15 November 2019 - Anders Søgaard: Seminar
Title: NLP is Impossible: Swinging the Pendulum Slightly Back
NLP is, technically, impossible. We simply cannot learn to predict properties of a drifting population that cannot be adequately defined. Or, should I say, we will never know when we do, and if we do, our predictions will quickly worsen. In practice, we should therefore not rely too heavily on finite samples, but dig a bit deeper in the toolbox of science: Our best practices rely on multiple finite samples, but are still, in the words of John Dewey, “badly handicapped”. What else hides in the toolbox of science? Well, if you want to know, and asking Siri didn’t help, go to your library and pick up all their John Dewey books - or add my talk to your calendar.
PhD, Dr.Phil, Professor at Dpt. of Computer Science, University of Copenhagen and a http://coastalcph.github.io/ member. While Anders Søgaard has published a lot of papers and won occasional awards at top venues in NLP/ML, his research lacks a clear focus, and it doesn’t really help that he is also a publishing poet. He is currently a visiting scientist at Google Research in Berlin, Germany.