Sümeyye Meryem Taşyürek

Research keywords: Multimodal Learning, Representation Learning, Generative Models, Embodied Communication

Bio:

Hi, I am Meryem! I am originally from Ankara, Türkiye. Before moving to Edinburgh to begin my PhD, I completed my BSc and MSc in Computer Engineering at Hacettepe University, where I also worked as a Research Assistant. During this time, I was involved in teaching and research projects in machine learning, natural language processing and computer vision. My master’s research explored transformer-based generative models and latent representation learning for generating sign language from spoken language text.

Through my previous work I had publications at FG, ICCV, and WACV, alongside a third-place award in the Sign Language Production Challenge at CVPR 2025. My research journey has gradually led me towards broader questions about how AI systems represent and communicate meaning, bringing together my interests in multimodal learning, generative modelling, and human cognition.

Outside academia, I enjoy reading about psychology and philosophy, particularly topics that help me better understand people and the way we think, behave, and interact. I also enjoy hiking, long walks, and exploring new places.

PhD research:

My research explores how cognitively informed approaches can improve the representation, generation, and evaluation of embodied and multimodal communication.

Generative models in this domain can struggle to preserve meaningful linguistic distinctions and variation, with a tendency to produce averaged outputs that compromise naturalness and communicative meaning. I am interested in understanding how representational and modelling choices contribute to these limitations, and how insights from human cognition can inform more expressive and reliable generative systems.

Building on my previous work in sign language production, I aim to explore these questions at the intersection of multimodal learning, representation learning, and human perception. Sign language offers a rich experimental setting for investigating how linguistic meaning is expressed through bodily movement and interpreted by humans.

A further goal is to develop evaluation approaches that go beyond conventional computational metrics to better capture semantic and perceptual aspects of generated communication, contributing to more interpretable and trustworthy multimodal AI systems.

Supervisors: Mirella Lapata, Alessandro Suglia