Researchers at the University of Waterloo have pioneered a novel machine-learning approach that identifies hate speech on social media with an impressive 88% accuracy, significantly reducing the workload and emotional toll on content moderators.
A novel tactic, referred to as the Multi-Modal Dialogue Transformer (mDT), excels at identifying relationships between textual content and images, while also providing insightful feedback that contextualises input, distinct from traditional hate speech detection approaches. By reducing false positives, this approach helps mitigate the risk of incorrectly labelling culturally sensitive content as hate speech, thereby preserving nuanced expressions and cultural diversity online.
According to Liam Hebert, a PhD scholar in laptop science at Waterloo University and the primary creator of the examine, “We genuinely hope that our findings will contribute to reducing the emotional toll on individuals who are currently forced to manually sift through hate speech.” “We propose that by adopting a community-centric approach to AI development, we can contribute to the creation of safer online environments for everyone.”
For several years, researchers have endeavored to develop models that decipher the significance of human dialogue, but these models have historically faltered in capturing the subtleties of conversations and contextual nuances. While earlier methods have struggled to identify hate speech with more than 74% accuracy, the Waterloo study achieved a notable milestone in this regard.
“With context being a vital component, it’s crucial to grasp the nuances of hate speech,” Hebert noted. While seemingly harmless, the phrase ‘That is gross!’ can drastically shift meanings when used in response to a photo of pizza with pineapple compared to one depicting an individual from a marginalized group.
While understanding this distinction comes naturally to humans, replicating it in a machine learning model – specifically, teaching an artificial intelligence to comprehend contextual connections within a dialogue that incorporates images and other multimedia elements – proves an incredibly daunting challenge.
Unlike previous attempts, the Waterloo team designed and fine-tuned their model on a dataset comprising not only remote hateful comments, but also the contextual information surrounding those comments. The AI-powered mannequin excelled in engaging with a staggering 8,266 Reddit discussions, garnering comprehensive feedback – 18,359 labelled comments – from a diverse range of 850 online communities.
“With over three billion people actively using social media every day,” Herbert said. “The impact of social media platforms has reached unprecedented levels.” To build areas where everyone is respected and feels secure, it’s essential to develop a robust system that can effectively detect and address hate speech on a massive scale.
Recently published in the proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence, the analysis “Multi-Modal Dialogue Transformer: Integrating Textual content, Photos and Graph Transformers to Detect Hate Speech on Social Media” made its debut.
