Day 2 - EM
joanappardal opened this issue · 0 comments
joanappardal commented
"Another observation is that the accuracy goes up from 32% to 38%. Note that normally you will run this algorithm for 200 iterations, we stopped earlier for time constraints. [...] The actual state of the art is around 71% for fully unsupervised (Grac ̧a, 2010; Berg-Kirkpatrick et al., 2010) and 80% (Das and Petrov, 2011) using parallel data and information from labels in the other language."
I would add the expected value for such number of iterations:
"For 100 iterations we could expect a value around ??% and ??% for 200."
Also the number of iterations used by the state of the art referred works.
Maybe a note on the use of distributed computation would also be useful...