TRAILS Trustworthy Machine Learning Seminar Series
Speaker: Marine Carpuat, a professor of computer science at the University of Maryland and member of TRAILS
Where: University of Maryland, CSIC 2120 and on Zoom: https://umd.zoom.us/j/93960331973?pwd=tVQ7gtFucJEkSYna38zePzkPpHm3J7.1
Talk: From Quality to Communication Success: Insights from Recent Research and Open Questions for AI Translation
Abstract: Machine translation benchmarks have never looked better, but how well do current tools help people communicate across language barriers? In this talk, I will share some insights from human-centered work on how lay users perceive and rely on machine translation, and what those findings imply for evaluation and system design. I will then turn to open questions on what it would take to put communication success at the center of LLM research & development.
[2] Measuring User’s Mental Models of Speech Translation in Human-AI Collaboration
Bio: Marine Carpuat is a Professor in the Department of Computer Science at the University of Maryland. Her research focuses on developing AI techniques that help people communicate across languages, and studying whether those systems actually succeed. Her work spans NLP methods, evaluation methodology, and human-centered studies of how people perceive and rely on AI-generated translations. Before joining UMD, she was a Research Officer at the National Research Council Canada. She received her Ph.D. from the Hong Kong University of Science and Technology and a diplôme d'ingénieur from the French grande école Supélec.
About the TRAILS Trustworthy Machine Learning Seminar Series: Each week a researcher from across the TRAILS community presents current work on what it takes to make AI systems worth trusting: measuring and evaluating their behavior, interpreting and steering them, keeping them robust and secure, protecting privacy, and understanding how people actually rely on them — with occasional excursions into the philosophy and psychology of trust. Talks are 40 minutes, followed by 20 minutes of discussion.