the first half of 2021 that were annotated by in total six annotators. With twelve different annotation categories, it is far more comprehensive than other datasets, and goes beyond just hate speech detection. The labels aim in particular also at toxicity, criminal relevance and discrimination types of comments. Furthermore, about half of the comments are from coherent parts of conversations, which opens the possibility to consider the comments’ contexts and do conversation analyses in order to research the contagion of offensive language in conversations.

Hate speech detection models are typically evaluated on held-out test sets. However, this risks painting an incomplete and potentially misleading picture of model performance because of increasingly well-documented systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, recent research has thus introduced functional tests for hate speech detection models. However, these tests currently only exist for English-language content, which means that they cannot support the development of more effective models in other languages spoken by billions across the world. To help address this issue, we introduce Multilingual HateCheck (MHC), a suite of functional tests for multilingual hate speech detection models. MHC covers 34 functionalities across ten languages, which is more languages than any other hate speech dataset. To illustrate MHC’s utility, we train and test a high-performing multilingual hate speech detection model, and reveal critical model weaknesses for monolingual and cross-lingual applications.

“Dogwhistles” are expressions intended by the speaker have two messages: a socially-unacceptable “in-group” message understood by a subset of listeners, and a benign message intended for the out-group. We take the result of a word-replacement survey of the Swedish population intended to reveal how dogwhistles are understood, and we show that the difficulty of annotating dogwhistles is reflected in the separability in the space of a sentence-transformer Swedish BERT trained on general data.

The subjectivity of automatic hate speech detection makes it a complex task, reflected in different and incomplete definitions in NLP. We present hate speech criteria, developed with insights from a law and social science expert, that help researchers create more explicit definitions and annotation guidelines on five aspects: (1) target groups and (2) dominance, (3) perpetrator characteristics, (4) explicit presence of negative interactions, and the (5) type of consequences/effects. Definitions can be structured so that they cover a more broad or more narrow phenomenon and conscious choices can be made on specifying criteria or leaving them open. We argue that the goal and exact task developers have in mind should determine how the scope of hate speech is defined. We provide an overview of the properties of datasets from hatespeechdata.com that may help select the most suitable dataset for a specific scenario.

Tasks such as toxicity detection, hate speech detection, and online harassment detection have been developed for identifying interactions involving offensive speech. In this work we articulate the need for a relational understanding of offensiveness to help distinguish denotative offensive speech from offensive speech serving as a mechanism through which marginalized communities resist oppressive social norms. Using examples from the queer community, we argue that evaluations of offensive speech must focus on the impacts of language use. We call this the cynic perspective– or a characteristic of language with roots in Cynic philosophy that pertains to employing offensive speech as a practice of resistance. We also explore the degree to which NLP systems may encounter limits to modeling relational context.

Online messaging is dynamic, influential, and highly contextual, and a single post may contain contrasting sentiments towards multiple entities, such as dehumanizing one actor while empathizing with another in the same message. These complexities are important to capture for understanding the systematic abuse voiced within an online community, or for determining whether individuals are advocating for abuse, opposing abuse, or simply reporting abuse. In this work, we describe a formulation of directed social regard (DSR) as a problem of multi-entity aspect-based sentiment analysis (ME-ABSA), which models the degree of intensity of multiple sentiments that are associated with entities described by a text document. Our DSR schema is informed by Bandura’s psychosocial theory of moral disengagement and by recent work in ABSA. We present a dataset of over 2,900 posts and sentences, comprising over 24,000 entities annotated for DSR over nine psychosocial dimensions by three annotators. We present a novel transformer-based ME-ABSA model for DSR, achieving favorable preliminary results on this dataset.

A common approach for testing fairness issues in text-based classifiers is through the use of counterfactuals: does the classifier output change if a sensitive attribute in the input is changed? Existing counterfactual generation methods typically rely on wordlists or templates, producing simple counterfactuals that fail to take into account grammar, context, or subtle sensitive attribute references, and could miss issues that the wordlist creators had not considered. In this paper, we introduce a task for generating counterfactuals that overcomes these shortcomings, and demonstrate how large language models (LLMs) can be leveraged to accomplish this task. We show that this LLM-based method can produce complex counterfactuals that existing methods cannot, comparing the performance of various counterfactual generation methods on the Civil Comments dataset and showing their value in evaluating a toxicity classifier.

Romania ranks almost last in Europe when it comes to gender equality in political representation, with about 10% fewer women in politics than the E.U. average. We proceed from the assumption that this underrepresentation is also influenced by the sexism and verbal abuse female politicians face in the public sphere, especially in online media. We collect a novel dataset with sexist comments in Romanian language from newspaper articles about Romanian female politicians and propose baseline models using classical machine learning models and fine-tuned pretrained transformer models for the classification of sexist language in the online medium.

The past decade has seen an abundance of work seeking to detect, characterize, and measure online hate speech. A related, but less studied problem, is the detection of identity groups targeted by that hate speech. Predictive accuracy on this task can supplement additional analyses beyond hate speech detection, motivating its study. Using the Measuring Hate Speech corpus, which provided annotations for targeted identity groups, we created neural network models to perform multi-label binary prediction of identity groups targeted by a comment. Specifically, we studied 8 broad identity groups and 12 identity sub-groups within race and gender identity. We found that these networks exhibited good predictive performance, achieving ROC AUCs of greater than 0.9 and PR AUCs of greater than 0.7 on several identity groups. We validated their performance on HateCheck and Gab Hate Corpora, finding that predictive performance generalized in most settings. We additionally examined the performance of the model on comments targeting multiple identity groups. Our results demonstrate the feasibility of simultaneously identifying targeted groups in social media comments.

Lexicons play an important role in content moderation often being the first line of defense. However, little or no literature exists in analyzing the representation of queer-related words in them. In this paper, we consider twelve well-known lexicons containing inappropriate words and analyze how gender and sexual minorities are represented in these lexicons. Our analyses reveal that several of these lexicons barely make any distinction between pejorative and non-pejorative queer-related words. We express concern that such unfettered usage of non-pejorative queer-related words may impact queer presence in mainstream discourse. Our analyses further reveal that the lexicons have poor overlap in queer-related words. We finally present a quantifiable measure of consistency and show that several of these lexicons are not consistent in how they include (or omit) queer-related words.

Online hate speech is a dangerous phenomenon that can (and should) be promptly counteracted properly. While Natural Language Processing supplies appropriate algorithms for trying to reach this objective, all research efforts are directed toward the English language. This strongly limits the classification power on non-English languages. In this paper, we test several learning frameworks for identifying hate speech in Italian text. We release HATE-ITA, a multi-language model trained on a large set of English data and available Italian datasets. HATE-ITA performs better than mono-lingual models and seems to adapt well also on language-specific slurs. We hope our findings will encourage the research in other mid-to-low resource communities and provide a valuable benchmarking tool for the Italian community.