* Add comment about zipf reweighting

2015-05-27 01:14:07 +02:00 · 2015-05-27 01:14:07 +02:00 · 0eec1d12af
parent 4d37b66c55
commit 0eec1d12af
1 changed files with 3 additions and 0 deletions
--- a/spacy/_ml.pyx
+++ b/spacy/_ml.pyx
@ -47,6 +47,9 @@ cdef class Model:
    @cython.cdivision
    @cython.boundscheck(False)
    cdef int regularize(self, Feature* feats, int n, int a=3) except -1:
+        # Use the Zipfian corruptions technique from here:
+        # http://www.aclweb.org/anthology/N13-1077
+        # This seems good for 0.1 - 0.3 % on OOD data.
        cdef int i
        cdef long[:] zipfs = numpy.random.zipf(a, n)
        for i in range(n):