Add option for weighting anchor word MI by anchor_strength when ranki…

…ng words for topics

corextopic/corextopic.py

@@ -529,7 +529,7 @@ def calculate_mis(self, theta, log_p_y):
         mis = self.h_x - p_y * binary_entropy(np.exp(theta[3]).T) - (1 - p_y) * binary_entropy(np.exp(theta[2]).T)
         return (mis - 1. / (2. * self.n_samples)).clip(0.)  # P-T bias correction
 
-    def get_topics(self, n_words=10, topic=None, print_words=True):
+    def get_topics(self, n_words=10, topic=None, print_words=True, weighted_rank=True):
         """
         Gets the top words for topics. If `words` was not provided to CorEx,
         then the indices of the top words are returned
@@ -543,6 +543,10 @@ def get_topics(self, n_words=10, topic=None, print_words=True):
             the top words for all topics
         print_words, bool
             Whether to return string words or integer indices for the topics
+        weighted_rank, bool
+            Whether to weight the mutual information of anchor words with their
+            anchor strengths. If anchoring was not used, then this does not
+            affect the output
 
         RETURNS
         -------
@@ -576,12 +580,25 @@ def get_topics(self, n_words=10, topic=None, print_words=True):
             # Get indices of which words belong to the topic
             inds = np.where(self.alpha[n] >= 1.)[0]
             # Sort topic words according to mutual information
-            inds = inds[np.argsort(-self.alpha[n,inds] * self.mis[n,inds])]
-            # Create topic to return
-            if print_words is True:
-                topic = [(self.col_index2word[ind], self.mis[n,ind], self.sign[n,ind]) for ind in inds[:n_words]]
+            if weighted_rank:
+                inds = inds[np.argsort(-self.alpha[n,inds] * self.mis[n,inds])]
             else:
-                topic = [(ind, self.mis[n,ind], self.sign[n,ind]) for ind in inds[:n_words]]
+                inds = inds[np.argsort(-self.mis[n,inds])]
+            # Create topic to return
+            topic = []
+            for ind in inds[:n_words]:
+                if print_words is True:
+                    word = self.col_index2word[ind]
+                else:
+                    word = ind
+
+                if weighted_rank:
+                    mi = self.alpha[n,inds] * self.mis[n,ind]
+                else:
+                    mi = self.mis[n,ind]
+
+                topic.append((word, mi, self.sign[n,ind]))
+
             # Add topic to list of topics if returning all topics. Otherwise, return topic
             if len(topic_ns) != 1:
                 topics.append(topic)