The BLP-2025 Task 1A requires Bengali hate speech classification into six categories. Traditional supervised approaches need extensive labeled datasets that are expensive for low-resource languages. We developed PromptGuard, a few-shot framework combining chi-square statistical analysis for keyword extraction with adaptive majority voting for decision-making. We explore statistical keyword selection versus random approaches and adaptive voting mechanisms that extend classification based on consensus quality. Chi-square keywords provide consistent improvements across categories, while adaptive voting benefits ambiguous cases requiring extended classification rounds. PromptGuard achieves a micro-F1 of 67.61, outperforming n-gram baselines (60.75) and random approaches (14.65). Ablation studies confirm chi-square-based keywords show the most consistent impact across all categories.

A key value proposition of machine learning is generalizability: the same methods and model architecture should be able to work across different domains and different contexts. While powerful, this generalization can sometimes go too far, and miss the importance of the specifics. In this work, we look at how fair machine learning has often treated as interchangeable the identity axis along which discrimination occurs. In other words, racism is measured and mitigated the same way as sexism, as ableism, as ageism. Disciplines outside of computer science have pointed out both the similarities and differences between these different forms of oppression, and in this work we draw out the implications for fair machine learning. While certainly not all aspects of fair machine learning need to be tailored to the specific form of oppression, there is a pressing need for greater attention to such specificity than is currently evident. Ultimately, context specificity can deepen our understanding of how to build more fair systems, widen our scope to include currently overlooked harms, and, almost paradoxically, also help to narrow our scope and counter the fear of an infinite number of group-specific methods of analysis.

The use of Large Language Models (LLMs) has proven to be a tool that could help in the automatic detection of sexism. Previous studies have shown that these models contain biases that do not accurately reflect reality, especially for minority groups. Despite various efforts to improve the detection of sexist content, this task remains a significant challenge due to its subjective nature and the biases present in automated models. We explore the capabilities of different LLMs to detect sexism in social media text using the EXIST 2024 tweet dataset. It includes annotations from six distinct profiles for each tweet, allowing us to evaluate to what extent LLMs can mimic these groups' perceptions in sexism detection. Additionally, we analyze the demographic biases present in the models and conduct a statistical analysis to identify which demographic characteristics (age, gender) contribute most effectively to this task. Our results show that, while LLMs can to some extent detect sexism when considering the overall opinion of populations, they do not accurately replicate the diversity of perceptions among different demographic groups. This highlights the need for better-calibrated models that account for the diversity of perspectives across different populations.

Sexism is generally defined as prejudice and discrimination based on sex or gender, affecting every sector of society, from social institutions to relationships and individual behavior. Social media platforms amplify the impact of sexism by conveying discriminatory content not only through text but also across multiple modalities, highlighting the critical need for a multimodal approach to the analysis of sexism online. With the rise of social media platforms where users share short videos, sexism is increasingly spreading through video content. Automatically detecting sexism in videos is a challenging task, as it requires analyzing the combination of verbal, audio, and visual elements to identify sexist content. In this study, (1) we introduce MuSeD, a new Multimodal Spanish dataset for Sexism Detection consisting of $\approx$ 11 hours of videos extracted from TikTok and BitChute; (2) we propose an innovative annotation framework for analyzing the contributions of textual, vocal, and visual modalities to the classification of content as either sexist or non-sexist; and (3) we evaluate a range of large language models (LLMs) and multimodal LLMs on the task of sexism detection. We find that visual information plays a key role in labeling sexist content for both humans and models. Models effectively detect explicit sexism; however, they struggle with implicit cases, such as stereotypes, instances where annotators also show low agreement. This highlights the inherent difficulty of the task, as identifying implicit sexism depends on the social and cultural context.

Anti-sexist speech, i.e., public expressions that challenge or resist gendered abuse and sexism, plays a vital role in shaping democratic debate online. Yet automated content moderation systems, increasingly powered by large language models (LLMs), may struggle to distinguish such resistance from the sexism it opposes. This study examines how five LLMs classify sexist, anti-sexist, and neutral political tweets from the UK, focusing on high-salience trigger events involving female Members of Parliament in the year 2022. Our analysis show that models frequently misclassify anti-sexist speech as harmful, particularly during politically charged events where rhetorical styles of harm and resistance converge. These errors risk silencing those who challenge sexism, with disproportionate consequences for marginalised voices. We argue that moderation design must move beyond binary harmful/not-harmful schemas, integrate human-in-the-loop review during sensitive events, and explicitly include counter-speech in training data. By linking feminist scholarship, event-based analysis, and model evaluation, this work highlights the sociotechnical challenges of safeguarding resistance speech in digital political spaces.

The global reach of social media has amplified the spread of hateful content, including implicit sexism, which is often overlooked by conventional detection methods. In this work, we introduce an Adaptive Supervised Contrastive lEarning framework for implicit sexism detectioN (ASCEND). A key innovation of our method is the incorporation of threshold-based contrastive learning: by computing cosine similarities between embeddings, we selectively treat only those sample pairs as positive if their similarity exceeds a learnable threshold. This mechanism refines the embedding space by robustly pulling together representations of semantically similar texts while pushing apart dissimilar ones, thus reducing false positives and negatives. The final classification is achieved by jointly optimizing a contrastive loss with a cross-entropy loss. Textual features are enhanced through a word-level attention module. Additionally, we employ sentiment, emotion, and toxicity features. Evaluations on the EXIST2021 and MLSC datasets demonstrate that ASCEND significantly outperforms existing methods, with average Macro F1 improvements of 9.86%, 29.63%, and 32.51% across multiple tasks, highlighting its efficacy in capturing the subtle cues of implicit sexist language.

Detecting deviant language such as sexism, or nuanced language such as metaphors or sarcasm, is crucial for enhancing the safety, clarity, and interpretation of online social discourse. While existing classifiers deliver strong results on these tasks, they often come with significant computational cost and high data demands. In this work, we propose \textbf{Cla}ss \textbf{D}istillation (ClaD), a novel training paradigm that targets the core challenge: distilling a small, well-defined target class from a highly diverse and heterogeneous background. ClaD integrates two key innovations: (i) a loss function informed by the structural properties of class distributions, based on Mahalanobis distance, and (ii) an interpretable decision algorithm optimized for class separation. Across three benchmark detection tasks -- sexism, metaphor, and sarcasm -- ClaD outperforms competitive baselines, and even with smaller language models and orders of magnitude fewer parameters, achieves performance comparable to several large language models (LLMs). These results demonstrate ClaD as an efficient tool for pragmatic language understanding tasks that require gleaning a small target class from a larger heterogeneous background.

This paper investigates hybrid intelligence and collaboration between researchers of sexism and Large Language Models (LLMs), with a four-component pipeline. First, nine sexism researchers answer questions about their knowledge of sexism and of LLMs. They then participate in two interactive experiments involving an LLM (GPT3.5). The first experiment has experts assessing the model's knowledge about sexism and suitability for use in research. The second experiment tasks them with creating three different definitions of sexism: an expert-written definition, an LLM-written one, and a co-created definition. Lastly, zero-shot classification experiments use the three definitions from each expert in a prompt template for sexism detection, evaluating GPT4o on 2.500 texts sampled from five sexism benchmarks. We then analyze the resulting 67.500 classification decisions. The LLM interactions lead to longer and more complex definitions of sexism. Expert-written definitions on average perform poorly compared to LLM-generated definitions. However, some experts do improve classification performance with their co-created definitions of sexism, also experts who are inexperienced in using LLMs.

-- Online hate speech can harmfully impact individuals and groups, specifically on non - moderated platforms such as 4chan where users can post anonymous content. This work focuses on analy s ing and measuring the prevalence of online hat e on 4chan's politically incorrect board (/pol/) using state - of - the - art Natural Language Processing (NLP) models, specifically transformer - based models such as RoBERTa and Detoxify . By leveraging these advanced models, we provide an in - depth analysis of hate speech dynamics and quantify the extent of online hate non - moderated platforms. The study advances understanding through multi - class classification of hate speech (racism, sexism, religion, etc.), while also incorporating the classification of toxic content (e.g., identity attacks and threats) and a further topic modelling analysis. The results show that 11.20% of this dataset is identified as containing hate in different categories. These evaluations show that online hate is manifested in various forms, confirming the complicated and volatile nature of detection in the wild. Index Terms -- Hate speech, machine learning, natural language processing (NLP), online hate, toxicity analysis. INTRODUCTION H E SPREAD of hate speech on online platforms has become a serious problem in our society. As digital communication becomes ubiquitous, platforms like 4chan, known for their anonymity and minimal moderation, have become hotspots for this harmful behaviour . This is particularly evident on its politically incorrect board, /pol/, a notorious board dedicated to discussing politics and current events, often associated with hate speech, extremist content, and conspiracy theories [1] . The anonymity provided by these platforms often encourages users to express extreme ideologies [2] . This issue raises significant concerns about the impact on at - risk and vulnerable groups as it can cause real - world harm, including psychological trauma. Therefore, a systematic approach is needed to measure and understand the prevalence and forms of online hate. Received 28 August 2024; revised 23 December 2024, 10 February 2025, and 6 March 2025; accepted 6 March 2025. This work is supported by the UK Research and Innovation (UKRI), through the Strategic Priority Fund as part of the Protecting Citizens Online programme (AGENCY: Assuring Citizen Agency in a World with Complex Online Harms, EP/W032481/2).

Resolving disagreement in manual annotation typically consists of removing unreliable annotators and using a label aggregation strategy such as majority vote or expert opinion to resolve disagreement. These may have the side-effect of silencing or under-representing minority but equally valid opinions. In this paper, we study the impact of standard label aggregation strategies on minority opinion representation in sexism detection. We investigate the quality and value of minority annotations, and then examine their effect on the class distributions in gold labels, as well as how this affects the behaviour of models trained on the resulting datasets. Finally, we discuss the potential biases introduced by each method and how they can be amplified by the models.