Start with improving the place results

analyze.py

@@ -1,3 +1,77 @@
+import re
+import string
+from nltk.tag.stanford import POSTagger
+
+tagger = POSTagger('./stanford-postagger-full-2014-10-26/models/german-fast.tagger',
+                   './stanford-postagger-full-2014-10-26/stanford-postagger-3.5.0.jar',
+                    'UTF-8')
+
+punctuation_regex = re.compile("[%s]" % re.escape(string.punctuation))
+
+def get_potential_places(article_place, article_body):
+    """
+    Returns a list of potential places as tuples with their part-of-speech tags
+    for later filtering
+    """
+    # remove punctuation
+    full_text = punctuation_regex.sub(" ", article_place + " " + article_body)
+
+    pos = tagger.tag(full_text.split())
+
+    # extract the places
+    places = []
+    is_matching = False
+    current_match = []
+    for tuple in pos:
+        if is_matching:
+            # when we're matching, the phrases we're looking for look like
+            # "Im S-Bahnhof Wedding"... the tags below mean
+            if tuple[1] in ("ART", "ADJA", "NN", "NE", "CARD"):
+                current_match.append(tuple)
+            else:
+                # we stop the match, so append the current match
+                places.append(current_match)
+                current_match = []
+                is_matching = False
+        else:
+            # start matching when we have a preposition
+            if tuple[1] in ("APPR", "APPRART"):
+                is_matching = True
+
+    return places
+
+def improve_potential_places(pos_tuples):
+    """
+    Improves the matches' quality so we don't have to look up the lat-lng of so
+    many mismatches
+    """
+    better_tuples = []
+    for tuple_list in pos_tuples:
+        # first, exluce empty lists
+        if tuple_list:
+            cleaner = []
+
+            index = -1
+            for tuple in tuple_list:
+                index += 1
+
+                # exclude articles ("the", "a"), they only introduce noise, but
+                # keep the wh
+                if tuple[1] is "ART":
+                    continue
+
+                # if we have numbers in the middle of our phrase, it's probably
+                # also not useful (as opposed to Krügerstr. 22)
+                if tuple[1] is "CARD" and index < len(tuple_list):
+                    cleaner_tuple = []
+                    break
+
+                cleaner.append(tuple)
+
+            better_tuples.append(cleaner)
+
+    return better_tuples
+
 def get_district(article_headline):
     """
     Returns a geo-coded version of a district an article is about, based on its

tagger.py

@@ -1,33 +1,17 @@
 from nltk.tag.stanford import POSTagger
 from models import Article
+from analyze import *
 
 tagger = POSTagger('./stanford-postagger-full-2014-10-26/models/german-fast.tagger',
                    './stanford-postagger-full-2014-10-26/stanford-postagger-3.5.0.jar',
                     'UTF-8')
 
 for article in Article.select().limit(100):
-    pos = tagger.tag((article.place + " " + article.description).split())
+    potential = get_potential_places(article.place, article.description)
-
+    places = improve_potential_places(potential)
-    # extract the places
-    places = []
-    is_matching = False
-    current_match = []
-    for tuple in pos:
-        if is_matching:
-            # when we're matching, the phrases we're looking for look like
-            # "Im S-Bahnhof Wedding"... the tags below mean
-            if tuple[1] in ("ART", "ADJA", "NN", "NE", "CARD"):
-                current_match.append(tuple)
-            else:
-                places.append(current_match)
-                current_match = []
-                is_matching = False
-        else:
-            # start matching when we have a preposition
-            if tuple[1] in ("APPR", "APPRART"):
-                is_matching = True
 
     print(article.place)
     print(article.description)
     print()
-    print("Relevant: " + str(places))
+    print("Potential: " + str(potential))
+    print("Improved:  " + str(places))