Papers
Topics
Authors
Recent
Search
2000 character limit reached

Multi-Modal Discussion Transformer: Integrating Text, Images and Graph Transformers to Detect Hate Speech on Social Media

Published 18 Jul 2023 in cs.CL, cs.LG, cs.MM, and cs.SI | (2307.09312v4)

Abstract: We present the Multi-Modal Discussion Transformer (mDT), a novel methodfor detecting hate speech in online social networks such as Reddit discussions. In contrast to traditional comment-only methods, our approach to labelling a comment as hate speech involves a holistic analysis of text and images grounded in the discussion context. This is done by leveraging graph transformers to capture the contextual relationships in the discussion surrounding a comment and grounding the interwoven fusion layers that combine text and image embeddings instead of processing modalities separately. To evaluate our work, we present a new dataset, HatefulDiscussions, comprising complete multi-modal discussions from multiple online communities on Reddit. We compare the performance of our model to baselines that only process individual comments and conduct extensive ablation studies.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (34)
  1. Prompting for Multimodal Hateful Meme Classification. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 321–332. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics.
  2. HateBERT: Retraining BERT for Abusive Language Detection in English. In Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), 17–25. Online: Association for Computational Linguistics.
  3. Title-and-Tag Contrastive Vision-and-Language Transformer for Social Media Popularity Prediction. In Proceedings of the 30th ACM International Conference on Multimedia, MM ’22, 7008–7012. New York, NY, USA: Association for Computing Machinery. ISBN 9781450392037.
  4. Hate Speech in Online Social Media. SIGWEB Newsl., (Autumn).
  5. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.
  6. Large Scale Crowdsourcing and Characterization of Twitter Abusive Behavior. Proceedings of the International AAAI Conference on Web and Social Media, 12(1).
  7. Abusive language detection in heterogeneous contexts: Dataset collection and the role of supervised attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AISI Track), volume 35, 14804–14812.
  8. Hanu, L.; and Unitary team. 2020. Detoxify. Github. https://github.com/unitaryai/detoxify.
  9. Qualitative Analysis of a Graph Transformer Approach to Addressing Hate Speech: Adapting to Dynamically Changing Content. Workshop in Artificial Intelligence for Social Good (AISG) at AAAI.
  10. Predicting Hateful Discussions on Reddit using Graph Transformer Networks and Communal Context. In 2022 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), 9–17.
  11. Sentiment analysis on product reviews using machine learning techniques. In Cognitive Informatics and Soft Computing: Proceeding of CISC 2017, 639–647. Springer.
  12. The hateful memes challenge: Detecting hate speech in multimodal memes.
  13. ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision. In Proceedings of the 38th International Conference on Machine Learning, 5583–5594. PMLR. ISSN: 2640-3498.
  14. Towards a Comprehensive Taxonomy and Large-Scale Annotated Corpus for Online Slur Usage. In Proceedings of the Fourth Workshop on Online Abuse and Harms, 138–149. Online: Association for Computational Linguistics.
  15. Multi-Modal Sarcasm Detection via Cross-Modal Graph Convolutional Network. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 1767–1777. Dublin, Ireland: Association for Computational Linguistics.
  16. ViLBERT: pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, 2, 13–23. Red Hook, NY, USA: Curran Associates Inc.
  17. Understanding Political Polarization via Jointly Modeling Users, Connections and Multimodal Contents on Heterogeneous Graphs. In Proceedings of the 30th ACM International Conference on Multimedia, MM ’22, 4072–4082. New York, NY, USA: Association for Computing Machinery. ISBN 9781450392037.
  18. Hatexplain: A benchmark dataset for explainable hate speech detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AISI Track), volume 35, 14867–14875.
  19. Meta. 2023. Meta Reports First Quarter 2023 Results. Meta Investor Relations. https://investor.fb.com/investor-news/press-release-details/2023/Meta-Reports-First-Quarter-2023-Results/default.aspx.
  20. Tweetnerd-end to end entity linking benchmark for tweets. Advances in Neural Information Processing Systems, 35: 1419–1433.
  21. Attention Bottlenecks for Multimodal Fusion. In Ranzato, M.; Beygelzimer, A.; Dauphin, Y.; Liang, P.; and Vaughan, J. W., eds., Advances in Neural Information Processing Systems, volume 34, 14200–14213. Curran Associates, Inc.
  22. Learning Reddit User Reputation Using Graphical Attention Networks. In Future Technologies Conference (FTC) 2020, Volume 1, 777–789. ISBN 978-3-030-63128-4.
  23. A Benchmark Dataset for Learning to Intervene in Online Hate Speech. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 4755–4764. Hong Kong, China: Association for Computational Linguistics.
  24. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748–8763. PMLR.
  25. Towards a multi-agent system for online hate speech detection. Second Workshop on Autonomous Agents for Social Good (AASG), AAMAS, 2021.
  26. Sentiment and Emotion-Aware Multi-Modal Complaint Identification. Proceedings of the AAAI Conference on Artificial Intelligence (AISI Track), 36(11): 12163–12171.
  27. An Efficient Multi-View Multimodal Data Processing Framework for Social Media Popularity Prediction. In Proceedings of the 30th ACM International Conference on Multimedia, MM ’22, 7200–7204. New York, NY, USA: Association for Computing Machinery. ISBN 9781450392037.
  28. Hate speech harms: a social justice discussion of disabled Norwegians’ experiences. Disability & Society, 34(3): 368–383.
  29. Introducing CAD: the Contextual Abuse Dataset. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2289–2303. Online: Association for Computational Linguistics.
  30. Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate Detection. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 1667–1682. Online: Association for Computational Linguistics.
  31. Online hate speech victimization and depressive symptoms among adolescents: The protective role of resilience. Cyberpsychology, Behavior, and Social Networking.
  32. Quantifying social organization and political polarization in online platforms. Nature, 600(7888): 264–268.
  33. Do transformers really perform badly for graph representation? Advances in Neural Information Processing Systems, 34: 28877–28888.
  34. Predicting the Type and Target of Offensive Posts in Social Media. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 1415–1420. Minneapolis, Minnesota: Association for Computational Linguistics.
Citations (8)
View on Semantic Scholar

Summary

  • The paper introduces a Multi-Modal Discussion Transformer that integrates text, images, and graph transformers to enhance hate speech detection on platforms like Reddit.
  • It employs innovative modality fusion using shared bottlenecks and graph transformers with hierarchical spatial encoding to capture the complete discussion context.
  • Empirical results on the HatefulDiscussions dataset demonstrate improved precision, recall, and F1 scores compared to traditional text-only and context-aware models.

Multi-Modal Discussion Transformer: Integrating Text, Images, and Graph Transformers for Hate Speech Detection on Social Media

The paper "Multi-Modal Discussion Transformer: Integrating Text, Images, and Graph Transformers to Detect Hate Speech on Social Media" introduces a novel approach to the complex issue of hate speech detection on platforms like Reddit. The current challenges with traditional methods are that they often focus solely on textual content, neglecting the contextual richness offered by multi-modal data, including images and the broader conversational context. The proposed Multi-Modal Discussion Transformer (mDT) addresses these limitations by employing a holistic method that integrates multiple modalities within the discussion threads of social media platforms.

Methodological Innovation

The mDT architecture stands out for its comprehensive integration of text and image data through a graph transformer framework. This is achieved via several key components:

  1. Modality Fusion: The paper presents an innovative mechanism utilizing shared modality bottlenecks, ensuring that both text and images compress information efficiently before fusing. This prevents the potential dilution of semantic content and optimizes multi-modal information exchange.
  2. Graph Transformer: Unlike traditional models that process each comment in isolation, mDT employs graph transformers to embed comments within the broader context of an entire discussion. This strategy effectively grounds the meaning of each comment in its conversational environment.
  3. Hierarchical Spatial Encoding: A novel feature of mDT is its hierarchical spatial encoding method, enhancing the model's ability to discern the structural relationships between comments in a tree-format discussion, such as those found on Reddit.
  4. Dataset Contributions: An integral aspect of this research is the introduction of the HatefulDiscussions dataset, comprising annotated multi-modal discussions sourced from various Reddit communities. This dataset serves as a benchmark for evaluating the mDT's performance in a real-world setting.

Empirical Assessment

The experiments conducted demonstrate the superiority of mDT over existing hate speech detection methods. Compared against text-only models like BERT-HateXplain and RoBERTa Dynabench, mDT showed improved precision, recall, and F1 scores, highlighting the efficacy of integrating images and contextual graph structures. Notably, mDT delivers a substantial improvement in F1 score over previous context-aware models such as the text-only Graphormer.

Implications and Future Directions

The integration of multi-modal data offers promising advancements in content moderation across online platforms. By considering both images and the interconnected nature of online discussions, the mDT enhances the nuanced detection of hate speech, which can vary significantly based on context. This capability not only mitigates false positives but also provides a tool for understanding community-specific language and behavior, such as reclaimed vernacular in marginalized communities.

Future work could extend this model to other platforms and explore additional modalities such as audio or video content. There is also potential for enriching models through named entity recognition, further enhancing contextual and real-world understanding. As AI continues to evolve, the framework proposed by mDT may aid in developing robust solutions that encapsulate the dynamic nature of digital communications.

In summary, the Multi-Modal Discussion Transformer presents a significant step forward in hate speech detection, directly addressing the constraints of past models by embracing the multi-faceted and interconnected nature of online discussions. This research not only contributes to computational methods but also holds potential social value in promoting healthier, inclusive online communities.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Sign Up to Generate

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Generate Now

Collections

Sign up for free to add this paper to one or more collections.

User Circle Single Streamline Icon: https://streamlinehq.com Sign Up

Tweets

Sign up for free to view the 3 tweets with 173 likes about this paper.

User Circle Single Streamline Icon: https://streamlinehq.com Sign Up for Free