Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024)

Yi-Ling Chung, Zeerak Talat, Debora Nozza, Flor Miriam Plaza-del-Arco, Paul Röttger, Aida Mostafazadeh Davani, Agostina CalabreseNAACL 2024

TL;DR

This document comprises the proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024), focusing on the critical challenges posed by online abuse and harm in the age of large language models (LLMs). It highlights the urgent need for computational tools to combat the amplification of harmful content and explores how LLMs can be both a source of new harms and a potential solution for detection and mitigation.

What Problem Does This Paper Address?

The proliferation of digital technologies has brought immense societal benefits, but it has also facilitated the widespread dissemination of abusive and harmful content, such as hate speech and harassment. The sheer volume of online content makes manual moderation infeasible, necessitating the development of sophisticated computational tools. However, effectively detecting and moderating online abuse and harms presents a complex set of technical, social, legal, and ethical challenges. This workshop specifically addresses the emerging threat of "online harms in the age of large language models," where powerful LLMs, if not properly safeguarded, can be easily exploited to generate toxic content or follow malicious instructions, exacerbating existing problems and creating new ones.

Methods

The workshop proceedings showcase a diverse range of research methodologies applied to the problem of online abuse and harms. These include:

  • Natural Language Processing (NLP) techniques: For tasks like hate speech detection, toxicity classification, and identifying radicalization indicators.
  • Machine Learning (ML) algorithms: Including supervised and unsupervised learning for classification, anomaly detection, and developing robust safety classifiers.
  • Data Augmentation: Techniques to improve model performance, especially in data-scarce scenarios.
  • Dataset Creation and Annotation: Development of new datasets for specific languages, types of abuse, or specialized tasks (e.g., visual culture, counterspeech).
  • Interdisciplinary Approaches: Combining insights from computational social sciences, law, politics, psychology, and cultural studies.
  • LLM-specific investigations: Research into prompt engineering, LLM-based toxicity detection, and the ethical implications of LLM deployment.
  • Adaptable and Continual Learning Frameworks: Developing models that can adapt to evolving forms of online harm.

Datasets

The papers presented at WOAH 2024 utilize and introduce a variety of datasets. While a comprehensive list is not provided in the introductory sections, the Table of Contents and individual paper titles indicate the use of:

  • Existing datasets: For various forms of online abuse (hate speech, toxicity, etc.) in different languages.
  • Newly introduced datasets:
    • EkoHate: For code-switched political discussions on Nigerian Twitter.
  • HausaHate: An expert-annotated corpus for Hausa hate speech detection.
  • VIDA: The Visual Incel Data Archive: A dataset for hate detection through visual culture.
  • The Uli Dataset: For experience-led annotation of online gender-based violence (oGBV).
  • AustroTox: A dataset for target-based Austrian German and English offensive language detection.
  • Annotated corpora: For specific tasks like counterspeech and toxicity classification.

Key Insights and Ideas

  • LLMs are a double-edged sword: While LLMs offer immense potential for improving content moderation and developing countermeasures, they also present significant risks of generating and amplifying harmful content if not carefully controlled.
  • The importance of context: Understanding the nuances of language, culture, and specific online communities is crucial for accurate detection of abuse, especially in code-switched or low-resource settings.
  • Data scarcity remains a challenge: Developing effective models for under-resourced languages and specific types of harm requires dedicated efforts in dataset creation and annotation.
  • Interdisciplinary collaboration is vital: Addressing online abuse requires a holistic approach that integrates technical solutions with insights from social sciences, law, and ethics.
  • Interpretable AI is key: Moving beyond black-box models to understand why a model makes a certain prediction is essential for building trust and improving detection systems.
  • The evolution of online harms: New forms of abuse and manipulation are constantly emerging, requiring continuous research and adaptation of detection methods.

Key Results

The proceedings showcase a wide array of research findings, with specific quantitative results detailed within individual papers. However, the introductory sections highlight:

  • Acceptance Rate: 33 out of 56 submissions were accepted for presentation, indicating a competitive selection process.
  • Diverse Research Areas: The accepted papers cover a broad spectrum of topics, from detecting radicalization indicators and conspiracy theories to analyzing cultural biases in hate speech datasets and developing LLM-based safety tools.
  • Focus on LLMs: A significant portion of the research directly addresses the implications of LLMs for online harms, including prompt engineering, toxicity detection, and the generation of harmful content.

Limitations and Future Work

The proceedings, by their nature, represent a snapshot of ongoing research. While specific limitations are detailed within individual papers, the overarching themes suggest several directions for future work:

  • Developing robust LLM safeguards: Continued research is needed to prevent LLMs from generating toxic content or following malicious instructions.
  • Improving LLM-based countermeasures: Exploring how LLMs can be leveraged to enhance content moderation, generate personalized counterspeech, and improve overall online safety.
  • Addressing low-resource languages and contexts: Expanding research and dataset development to cover a wider range of languages and cultural nuances.
  • Enhancing interpretability and explainability: Developing methods to understand the decision-making processes of AI models used for harm detection.
  • Investigating cross-cultural differences: Further research into how online harms manifest and are perceived across different cultural contexts.
  • Real-world deployment and evaluation: Moving beyond lab settings to test and refine solutions in real-world online environments.

Summary updated

BibTeX
@inproceedings{chung2024proceedings,
  author = {Yi-Ling Chung and Zeerak Talat and Debora Nozza and Flor Miriam Plaza-del-Arco and Paul Röttger and Aida Mostafazadeh Davani and Agostina Calabrese},
  title = {{Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024)}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure