You are here
Language Modeling at Web Scale
Kenneth Heafield is a Lecturer at the University of Edinburgh.
Motivated by machine translation problems, he prefers to build scalable
systems instead of approximating. He is the author of the KenLM toolkit
that scales statistical language models to trillions of words and runs
Google-scale problems on a single machine. He completed his PhD at
Carnegie Mellon and worked at Stanford as a postdoctoral scholar.
Language models predict what you will type and are essential to machine translation. Building and applying these models presents a common trade-off between quality and computational cost. This talk takes an uncompromising approach to quality by making better systems with streaming algorithms, memory latency, quality approximations, and optionally specialized hardware.