* Remove POS alignment stuff. Now use training data based on raw text, instead of clumsy detokenization stuff

spacy/pos_util.py

@@ -6,56 +6,35 @@ from .en import EN
 from .pos import Tagger
 
 
-def realign_tagged(token_rules, tagged_line, sep='/'):
+def read_gold(file_):
-    words, pos = zip(*[token.rsplit(sep, 1) for token in tagged_line.split()])
+    paras = file_.read().strip().split('\n\n')
-    positions = util.detokenize(token_rules, words)
+    golds = []
-    aligned = []
+    for para in paras:
-    for group in positions:
+        if not para.strip():
-        w_group = [words[i] for i in group]
-        p_group = [pos[i] for i in group]
-        aligned.append('<SEP>'.join(w_group) + sep + '_'.join(p_group))
-    return ' '.join(aligned)
-
-
-def read_tagged(detoken_rules, file_, sep='/'):
-    sentences = []
-    for line in file_:
-        if not line.strip():
             continue
-        line = realign_tagged(detoken_rules, line, sep=sep)
+        lines = para.strip().split('\n')
-        tokens, tags = _parse_line(line, sep)
+        raw = lines.pop(0)
-        assert len(tokens) == len(tags)
+        gold_toks = lines.pop(0)
-        sentences.append((tokens, tags))
+        tokens = EN.tokenize(raw)
-    return sentences
-
-
-def _parse_line(line, sep):
-    words = []
         tags = []
-    for token_str in line.split():
+        conll_toks = []
-        word, pos = token_str.rsplit(sep, 1)
+        for line in lines:
-        word = word.replace('<SEP>', '')
+            pieces = line.split()
-        subtokens = EN.tokenize(word)
+            conll_toks.append((int(pieces[0]), len(pieces[1]), pieces[3]))
-        subtags = pos.split('_')
+        for i, token in enumerate(tokens):
-        while len(subtags) < len(subtokens):
+            if not conll_toks:
-            subtags.append('NULL')
+                tags.append('NULL')
-        assert len(subtags) == len(subtokens), [t.string for t in subtokens]
+            elif token.idx == conll_toks[0][0]:
-        words.append(word)
+                tags.append(conll_toks[0][2])
-        tags.extend([Tagger.encode_pos(ptb_to_univ(pos)) for pos in subtags])
+                conll_toks.pop(0)
-    tokens = EN.tokenize(' '.join(words)), tags
+            elif token.idx < conll_toks[0]:
-    return tokens
+                tags.append('NULL')
-
+            else:
-
+                conll_toks.pop(0)
-def get_tagdict(train_sents):
+        assert len(tags) == len(tokens)
-    tagdict = {}
+        tags = [Tagger.encode_pos(t) for t in tags]
-    for tokens, tags in train_sents:
+        golds.append((tokens, tags))
-        for i, tag in enumerate(tags):
+    return golds
-            if tag == 'NULL':
-                continue
-            word = tokens.string(i)
-            tagdict.setdefault(word, {}).setdefault(tag, 0)
-            tagdict[word][tag] += 1
-    return tagdict
 
 
 def ptb_to_univ(tag):