An increasingly common expression of online hate speech is multimodal in nature and comes in the form of memes. Designing systems that can automatically detect hateful content is of paramount importance if we are to mitigate its undesirable effects on the society at large. As memes convey a message using both images and text, effective detection depends on models that are capable of multimodal reasoning and joint visual and language understanding.Dr Helen Yannakoudakis
14 May 2021
King's success in hateful memes challenge
King’s Lecturer in Computer Science, Dr Helen Yannakoudakis, was part of a team that enjoyed success in Facebook AI’s ‘Hateful Memes Challenge’ competition.
The ‘Hateful Memes Challenge’ competition set by Facebook AI, Getty Images and DrivenData addressed the difficulty of using AI to decide if a meme is offensive. AI can identify ‘hateful’ text or ‘hateful’ images, but this becomes more complex when images and text that might be inoffensive on their own are combined to make a meme.
More than 3,000 people from 150 countries took part in the competition. Teams were presented with a unique dataset of over 10,000 memes, and set with the goal of developing multimodal machine learning models to automatically identify hateful content.
The ‘Kingsterdam’ team, comprised of researchers from King’s College London and the University of Amsterdam, placed fourth in the competition. The team won a cash prize for their entry, and presented their findings at the 2020 Conference on Neural Information Processing Systems – a leading conference for machine learning and AI.
Dr Helen Yannakoudakis, Lecturer in Natural Language Processing and Machine Learning at the Department of Informatics and lead of the Kingsterdam team, commented:
Helen elaborated further on the team's approach:
We developed a framework of training algorithms and data manipulation strategies that allow us to better detect multimodal hate in unbalanced data distributions by leveraging signal from pairs of hateful examples and their benign counterparts (where either the image or the text of a meme is changed to transform it from hateful to non-hateful), and better guide machine learning for memes that require subtle reasoning and understanding.Dr Helen Yannakoudakis
Find out more about details of the team’s approach.