Seminari di Ingegneria Informatica - Vilém Zouhar: Theory and Practice of Efficient Evaluation
Speaker:
Vilém Zouhar (ETH Zurich)
Data dell'evento:
Venerdì, 9 October, 2026 - 11:30
Luogo:
Zoom Link: https://uniroma1.zoom.us/j/85485829872
Contatto:
Stefano Leonardi
Speaker: Vilém Zouhar (ETH Zurich)
Title: Theory and Practice of Efficient Evaluation
Zoom Link: https://uniroma1.zoom.us/j/85485829872
Abstract: Science without evaluation is limited to blind exploration, though good evaluation comes at a cost. As NLP reaches human parity, previous evaluation practices are becoming obsolete. We cover various perspectives on NLP evaluation, ranging from human to automatic evaluation and from theoretical to practical aspects, such as how to construct challenging and efficient testsets.
Bio: Vilém is last-year PhD at ETH Zurich and a Google PhD Fellow. He works on both the practical and theoretical aspects of evaluation (human annotations, automated metrics, methods), especially in the context of multilingual NLP. He organizes WMT human evaluation and recently the Last Translation Benchmark efforts.