Enhancing harmful content detection in memes using multimodal machine learning models

Abstract

Memes are a popular way of sharing messages with large audiences. They are often funny or informative, but memes can also spread hate speech and harm- ful content. This represents a challenge for social networks, which need to detect and moderate this type of content to provide a safer online environment. This thesis focuses on improving the detection of harmful content in memes using ad- vanced machine learning methods, including canonical transformer-based models and recent multimodal large language models (MLLMs). First, the work presents a literature review that highlights the strengths, weaknesses, and challenges of current methods. It also introduces a new taxonomy in order to categorize current methods, making it easier for researchers to compare and improve detection tech- niques. Then, this work introduces a novel approach to enhance the performance of canonical multimodal transformer models. This is done by adding a specific module called Compact Parameter Blocks into the encoder segments of these mod- els. The experimental results demonstrate that the proposed method outperforms several more complex approaches. The thesis also explores the use of generative models, such as MLLMs, to detect aggressive memes by using specific prompts to guide the models. The results show that these models can identify aggressive memes, despite not being explicitly designed for this task. However, the perfor- mance of MLLMs decreases significantly when the level of multimodality reasoning required is very high. This research contributes to the field by providing practi- cal improvements to current detection methods and by exploring new approaches using generative models. These advances are important in creating safer online environments while respecting freedom of expression.