Supplementary appendix to "Learning to Extract International Relations from Political Context" (ACL 2013)
Brendan T. O’Connor, Brandon Stewart, Noah A. Smith · 2013
k ∑ w∈M,v∈M,w 6=v nw,knv,k To properly compute a posterior expectation using multiple samples,Q/N should be re-evaluated on several complete samples and then averaged. However, we found little variation between samples so used only one. We also tried evaluating a singleQ/N where nwk and nvk are averaged counts from multiple samples—using this corresponds to a factored, mean-field-like approximation to the posterior—but it also was very similar to using a single sample. The implementation is in verbdict/score.py.