From 0eec1d12affa9c8301612cfb5cddf706d2628e2b Mon Sep 17 00:00:00 2001
From: Matthew Honnibal <honnibal@gmail.com>
Date: Wed, 27 May 2015 01:14:07 +0200
Subject: [PATCH] * Add comment about zipf reweighting

---
 spacy/_ml.pyx | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/spacy/_ml.pyx b/spacy/_ml.pyx
index a2b943589..3a439e2ba 100644
--- 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):