Merge branch 'develop' into spacy.io

README.md

@@ -54,9 +54,9 @@ valuable if it's shared publicly, so that more people can benefit from it.
 
 | Type                     | Platforms                                              |
 | ------------------------ | ------------------------------------------------------ |
-| 🚨**Bug Reports**        | [GitHub Issue Tracker]                                 |
+| 🚨 **Bug Reports**       | [GitHub Issue Tracker]                                 |
 | 🎁 **Feature Requests**  | [GitHub Issue Tracker]                                 |
-| 👩‍💻**Usage Questions**    | [Stack Overflow] · [Gitter Chat] · [Reddit User Group] |
+| 👩‍💻 **Usage Questions**   | [Stack Overflow] · [Gitter Chat] · [Reddit User Group] |
 | 🗯 **General Discussion** | [Gitter Chat] · [Reddit User Group]                    |
 
 [github issue tracker]: https://github.com/explosion/spaCy/issues

examples/training/rehearsal.py

@@ -4,7 +4,7 @@ import random
 import srsly
 import spacy
 from spacy.gold import GoldParse
-from spacy.util import minibatch
+from spacy.util import minibatch, compounding
 
 
 LABEL = "ANIMAL"
@@ -54,9 +54,17 @@ def main(model_name, unlabelled_loc):
     nlp.get_pipe("ner").add_label(LABEL)
     raw_docs = list(read_raw_data(nlp, unlabelled_loc))
     optimizer = nlp.resume_training()
+    # Avoid use of Adam when resuming training. I don't understand this well
+    # yet, but I'm getting weird results from Adam. Try commenting out the
+    # nlp.update(), and using Adam -- you'll find the models drift apart.
+    # I guess Adam is losing precision, introducing gradient noise?
+    optimizer.alpha = 0.1
+    optimizer.b1 = 0.0
+    optimizer.b2 = 0.0
 
     # get names of other pipes to disable them during training
     other_pipes = [pipe for pipe in nlp.pipe_names if pipe != "ner"]
+    sizes = compounding(1.0, 4.0, 1.001)
     with nlp.disable_pipes(*other_pipes):
         for itn in range(n_iter):
             random.shuffle(TRAIN_DATA)
@@ -64,13 +72,22 @@ def main(model_name, unlabelled_loc):
             losses = {}
             r_losses = {}
             # batch up the examples using spaCy's minibatch
-            raw_batches = minibatch(raw_docs, size=batch_size)
+            raw_batches = minibatch(raw_docs, size=4)
-            for doc, gold in TRAIN_DATA:
+            for batch in minibatch(TRAIN_DATA, size=sizes):
-                nlp.update([doc], [gold], sgd=optimizer, drop=dropout, losses=losses)
+                docs, golds = zip(*batch)
+                nlp.update(docs, golds, sgd=optimizer, drop=dropout, losses=losses)
                 raw_batch = list(next(raw_batches))
                 nlp.rehearse(raw_batch, sgd=optimizer, losses=r_losses)
             print("Losses", losses)
             print("R. Losses", r_losses)
+    print(nlp.get_pipe('ner').model.unseen_classes)
+    test_text = "Do you like horses?"
+    doc = nlp(test_text)
+    print("Entities in '%s'" % test_text)
+    for ent in doc.ents:
+        print(ent.label_, ent.text)
+
+
 
 
 if __name__ == "__main__":

examples/training/train_new_entity_type.py

@@ -45,19 +45,19 @@ LABEL = "ANIMAL"
 TRAIN_DATA = [
     (
         "Horses are too tall and they pretend to care about your feelings",
-        {"entities": [(0, 6, "ANIMAL")]},
+        {"entities": [(0, 6, LABEL)]},
     ),
     ("Do they bite?", {"entities": []}),
     (
         "horses are too tall and they pretend to care about your feelings",
-        {"entities": [(0, 6, "ANIMAL")]},
+        {"entities": [(0, 6, LABEL)]},
     ),
-    ("horses pretend to care about your feelings", {"entities": [(0, 6, "ANIMAL")]}),
+    ("horses pretend to care about your feelings", {"entities": [(0, 6, LABEL)]}),
     (
         "they pretend to care about your feelings, those horses",
-        {"entities": [(48, 54, "ANIMAL")]},
+        {"entities": [(48, 54, LABEL)]},
     ),
-    ("horses?", {"entities": [(0, 6, "ANIMAL")]}),
+    ("horses?", {"entities": [(0, 6, LABEL)]}),
 ]
 
 
@@ -67,8 +67,9 @@ TRAIN_DATA = [
     output_dir=("Optional output directory", "option", "o", Path),
     n_iter=("Number of training iterations", "option", "n", int),
 )
-def main(model=None, new_model_name="animal", output_dir=None, n_iter=10):
+def main(model=None, new_model_name="animal", output_dir=None, n_iter=30):
     """Set up the pipeline and entity recognizer, and train the new entity."""
+    random.seed(0)
     if model is not None:
         nlp = spacy.load(model)  # load existing spaCy model
         print("Loaded model '%s'" % model)
@@ -85,21 +86,22 @@ def main(model=None, new_model_name="animal", output_dir=None, n_iter=10):
         ner = nlp.get_pipe("ner")
 
     ner.add_label(LABEL)  # add new entity label to entity recognizer
+    # Adding extraneous labels shouldn't mess anything up
+    ner.add_label('VEGETABLE')
     if model is None:
         optimizer = nlp.begin_training()
     else:
-        # Note that 'begin_training' initializes the models, so it'll zero out
+        optimizer = nlp.resume_training()
-        # existing entity types.
+    move_names = list(ner.move_names)
-        optimizer = nlp.entity.create_optimizer()
-
     # get names of other pipes to disable them during training
     other_pipes = [pipe for pipe in nlp.pipe_names if pipe != "ner"]
     with nlp.disable_pipes(*other_pipes):  # only train NER
+        sizes = compounding(1.0, 4.0, 1.001)
+        # batch up the examples using spaCy's minibatch
         for itn in range(n_iter):
             random.shuffle(TRAIN_DATA)
+            batches = minibatch(TRAIN_DATA, size=sizes)
             losses = {}
-            # batch up the examples using spaCy's minibatch
-            batches = minibatch(TRAIN_DATA, size=compounding(4.0, 32.0, 1.001))
             for batch in batches:
                 texts, annotations = zip(*batch)
                 nlp.update(texts, annotations, sgd=optimizer, drop=0.35, losses=losses)
@@ -124,6 +126,8 @@ def main(model=None, new_model_name="animal", output_dir=None, n_iter=10):
         # test the saved model
         print("Loading from", output_dir)
         nlp2 = spacy.load(output_dir)
+        # Check the classes have loaded back consistently
+        assert nlp2.get_pipe('ner').move_names == move_names
         doc2 = nlp2(test_text)
         for ent in doc2.ents:
             print(ent.label_, ent.text)

spacy/_ml.py

@@ -571,8 +571,6 @@ def build_text_classifier(nr_class, width=64, **cfg):
                 zero_init(Affine(nr_class, nr_class * 2, drop_factor=0.0))
                 >> logistic
             )
-
-
         model = (
             (linear_model | cnn_model)
             >> output_layer

spacy/about.py

@@ -4,7 +4,7 @@
 # fmt: off
 
 __title__ = "spacy-nightly"
-__version__ = "2.1.0a9.dev1"
+__version__ = "2.1.0a9.dev2"
 __summary__ = "Industrial-strength Natural Language Processing (NLP) with Python and Cython"
 __uri__ = "https://spacy.io"
 __author__ = "Explosion AI"

spacy/errors.py

@@ -290,7 +290,8 @@ class Errors(object):
             "NBOR_RELOP.")
     E101 = ("NODE_NAME should be a new node and NBOR_NAME should already have "
             "have been declared in previous edges.")
-    E102 = ("Can't merge non-disjoint spans. '{token}' is already part of tokens to merge")
+    E102 = ("Can't merge non-disjoint spans. '{token}' is already part of "
+            "tokens to merge.")
     E103 = ("Trying to set conflicting doc.ents: '{span1}' and '{span2}'. A token"
             " can only be part of one entity, so make sure the entities you're "
             "setting don't overlap.")
@@ -318,12 +319,12 @@ class Errors(object):
             "So instead of pickling the span, pickle the Doc it belongs to or "
             "use Span.as_doc to convert the span to a standalone Doc object.")
     E113 = ("The newly split token can only have one root (head = 0).")
-    E114 = ("The newly split token needs to have a root (head = 0)")
+    E114 = ("The newly split token needs to have a root (head = 0).")
-    E115 = ("All subtokens must have associated heads")
+    E115 = ("All subtokens must have associated heads.")
     E116 = ("Cannot currently add labels to pre-trained text classifier. Add "
             "labels before training begins. This functionality was available "
             "in previous versions, but had significant bugs that led to poor "
-            "performance")
+            "performance.")
     E117 = ("The newly split tokens must match the text of the original token. "
             "New orths: {new}. Old text: {old}.")
 

spacy/lang/char_classes.py

@@ -24,52 +24,68 @@ _latin_l_supplement = r"\u00DF-\u00F6\u00F8-\u00FF"
 _latin_supplement = r"\u00C0-\u00D6\u00D8-\u00F6\u00F8-\u00FF"
 
 # letters with diacritics - Catalan, Czech, Latin, Latvian, Lithuanian, Polish, Slovak, Turkish, Welsh
-_latin_u_extendedA = r"\u0100\u0102\u0104\u0106\u0108\u010A\u010C\u010E\u0110\u0112\u0114\u0116\u0118\u011A\u011C" \
+_latin_u_extendedA = (
-                         r"\u011E\u0120\u0122\u0124\u0126\u0128\u012A\u012C\u012E\u0130\u0132\u0134\u0136\u0139\u013B" \
+    r"\u0100\u0102\u0104\u0106\u0108\u010A\u010C\u010E\u0110\u0112\u0114\u0116\u0118\u011A\u011C"
-                         r"\u013D\u013F\u0141\u0143\u0145\u0147\u014A\u014C\u014E\u0150\u0152\u0154\u0156\u0158" \
+    r"\u011E\u0120\u0122\u0124\u0126\u0128\u012A\u012C\u012E\u0130\u0132\u0134\u0136\u0139\u013B"
-                         r"\u015A\u015C\u015E\u0160\u0162\u0164\u0166\u0168\u016A\u016C\u016E\u0170\u0172\u0174\u0176" \
+    r"\u013D\u013F\u0141\u0143\u0145\u0147\u014A\u014C\u014E\u0150\u0152\u0154\u0156\u0158"
-                         r"\u0178\u0179\u017B\u017D"
+    r"\u015A\u015C\u015E\u0160\u0162\u0164\u0166\u0168\u016A\u016C\u016E\u0170\u0172\u0174\u0176"
-_latin_l_extendedA = r"\u0101\u0103\u0105\u0107\u0109\u010B\u010D\u010F\u0111\u0113\u0115\u0117\u0119\u011B\u011D" \
+    r"\u0178\u0179\u017B\u017D"
-                        r"\u011F\u0121\u0123\u0125\u0127\u0129\u012B\u012D\u012F\u0131\u0133\u0135\u0137\u0138\u013A" \
+)
-                        r"\u013C\u013E\u0140\u0142\u0144\u0146\u0148\u0149\u014B\u014D\u014F\u0151\u0153\u0155\u0157" \
+_latin_l_extendedA = (
-                        r"\u0159\u015B\u015D\u015F\u0161\u0163\u0165\u0167\u0169\u016B\u016D\u016F\u0171\u0173\u0175" \
+    r"\u0101\u0103\u0105\u0107\u0109\u010B\u010D\u010F\u0111\u0113\u0115\u0117\u0119\u011B\u011D"
-                        r"\u0177\u017A\u017C\u017E\u017F"
+    r"\u011F\u0121\u0123\u0125\u0127\u0129\u012B\u012D\u012F\u0131\u0133\u0135\u0137\u0138\u013A"
+    r"\u013C\u013E\u0140\u0142\u0144\u0146\u0148\u0149\u014B\u014D\u014F\u0151\u0153\u0155\u0157"
+    r"\u0159\u015B\u015D\u015F\u0161\u0163\u0165\u0167\u0169\u016B\u016D\u016F\u0171\u0173\u0175"
+    r"\u0177\u017A\u017C\u017E\u017F"
+)
 _latin_extendedA = r"\u0100-\u017F"
 
 # special characters - Khoisan, Pan-Nigerian, Pinyin, Romanian
 # those that are a combination of both upper and lower letters are only included in the group _latin_extendedB
-_latin_u_extendedB = r"\u0181\u0182\u0184\u0186\u0187\u0189-\u018B\u018E-\u0191\u0193\u0194\u0196-\u0198\u019C" \
+_latin_u_extendedB = (
-                         r"\u019D\u019F\u01A0\u01A2\u01A4\u01A6\u01A7\u01A9\u01AC\u01AE\u01AF\u01B1-\u01B3\u01B5" \
+    r"\u0181\u0182\u0184\u0186\u0187\u0189-\u018B\u018E-\u0191\u0193\u0194\u0196-\u0198\u019C"
-                         r"\u01B7\u01B8\u01BC\u01C4\u01C7\u01CA\u01CD\u01CF\u01D1\u01D3\u01D5\u01D7\u01D9\u01DB" \
+    r"\u019D\u019F\u01A0\u01A2\u01A4\u01A6\u01A7\u01A9\u01AC\u01AE\u01AF\u01B1-\u01B3\u01B5"
-                         r"\u01DE\u01E0\u01E2\u01E4\u01E6\u01E8\u01EA\u01EC\u01EE\u01F1\u01F4\u01F6-\u01F8\u01FA" \
+    r"\u01B7\u01B8\u01BC\u01C4\u01C7\u01CA\u01CD\u01CF\u01D1\u01D3\u01D5\u01D7\u01D9\u01DB"
-                         r"\u01FC\u01FE\u0200\u0202\u0204\u0206\u0208\u020A\u020C\u020E\u0210\u0212\u0214\u0216" \
+    r"\u01DE\u01E0\u01E2\u01E4\u01E6\u01E8\u01EA\u01EC\u01EE\u01F1\u01F4\u01F6-\u01F8\u01FA"
-                         r"\u0218\u021A\u021C\u021E\u0220\u0222\u0224\u0226\u0228\u022A\u022C\u022E\u0230\u0232" \
+    r"\u01FC\u01FE\u0200\u0202\u0204\u0206\u0208\u020A\u020C\u020E\u0210\u0212\u0214\u0216"
-                         r"\u023A\u023B\u023D\u023E\u0241\u0243-\u0246\u0248\u024A\u024C\u024E"
+    r"\u0218\u021A\u021C\u021E\u0220\u0222\u0224\u0226\u0228\u022A\u022C\u022E\u0230\u0232"
-_latin_l_extendedB = r"\u0180\u0183\u0185\u0188\u018C\u018D\u0192\u0195\u0199-\u019B\u019E\u01A1\u01A3\u01A5" \
+    r"\u023A\u023B\u023D\u023E\u0241\u0243-\u0246\u0248\u024A\u024C\u024E"
-                         r"\u01A8\u01AA\u01AB\u01AD\u01B0\u01B4\u01B6\u01B9\u01BA\u01BD-\u01BF\u01C6\u01C9\u01CC" \
+)
-                         r"\u01CE\u01D0\u01D2\u01D4\u01D6\u01D8\u01DA\u01DC\u01DD\u01DF\u01E1\u01E3\u01E5\u01E7" \
+_latin_l_extendedB = (
-                         r"\u01E9\u01EB\u01ED\u01EF\u01F0\u01F3\u01F5\u01F9\u01FB\u01FD\u01FF\u0201\u0203\u0205" \
+    r"\u0180\u0183\u0185\u0188\u018C\u018D\u0192\u0195\u0199-\u019B\u019E\u01A1\u01A3\u01A5"
-                         r"\u0207\u0209\u020B\u020D\u020F\u0211\u0213\u0215\u0217\u0219\u021B\u021D\u021F\u0221" \
+    r"\u01A8\u01AA\u01AB\u01AD\u01B0\u01B4\u01B6\u01B9\u01BA\u01BD-\u01BF\u01C6\u01C9\u01CC"
-                         r"\u0223\u0225\u0227\u0229\u022B\u022D\u022F\u0231\u0233-\u0239\u023C\u023F\u0240\u0242" \
+    r"\u01CE\u01D0\u01D2\u01D4\u01D6\u01D8\u01DA\u01DC\u01DD\u01DF\u01E1\u01E3\u01E5\u01E7"
-                         r"\u0247\u0249\u024B\u024D\u024F"
+    r"\u01E9\u01EB\u01ED\u01EF\u01F0\u01F3\u01F5\u01F9\u01FB\u01FD\u01FF\u0201\u0203\u0205"
+    r"\u0207\u0209\u020B\u020D\u020F\u0211\u0213\u0215\u0217\u0219\u021B\u021D\u021F\u0221"
+    r"\u0223\u0225\u0227\u0229\u022B\u022D\u022F\u0231\u0233-\u0239\u023C\u023F\u0240\u0242"
+    r"\u0247\u0249\u024B\u024D\u024F"
+)
 _latin_extendedB = r"\u0180-\u01BF\u01C4-\u024F"
 
 # special characters - Uighur, Uralic Phonetic
-_latin_u_extendedC = r"\u2C60\u2C62-\u2C64\u2C67\u2C69\u2C6B\u2C6D-\u2C70\u2C72\u2C75\u2C7E\u2C7F"
+_latin_u_extendedC = (
-_latin_l_extendedC = r"\u2C61\u2C65\u2C66\u2C68\u2C6A\u2C6C\u2C71\u2C73\u2C74\u2C76-\u2C7B"
+    r"\u2C60\u2C62-\u2C64\u2C67\u2C69\u2C6B\u2C6D-\u2C70\u2C72\u2C75\u2C7E\u2C7F"
+)
+_latin_l_extendedC = (
+    r"\u2C61\u2C65\u2C66\u2C68\u2C6A\u2C6C\u2C71\u2C73\u2C74\u2C76-\u2C7B"
+)
 _latin_extendedC = r"\u2C60-\u2C7B\u2C7E\u2C7F"
 
 # special characters - phonetic, Mayan, Medieval
-_latin_u_extendedD = r"\uA722\uA724\uA726\uA728\uA72A\uA72C\uA72E\uA732\uA734\uA736\uA738\uA73A\uA73C" \
+_latin_u_extendedD = (
-                         r"\uA73E\uA740\uA742\uA744\uA746\uA748\uA74A\uA74C\uA74E\uA750\uA752\uA754\uA756\uA758" \
+    r"\uA722\uA724\uA726\uA728\uA72A\uA72C\uA72E\uA732\uA734\uA736\uA738\uA73A\uA73C"
-                         r"\uA75A\uA75C\uA75E\uA760\uA762\uA764\uA766\uA768\uA76A\uA76C\uA76E\uA779\uA77B\uA77D" \
+    r"\uA73E\uA740\uA742\uA744\uA746\uA748\uA74A\uA74C\uA74E\uA750\uA752\uA754\uA756\uA758"
-                         r"\uA77E\uA780\uA782\uA784\uA786\uA78B\uA78D\uA790\uA792\uA796\uA798\uA79A\uA79C\uA79E" \
+    r"\uA75A\uA75C\uA75E\uA760\uA762\uA764\uA766\uA768\uA76A\uA76C\uA76E\uA779\uA77B\uA77D"
-                         r"\uA7A0\uA7A2\uA7A4\uA7A6\uA7A8\uA7AA-\uA7AE\uA7B0-\uA7B4\uA7B6\uA7B8"
+    r"\uA77E\uA780\uA782\uA784\uA786\uA78B\uA78D\uA790\uA792\uA796\uA798\uA79A\uA79C\uA79E"
-_latin_l_extendedD = r"\uA723\uA725\uA727\uA729\uA72B\uA72D\uA72F-\uA731\uA733\uA735\uA737\uA739\uA73B\uA73D" \
+    r"\uA7A0\uA7A2\uA7A4\uA7A6\uA7A8\uA7AA-\uA7AE\uA7B0-\uA7B4\uA7B6\uA7B8"
-                         r"\uA73F\uA741\uA743\uA745\uA747\uA749\uA74B\uA74D\uA74F\uA751\uA753\uA755\uA757\uA759" \
+)
-                         r"\uA75B\uA75D\uA75F\uA761\uA763\uA765\uA767\uA769\uA76B\uA76D\uA76F\uA771-\uA778\uA77A" \
+_latin_l_extendedD = (
-                         r"\uA77C\uA77F\uA781\uA783\uA785\uA787\uA78C\uA78E\uA791\uA793-\uA795\uA797\uA799\uA79B" \
+    r"\uA723\uA725\uA727\uA729\uA72B\uA72D\uA72F-\uA731\uA733\uA735\uA737\uA739\uA73B\uA73D"
-                         r"\uA79D\uA79F\uA7A1\uA7A3\uA7A5\uA7A7\uA7A9\uA7AF\uA7B5\uA7B7\uA7B9\uA7FA"
+    r"\uA73F\uA741\uA743\uA745\uA747\uA749\uA74B\uA74D\uA74F\uA751\uA753\uA755\uA757\uA759"
+    r"\uA75B\uA75D\uA75F\uA761\uA763\uA765\uA767\uA769\uA76B\uA76D\uA76F\uA771-\uA778\uA77A"
+    r"\uA77C\uA77F\uA781\uA783\uA785\uA787\uA78C\uA78E\uA791\uA793-\uA795\uA797\uA799\uA79B"
+    r"\uA79D\uA79F\uA7A1\uA7A3\uA7A5\uA7A7\uA7A9\uA7AF\uA7B5\uA7B7\uA7B9\uA7FA"
+)
 _latin_extendedD = r"\uA722-\uA76F\uA771-\uA787\uA78B-\uA78E\uA790-\uA7B9\uA7FA"
 
 # special characters - phonetic Teuthonista and Sakha
@@ -81,42 +97,80 @@ _latin_l_phonetic = r"\u0250-\u02AF\u1D00-\u1D25\u1D6B-\u1D77\u1D79-\u1D9A"
 _latin_phonetic = _latin_l_phonetic
 
 # letters with multiple diacritics - Vietnamese
-_latin_u_diacritics = r"\u1E00\u1E02\u1E04\u1E06\u1E08\u1E0A\u1E0C\u1E0E\u1E10\u1E12\u1E14\u1E16\u1E18\u1E1A" \
+_latin_u_diacritics = (
-                          r"\u1E1C\u1E1E\u1E20\u1E22\u1E24\u1E26\u1E28\u1E2A\u1E2C\u1E2E\u1E30\u1E32\u1E34\u1E36" \
+    r"\u1E00\u1E02\u1E04\u1E06\u1E08\u1E0A\u1E0C\u1E0E\u1E10\u1E12\u1E14\u1E16\u1E18\u1E1A"
-                          r"\u1E38\u1E3A\u1E3C\u1E3E\u1E40\u1E42\u1E44\u1E46\u1E48\u1E4A\u1E4C\u1E4E\u1E50\u1E52" \
+    r"\u1E1C\u1E1E\u1E20\u1E22\u1E24\u1E26\u1E28\u1E2A\u1E2C\u1E2E\u1E30\u1E32\u1E34\u1E36"
-                          r"\u1E54\u1E56\u1E58\u1E5A\u1E5C\u1E5E\u1E60\u1E62\u1E64\u1E66\u1E68\u1E6A\u1E6C\u1E6E" \
+    r"\u1E38\u1E3A\u1E3C\u1E3E\u1E40\u1E42\u1E44\u1E46\u1E48\u1E4A\u1E4C\u1E4E\u1E50\u1E52"
-                          r"\u1E70\u1E72\u1E74\u1E76\u1E78\u1E7A\u1E7C\u1E7E\u1E80\u1E82\u1E84\u1E86\u1E88\u1E8A" \
+    r"\u1E54\u1E56\u1E58\u1E5A\u1E5C\u1E5E\u1E60\u1E62\u1E64\u1E66\u1E68\u1E6A\u1E6C\u1E6E"
-                          r"\u1E8C\u1E8E\u1E90\u1E92\u1E94\u1E9E\u1EA0\u1EA2\u1EA4\u1EA6\u1EA8\u1EAA\u1EAC\u1EAE" \
+    r"\u1E70\u1E72\u1E74\u1E76\u1E78\u1E7A\u1E7C\u1E7E\u1E80\u1E82\u1E84\u1E86\u1E88\u1E8A"
-                          r"\u1EB0\u1EB2\u1EB4\u1EB6\u1EB8\u1EBA\u1EBC\u1EBE\u1EC0\u1EC2\u1EC4\u1EC6\u1EC8" \
+    r"\u1E8C\u1E8E\u1E90\u1E92\u1E94\u1E9E\u1EA0\u1EA2\u1EA4\u1EA6\u1EA8\u1EAA\u1EAC\u1EAE"
-                          r"\u1ECA\u1ECC\u1ECE\u1ED0\u1ED2\u1ED4\u1ED6\u1ED8\u1EDA\u1EDC\u1EDE\u1EE0\u1EE2\u1EE4" \
+    r"\u1EB0\u1EB2\u1EB4\u1EB6\u1EB8\u1EBA\u1EBC\u1EBE\u1EC0\u1EC2\u1EC4\u1EC6\u1EC8"
-                          r"\u1EE6\u1EE8\u1EEA\u1EEC\u1EEE\u1EF0\u1EF2\u1EF4\u1EF6\u1EF8\u1EFA\u1EFC\u1EFE"
+    r"\u1ECA\u1ECC\u1ECE\u1ED0\u1ED2\u1ED4\u1ED6\u1ED8\u1EDA\u1EDC\u1EDE\u1EE0\u1EE2\u1EE4"
-_latin_l_diacritics = r"\u1E01\u1E03\u1E05\u1E07\u1E09\u1E0B\u1E0D\u1E0F\u1E11\u1E13\u1E15\u1E17\u1E19\u1E1B" \
+    r"\u1EE6\u1EE8\u1EEA\u1EEC\u1EEE\u1EF0\u1EF2\u1EF4\u1EF6\u1EF8\u1EFA\u1EFC\u1EFE"
-                          r"\u1E1D\u1E1F\u1E21\u1E23\u1E25\u1E27\u1E29\u1E2B\u1E2D\u1E2F\u1E31\u1E33\u1E35\u1E37" \
+)
-                          r"\u1E39\u1E3B\u1E3D\u1E3F\u1E41\u1E43\u1E45\u1E47\u1E49\u1E4B\u1E4D\u1E4F\u1E51\u1E53" \
+_latin_l_diacritics = (
-                          r"\u1E55\u1E57\u1E59\u1E5B\u1E5D\u1E5F\u1E61\u1E63\u1E65\u1E67\u1E69\u1E6B\u1E6D\u1E6F" \
+    r"\u1E01\u1E03\u1E05\u1E07\u1E09\u1E0B\u1E0D\u1E0F\u1E11\u1E13\u1E15\u1E17\u1E19\u1E1B"
-                          r"\u1E71\u1E73\u1E75\u1E77\u1E79\u1E7B\u1E7D\u1E7F\u1E81\u1E83\u1E85\u1E87\u1E89\u1E8B" \
+    r"\u1E1D\u1E1F\u1E21\u1E23\u1E25\u1E27\u1E29\u1E2B\u1E2D\u1E2F\u1E31\u1E33\u1E35\u1E37"
-                          r"\u1E8D\u1E8F\u1E91\u1E93\u1E95-\u1E9D\u1E9F\u1EA1\u1EA3\u1EA5\u1EA7\u1EA9\u1EAB\u1EAD" \
+    r"\u1E39\u1E3B\u1E3D\u1E3F\u1E41\u1E43\u1E45\u1E47\u1E49\u1E4B\u1E4D\u1E4F\u1E51\u1E53"
-                          r"\u1EAF\u1EB1\u1EB3\u1EB5\u1EB7\u1EB9\u1EBB\u1EBD\u1EBF\u1EC1\u1EC3\u1EC5\u1EC7\u1EC9" \
+    r"\u1E55\u1E57\u1E59\u1E5B\u1E5D\u1E5F\u1E61\u1E63\u1E65\u1E67\u1E69\u1E6B\u1E6D\u1E6F"
-                          r"\u1ECB\u1ECD\u1ECF\u1ED1\u1ED3\u1ED5\u1ED7\u1ED9\u1EDB\u1EDD\u1EDF\u1EE1\u1EE3\u1EE5" \
+    r"\u1E71\u1E73\u1E75\u1E77\u1E79\u1E7B\u1E7D\u1E7F\u1E81\u1E83\u1E85\u1E87\u1E89\u1E8B"
-                          r"\u1EE7\u1EE9\u1EEB\u1EED\u1EEF\u1EF1\u1EF3\u1EF5\u1EF7\u1EF9\u1EFB\u1EFD\u1EFF"
+    r"\u1E8D\u1E8F\u1E91\u1E93\u1E95-\u1E9D\u1E9F\u1EA1\u1EA3\u1EA5\u1EA7\u1EA9\u1EAB\u1EAD"
+    r"\u1EAF\u1EB1\u1EB3\u1EB5\u1EB7\u1EB9\u1EBB\u1EBD\u1EBF\u1EC1\u1EC3\u1EC5\u1EC7\u1EC9"
+    r"\u1ECB\u1ECD\u1ECF\u1ED1\u1ED3\u1ED5\u1ED7\u1ED9\u1EDB\u1EDD\u1EDF\u1EE1\u1EE3\u1EE5"
+    r"\u1EE7\u1EE9\u1EEB\u1EED\u1EEF\u1EF1\u1EF3\u1EF5\u1EF7\u1EF9\u1EFB\u1EFD\u1EFF"
+)
 _latin_diacritics = r"\u1E00-\u1EFF"
 
 # all lower latin classes
-LATIN_LOWER_BASIC = _latin_l_standard + _latin_l_standard_fullwidth + _latin_l_supplement + _latin_l_extendedA
+LATIN_LOWER_BASIC = (
-LATIN_LOWER = LATIN_LOWER_BASIC + _latin_l_extendedB + _latin_l_extendedC + _latin_l_extendedD + _latin_l_extendedE \
+    _latin_l_standard
-               + _latin_l_phonetic + _latin_l_diacritics
+    + _latin_l_standard_fullwidth
+    + _latin_l_supplement
+    + _latin_l_extendedA
+)
+LATIN_LOWER = (
+    LATIN_LOWER_BASIC
+    + _latin_l_extendedB
+    + _latin_l_extendedC
+    + _latin_l_extendedD
+    + _latin_l_extendedE
+    + _latin_l_phonetic
+    + _latin_l_diacritics
+)
 
 # all upper latin classes
-LATIN_UPPER_BASIC = _latin_u_standard + _latin_u_standard_fullwidth + _latin_u_supplement + _latin_u_extendedA
+LATIN_UPPER_BASIC = (
-LATIN_UPPER = LATIN_UPPER_BASIC + _latin_u_extendedB + _latin_u_extendedC + _latin_u_extendedD + _latin_u_diacritics
+    _latin_u_standard
+    + _latin_u_standard_fullwidth
+    + _latin_u_supplement
+    + _latin_u_extendedA
+)
+LATIN_UPPER = (
+    LATIN_UPPER_BASIC
+    + _latin_u_extendedB
+    + _latin_u_extendedC
+    + _latin_u_extendedD
+    + _latin_u_diacritics
+)
 
 # all latin classes
-LATIN_BASIC = _latin_standard + _latin_standard_fullwidth + _latin_supplement + _latin_extendedA
+LATIN_BASIC = (
-LATIN = LATIN_BASIC + _latin_extendedB + _latin_extendedC + _latin_extendedD + _latin_extendedE \
+    _latin_standard + _latin_standard_fullwidth + _latin_supplement + _latin_extendedA
-         + _latin_phonetic + _latin_diacritics
+)
+LATIN = (
+    LATIN_BASIC
+    + _latin_extendedB
+    + _latin_extendedC
+    + _latin_extendedD
+    + _latin_extendedE
+    + _latin_phonetic
+    + _latin_diacritics
+)
 
-_persian = r"\u0620-\u064A\u066E-\u06D5\u06E5-\u06FF\u0750-\u077F\u08A0-\u08BD" \
+_persian = (
-           r"\uFB50-\uFBB1\uFBD3-\uFD3D\uFD50-\uFDC7\uFDF0-\uFDFB\uFE70-\uFEFC\U0001EE00-\U0001EEBB"
+    r"\u0620-\u064A\u066E-\u06D5\u06E5-\u06FF\u0750-\u077F\u08A0-\u08BD"
+    r"\uFB50-\uFBB1\uFBD3-\uFD3D\uFD50-\uFDC7\uFDF0-\uFDFB\uFE70-\uFEFC\U0001EE00-\U0001EEBB"
+)
 
 _russian_lower = r"ёа-я"
 _russian_upper = r"ЁА-Я"
@@ -165,33 +219,35 @@ _hyphens = "- – — -- --- —— ~"
 
 # Various symbols like dingbats, but also emoji
 # Details: https://www.compart.com/en/unicode/category/So
-_other_symbols = r"\u00A6\u00A9\u00AE\u00B0\u0482\u058D\u058E\u060E\u060F\u06DE\u06E9\u06FD\u06FE\u07F6\u09FA\u0B70" \
+_other_symbols = (
-                 r"\u0BF3-\u0BF8\u0BFA\u0C7F\u0D4F\u0D79\u0F01-\u0F03\u0F13\u0F15-\u0F17\u0F1A-\u0F1F\u0F34" \
+    r"\u00A6\u00A9\u00AE\u00B0\u0482\u058D\u058E\u060E\u060F\u06DE\u06E9\u06FD\u06FE\u07F6\u09FA\u0B70"
-                 r"\u0F36\u0F38\u0FBE-\u0FC5\u0FC7-\u0FCC\u0FCE\u0FCF\u0FD5-\u0FD8\u109E\u109F\u1390-\u1399" \
+    r"\u0BF3-\u0BF8\u0BFA\u0C7F\u0D4F\u0D79\u0F01-\u0F03\u0F13\u0F15-\u0F17\u0F1A-\u0F1F\u0F34"
-                 r"\u1940\u19DE-\u19FF\u1B61-\u1B6A\u1B74-\u1B7C\u2100\u2101\u2103-\u2106\u2108\u2109\u2114\u2116" \
+    r"\u0F36\u0F38\u0FBE-\u0FC5\u0FC7-\u0FCC\u0FCE\u0FCF\u0FD5-\u0FD8\u109E\u109F\u1390-\u1399"
-                 r"\u2117\u211E-\u2123\u2125\u2127\u2129\u212E\u213A\u213B\u214A\u214C\u214D\u214F\u218A\u218B" \
+    r"\u1940\u19DE-\u19FF\u1B61-\u1B6A\u1B74-\u1B7C\u2100\u2101\u2103-\u2106\u2108\u2109\u2114\u2116"
-                 r"\u2195-\u2199\u219C-\u219F\u21A1\u21A2\u21A4\u21A5\u21A7-\u21AD\u21AF-\u21CD\u21D0\u21D1\u21D3" \
+    r"\u2117\u211E-\u2123\u2125\u2127\u2129\u212E\u213A\u213B\u214A\u214C\u214D\u214F\u218A\u218B"
-                 r"\u21D5-\u21F3\u2300-\u2307\u230C-\u231F\u2322-\u2328\u232B-\u237B\u237D-\u239A\u23B4-\u23DB" \
+    r"\u2195-\u2199\u219C-\u219F\u21A1\u21A2\u21A4\u21A5\u21A7-\u21AD\u21AF-\u21CD\u21D0\u21D1\u21D3"
-                 r"\u23E2-\u2426\u2440-\u244A\u249C-\u24E9\u2500-\u25B6\u25B8-\u25C0\u25C2-\u25F7\u2600-\u266E" \
+    r"\u21D5-\u21F3\u2300-\u2307\u230C-\u231F\u2322-\u2328\u232B-\u237B\u237D-\u239A\u23B4-\u23DB"
-                 r"\u2670-\u2767\u2794-\u27BF\u2800-\u28FF\u2B00-\u2B2F\u2B45\u2B46\u2B4D-\u2B73\u2B76-\u2B95" \
+    r"\u23E2-\u2426\u2440-\u244A\u249C-\u24E9\u2500-\u25B6\u25B8-\u25C0\u25C2-\u25F7\u2600-\u266E"
-                 r"\u2B98-\u2BC8\u2BCA-\u2BFE\u2CE5-\u2CEA\u2E80-\u2E99\u2E9B-\u2EF3\u2F00-\u2FD5\u2FF0-\u2FFB" \
+    r"\u2670-\u2767\u2794-\u27BF\u2800-\u28FF\u2B00-\u2B2F\u2B45\u2B46\u2B4D-\u2B73\u2B76-\u2B95"
-                 r"\u3004\u3012\u3013\u3020\u3036\u3037\u303E\u303F\u3190\u3191\u3196-\u319F\u31C0-\u31E3" \
+    r"\u2B98-\u2BC8\u2BCA-\u2BFE\u2CE5-\u2CEA\u2E80-\u2E99\u2E9B-\u2EF3\u2F00-\u2FD5\u2FF0-\u2FFB"
-                 r"\u3200-\u321E\u322A-\u3247\u3250\u3260-\u327F\u328A-\u32B0\u32C0-\u32FE\u3300-\u33FF\u4DC0-\u4DFF" \
+    r"\u3004\u3012\u3013\u3020\u3036\u3037\u303E\u303F\u3190\u3191\u3196-\u319F\u31C0-\u31E3"
-                 r"\uA490-\uA4C6\uA828-\uA82B\uA836\uA837\uA839\uAA77-\uAA79\uFDFD\uFFE4\uFFE8\uFFED\uFFEE\uFFFC" \
+    r"\u3200-\u321E\u322A-\u3247\u3250\u3260-\u327F\u328A-\u32B0\u32C0-\u32FE\u3300-\u33FF\u4DC0-\u4DFF"
-                 r"\uFFFD\U00010137-\U0001013F\U00010179-\U00010189\U0001018C-\U0001018E\U00010190-\U0001019B" \
+    r"\uA490-\uA4C6\uA828-\uA82B\uA836\uA837\uA839\uAA77-\uAA79\uFDFD\uFFE4\uFFE8\uFFED\uFFEE\uFFFC"
-                 r"\U000101A0\U000101D0-\U000101FC\U00010877\U00010878\U00010AC8\U0001173F\U00016B3C-\U00016B3F" \
+    r"\uFFFD\U00010137-\U0001013F\U00010179-\U00010189\U0001018C-\U0001018E\U00010190-\U0001019B"
-                 r"\U00016B45\U0001BC9C\U0001D000-\U0001D0F5\U0001D100-\U0001D126\U0001D129-\U0001D164" \
+    r"\U000101A0\U000101D0-\U000101FC\U00010877\U00010878\U00010AC8\U0001173F\U00016B3C-\U00016B3F"
-                 r"\U0001D16A-\U0001D16C\U0001D183\U0001D184\U0001D18C-\U0001D1A9\U0001D1AE-\U0001D1E8" \
+    r"\U00016B45\U0001BC9C\U0001D000-\U0001D0F5\U0001D100-\U0001D126\U0001D129-\U0001D164"
-                 r"\U0001D200-\U0001D241\U0001D245\U0001D300-\U0001D356\U0001D800-\U0001D9FF\U0001DA37-\U0001DA3A" \
+    r"\U0001D16A-\U0001D16C\U0001D183\U0001D184\U0001D18C-\U0001D1A9\U0001D1AE-\U0001D1E8"
-                 r"\U0001DA6D-\U0001DA74\U0001DA76-\U0001DA83\U0001DA85\U0001DA86\U0001ECAC\U0001F000-\U0001F02B" \
+    r"\U0001D200-\U0001D241\U0001D245\U0001D300-\U0001D356\U0001D800-\U0001D9FF\U0001DA37-\U0001DA3A"
-                 r"\U0001F030-\U0001F093\U0001F0A0-\U0001F0AE\U0001F0B1-\U0001F0BF\U0001F0C1-\U0001F0CF" \
+    r"\U0001DA6D-\U0001DA74\U0001DA76-\U0001DA83\U0001DA85\U0001DA86\U0001ECAC\U0001F000-\U0001F02B"
-                 r"\U0001F0D1-\U0001F0F5\U0001F110-\U0001F16B\U0001F170-\U0001F1AC\U0001F1E6-\U0001F202" \
+    r"\U0001F030-\U0001F093\U0001F0A0-\U0001F0AE\U0001F0B1-\U0001F0BF\U0001F0C1-\U0001F0CF"
-                 r"\U0001F210-\U0001F23B\U0001F240-\U0001F248\U0001F250\U0001F251\U0001F260-\U0001F265" \
+    r"\U0001F0D1-\U0001F0F5\U0001F110-\U0001F16B\U0001F170-\U0001F1AC\U0001F1E6-\U0001F202"
-                 r"\U0001F300-\U0001F3FA\U0001F400-\U0001F6D4\U0001F6E0-\U0001F6EC\U0001F6F0-\U0001F6F9" \
+    r"\U0001F210-\U0001F23B\U0001F240-\U0001F248\U0001F250\U0001F251\U0001F260-\U0001F265"
-                 r"\U0001F700-\U0001F773\U0001F780-\U0001F7D8\U0001F800-\U0001F80B\U0001F810-\U0001F847" \
+    r"\U0001F300-\U0001F3FA\U0001F400-\U0001F6D4\U0001F6E0-\U0001F6EC\U0001F6F0-\U0001F6F9"
-                 r"\U0001F850-\U0001F859\U0001F860-\U0001F887\U0001F890-\U0001F8AD\U0001F900-\U0001F90B" \
+    r"\U0001F700-\U0001F773\U0001F780-\U0001F7D8\U0001F800-\U0001F80B\U0001F810-\U0001F847"
-                 r"\U0001F910-\U0001F93E\U0001F940-\U0001F970\U0001F973-\U0001F976\U0001F97A\U0001F97C-\U0001F9A2" \
+    r"\U0001F850-\U0001F859\U0001F860-\U0001F887\U0001F890-\U0001F8AD\U0001F900-\U0001F90B"
-                 r"\U0001F9B0-\U0001F9B9\U0001F9C0-\U0001F9C2\U0001F9D0-\U0001F9FF\U0001FA60-\U0001FA6D"
+    r"\U0001F910-\U0001F93E\U0001F940-\U0001F970\U0001F973-\U0001F976\U0001F97A\U0001F97C-\U0001F9A2"
+    r"\U0001F9B0-\U0001F9B9\U0001F9C0-\U0001F9C2\U0001F9D0-\U0001F9FF\U0001FA60-\U0001FA6D"
+)
 
 UNITS = merge_chars(_units)
 CURRENCY = merge_chars(_currency)

spacy/syntax/_parser_model.pxd

@@ -19,6 +19,7 @@ cdef struct WeightsC:
     const float* feat_bias
     const float* hidden_bias
     const float* hidden_weights
+    const float* seen_classes
 
 
 cdef struct ActivationsC:

spacy/syntax/_parser_model.pyx

@@ -44,8 +44,10 @@ cdef WeightsC get_c_weights(model) except *:
     output.feat_bias = <const float*>state2vec.bias.data
     cdef np.ndarray vec2scores_W = model.vec2scores.W
     cdef np.ndarray vec2scores_b = model.vec2scores.b
+    cdef np.ndarray class_mask = model._class_mask
     output.hidden_weights = <const float*>vec2scores_W.data
     output.hidden_bias = <const float*>vec2scores_b.data
+    output.seen_classes = <const float*>class_mask.data
     return output
 
 
@@ -115,6 +117,16 @@ cdef void predict_states(ActivationsC* A, StateC** states,
     for i in range(n.states):
         VecVec.add_i(&A.scores[i*n.classes],
             W.hidden_bias, 1., n.classes)
+    # Set unseen classes to minimum value
+    i = 0
+    min_ = A.scores[0]
+    for i in range(1, n.states * n.classes):
+        if A.scores[i] < min_:
+            min_ = A.scores[i]
+    for i in range(n.states):
+        for j in range(n.classes):
+            if not W.seen_classes[j]:
+                A.scores[i*n.classes+j] = min_
 
 
 cdef void sum_state_features(float* output,
@@ -189,12 +201,17 @@ cdef int arg_max_if_valid(const weight_t* scores, const int* is_valid, int n) no
 
 
 class ParserModel(Model):
-    def __init__(self, tok2vec, lower_model, upper_model):
+    def __init__(self, tok2vec, lower_model, upper_model, unseen_classes=None):
         Model.__init__(self)
         self._layers = [tok2vec, lower_model, upper_model]
+        self.unseen_classes = set()
+        if unseen_classes:
+            for class_ in unseen_classes:
+                self.unseen_classes.add(class_)
 
     def begin_update(self, docs, drop=0.):
-        step_model = ParserStepModel(docs, self._layers, drop=drop)
+        step_model = ParserStepModel(docs, self._layers, drop=drop,
+                        unseen_classes=self.unseen_classes)
         def finish_parser_update(golds, sgd=None):
             step_model.make_updates(sgd)
             return None
@@ -207,9 +224,8 @@ class ParserModel(Model):
 
         with Model.use_device('cpu'):
             larger = Affine(new_output, smaller.nI)
-        # Set nan as value for unseen classes, to prevent prediction.
+        larger.W.fill(0.0)
-        larger.W.fill(self.ops.xp.nan)
+        larger.b.fill(0.0)
-        larger.b.fill(self.ops.xp.nan)
         # It seems very unhappy if I pass these as smaller.W?
         # Seems to segfault. Maybe it's a descriptor protocol thing?
         smaller_W = smaller.W
@@ -221,6 +237,8 @@ class ParserModel(Model):
         larger_W[:smaller.nO] = smaller_W
         larger_b[:smaller.nO] = smaller_b
         self._layers[-1] = larger
+        for i in range(smaller.nO, new_output):
+            self.unseen_classes.add(i)
 
     def begin_training(self, X, y=None):
         self.lower.begin_training(X, y=y)
@@ -239,18 +257,32 @@ class ParserModel(Model):
 
 
 class ParserStepModel(Model):
-    def __init__(self, docs, layers, drop=0.):
+    def __init__(self, docs, layers, unseen_classes=None, drop=0.):
         self.tokvecs, self.bp_tokvecs = layers[0].begin_update(docs, drop=drop)
         self.state2vec = precompute_hiddens(len(docs), self.tokvecs, layers[1],
                                             drop=drop)
         self.vec2scores = layers[-1]
         self.cuda_stream = util.get_cuda_stream()
         self.backprops = []
+        self._class_mask = numpy.zeros((self.vec2scores.nO,), dtype='f')
+        self._class_mask.fill(1)
+        if unseen_classes is not None:
+            for class_ in unseen_classes:
+                self._class_mask[class_] = 0.
 
     @property
     def nO(self):
         return self.state2vec.nO
 
+    def class_is_unseen(self, class_):
+        return self._class_mask[class_]
+
+    def mark_class_unseen(self, class_):
+        self._class_mask[class_] = 0
+
+    def mark_class_seen(self, class_):
+        self._class_mask[class_] = 1
+
     def begin_update(self, states, drop=0.):
         token_ids = self.get_token_ids(states)
         vector, get_d_tokvecs = self.state2vec.begin_update(token_ids, drop=0.0)
@@ -258,24 +290,12 @@ class ParserStepModel(Model):
         if mask is not None:
             vector *= mask
         scores, get_d_vector = self.vec2scores.begin_update(vector, drop=drop)
-        # We can have nans from unseen classes.
+        # If the class is unseen, make sure its score is minimum
-        # For backprop purposes, we want to treat unseen classes as having the
+        scores[:, self._class_mask == 0] = numpy.nanmin(scores)
-        # lowest score.
-        # numpy's nan_to_num function doesn't take a value, and nan is replaced
-        # by 0...-inf is replaced by minimum, so we go via that. Ugly to the max.
-        # Note that scores is always a numpy array! Should fix #3112
-        scores[numpy.isnan(scores)] = -numpy.inf
-        numpy.nan_to_num(scores, copy=False)
 
         def backprop_parser_step(d_scores, sgd=None):
-            # If we have a non-zero gradient for a previously unseen class,
+            # Zero vectors for unseen classes
-            # replace the weight with 0.
+            d_scores *= self._class_mask
-            new_classes = self.vec2scores.ops.xp.logical_and(
-                self.vec2scores.ops.xp.isnan(self.vec2scores.b),
-                d_scores.any(axis=0)
-            )
-            self.vec2scores.b[new_classes] = 0.
-            self.vec2scores.W[new_classes] = 0.
             d_vector = get_d_vector(d_scores, sgd=sgd)
             if mask is not None:
                 d_vector *= mask

spacy/syntax/nn_parser.pyx

@@ -163,6 +163,8 @@ cdef class Parser:
             added = self.moves.add_action(action, label)
             if added:
                 resized = True
+        if resized and "nr_class" in self.cfg:
+            self.cfg["nr_class"] = self.moves.n_moves
         if self.model not in (True, False, None) and resized:
             self.model.resize_output(self.moves.n_moves)
 
@@ -435,22 +437,22 @@ cdef class Parser:
         if self._rehearsal_model is None:
             return None
         losses.setdefault(self.name, 0.)
+ 
         states = self.moves.init_batch(docs)
         # This is pretty dirty, but the NER can resize itself in init_batch,
         # if labels are missing. We therefore have to check whether we need to
         # expand our model output.
         self.model.resize_output(self.moves.n_moves)
+        self._rehearsal_model.resize_output(self.moves.n_moves)
         # Prepare the stepwise model, and get the callback for finishing the batch
-        tutor = self._rehearsal_model(docs)
+        tutor, _ = self._rehearsal_model.begin_update(docs, drop=0.0)
         model, finish_update = self.model.begin_update(docs, drop=0.0)
         n_scores = 0.
         loss = 0.
-        non_zeroed_classes = self._rehearsal_model.upper.W.any(axis=1)
         while states:
-            targets, _ = tutor.begin_update(states)
+            targets, _ = tutor.begin_update(states, drop=0.)
-            guesses, backprop = model.begin_update(states)
+            guesses, backprop = model.begin_update(states, drop=0.)
-            d_scores = (targets - guesses) / targets.shape[0]
+            d_scores = (guesses - targets) / targets.shape[0]
-            d_scores *= non_zeroed_classes
             # If all weights for an output are 0 in the original model, don't
             # supervise that output. This allows us to add classes.
             loss += (d_scores**2).sum()
@@ -543,6 +545,9 @@ cdef class Parser:
             memset(is_valid, 0, self.moves.n_moves * sizeof(int))
             memset(costs, 0, self.moves.n_moves * sizeof(float))
             self.moves.set_costs(is_valid, costs, state, gold)
+            for j in range(self.moves.n_moves):
+                if costs[j] <= 0.0 and j in self.model.unseen_classes:
+                    self.model.unseen_classes.remove(j)
             cpu_log_loss(c_d_scores,
                 costs, is_valid, &scores[i, 0], d_scores.shape[1])
             c_d_scores += d_scores.shape[1]

spacy/syntax/transition_system.pyx

@@ -147,6 +147,8 @@ cdef class TransitionSystem:
     def initialize_actions(self, labels_by_action, min_freq=None):
         self.labels = {}
         self.n_moves = 0
+        added_labels = []
+        added_actions = {}
         for action, label_freqs in sorted(labels_by_action.items()):
             action = int(action)
             # Make sure we take a copy here, and that we get a Counter
@@ -157,6 +159,15 @@ cdef class TransitionSystem:
             sorted_labels.sort()
             sorted_labels.reverse()
             for freq, label_str in sorted_labels:
+                if freq < 0:
+                    added_labels.append((freq, label_str))
+                    added_actions.setdefault(label_str, []).append(action)
+                else:
+                    self.add_action(int(action), label_str)
+                    self.labels[action][label_str] = freq
+        added_labels.sort(reverse=True)
+        for freq, label_str in added_labels:
+            for action in added_actions[label_str]:
                 self.add_action(int(action), label_str)
                 self.labels[action][label_str] = freq
 

spacy/tests/doc/test_doc_api.py

@@ -6,7 +6,6 @@ import pytest
 import numpy
 from spacy.tokens import Doc
 from spacy.vocab import Vocab
-from spacy.attrs import LEMMA
 from spacy.errors import ModelsWarning
 
 from ..util import get_doc
@@ -139,81 +138,6 @@ def test_doc_api_set_ents(en_tokenizer):
     assert tokens.ents[0].end == 4
 
 
-def test_doc_api_merge(en_tokenizer):
-    text = "WKRO played songs by the beach boys all night"
-    attrs = {"tag": "NAMED", "lemma": "LEMMA", "ent_type": "TYPE"}
-    # merge both with bulk merge
-    doc = en_tokenizer(text)
-    assert len(doc) == 9
-    with doc.retokenize() as retokenizer:
-        retokenizer.merge(doc[4:7], attrs=attrs)
-        retokenizer.merge(doc[7:9], attrs=attrs)
-    assert len(doc) == 6
-    assert doc[4].text == "the beach boys"
-    assert doc[4].text_with_ws == "the beach boys "
-    assert doc[4].tag_ == "NAMED"
-    assert doc[5].text == "all night"
-    assert doc[5].text_with_ws == "all night"
-    assert doc[5].tag_ == "NAMED"
-
-
-def test_doc_api_merge_children(en_tokenizer):
-    """Test that attachments work correctly after merging."""
-    text = "WKRO played songs by the beach boys all night"
-    attrs = {"tag": "NAMED", "lemma": "LEMMA", "ent_type": "TYPE"}
-    doc = en_tokenizer(text)
-    assert len(doc) == 9
-    with doc.retokenize() as retokenizer:
-        retokenizer.merge(doc[4:7], attrs=attrs)
-    for word in doc:
-        if word.i < word.head.i:
-            assert word in list(word.head.lefts)
-        elif word.i > word.head.i:
-            assert word in list(word.head.rights)
-
-
-def test_doc_api_merge_hang(en_tokenizer):
-    text = "through North and South Carolina"
-    doc = en_tokenizer(text)
-    with doc.retokenize() as retokenizer:
-        retokenizer.merge(doc[3:5], attrs={"lemma": "", "ent_type": "ORG"})
-        retokenizer.merge(doc[1:2], attrs={"lemma": "", "ent_type": "ORG"})
-
-
-def test_doc_api_retokenizer(en_tokenizer):
-    doc = en_tokenizer("WKRO played songs by the beach boys all night")
-    with doc.retokenize() as retokenizer:
-        retokenizer.merge(doc[4:7])
-    assert len(doc) == 7
-    assert doc[4].text == "the beach boys"
-
-
-def test_doc_api_retokenizer_attrs(en_tokenizer):
-    doc = en_tokenizer("WKRO played songs by the beach boys all night")
-    # test both string and integer attributes and values
-    attrs = {LEMMA: "boys", "ENT_TYPE": doc.vocab.strings["ORG"]}
-    with doc.retokenize() as retokenizer:
-        retokenizer.merge(doc[4:7], attrs=attrs)
-    assert len(doc) == 7
-    assert doc[4].text == "the beach boys"
-    assert doc[4].lemma_ == "boys"
-    assert doc[4].ent_type_ == "ORG"
-
-
-@pytest.mark.xfail
-def test_doc_api_retokenizer_lex_attrs(en_tokenizer):
-    """Test that lexical attributes can be changed (see #2390)."""
-    doc = en_tokenizer("WKRO played beach boys songs")
-    assert not any(token.is_stop for token in doc)
-    with doc.retokenize() as retokenizer:
-        retokenizer.merge(doc[2:4], attrs={"LEMMA": "boys", "IS_STOP": True})
-    assert doc[2].text == "beach boys"
-    assert doc[2].lemma_ == "boys"
-    assert doc[2].is_stop
-    new_doc = Doc(doc.vocab, words=["beach boys"])
-    assert new_doc[0].is_stop
-
-
 def test_doc_api_sents_empty_string(en_tokenizer):
     doc = en_tokenizer("")
     doc.is_parsed = True

spacy/tests/doc/test_retokenize_merge.py

@@ -1,14 +1,89 @@
 # coding: utf-8
 from __future__ import unicode_literals
 
+import pytest
+from spacy.attrs import LEMMA
 from spacy.vocab import Vocab
 from spacy.tokens import Doc
-import pytest
 
 from ..util import get_doc
 
 
-def test_spans_merge_tokens(en_tokenizer):
+def test_doc_retokenize_merge(en_tokenizer):
+    text = "WKRO played songs by the beach boys all night"
+    attrs = {"tag": "NAMED", "lemma": "LEMMA", "ent_type": "TYPE"}
+    doc = en_tokenizer(text)
+    assert len(doc) == 9
+    with doc.retokenize() as retokenizer:
+        retokenizer.merge(doc[4:7], attrs=attrs)
+        retokenizer.merge(doc[7:9], attrs=attrs)
+    assert len(doc) == 6
+    assert doc[4].text == "the beach boys"
+    assert doc[4].text_with_ws == "the beach boys "
+    assert doc[4].tag_ == "NAMED"
+    assert doc[5].text == "all night"
+    assert doc[5].text_with_ws == "all night"
+    assert doc[5].tag_ == "NAMED"
+
+
+def test_doc_retokenize_merge_children(en_tokenizer):
+    """Test that attachments work correctly after merging."""
+    text = "WKRO played songs by the beach boys all night"
+    attrs = {"tag": "NAMED", "lemma": "LEMMA", "ent_type": "TYPE"}
+    doc = en_tokenizer(text)
+    assert len(doc) == 9
+    with doc.retokenize() as retokenizer:
+        retokenizer.merge(doc[4:7], attrs=attrs)
+    for word in doc:
+        if word.i < word.head.i:
+            assert word in list(word.head.lefts)
+        elif word.i > word.head.i:
+            assert word in list(word.head.rights)
+
+
+def test_doc_retokenize_merge_hang(en_tokenizer):
+    text = "through North and South Carolina"
+    doc = en_tokenizer(text)
+    with doc.retokenize() as retokenizer:
+        retokenizer.merge(doc[3:5], attrs={"lemma": "", "ent_type": "ORG"})
+        retokenizer.merge(doc[1:2], attrs={"lemma": "", "ent_type": "ORG"})
+
+
+def test_doc_retokenize_retokenizer(en_tokenizer):
+    doc = en_tokenizer("WKRO played songs by the beach boys all night")
+    with doc.retokenize() as retokenizer:
+        retokenizer.merge(doc[4:7])
+    assert len(doc) == 7
+    assert doc[4].text == "the beach boys"
+
+
+def test_doc_retokenize_retokenizer_attrs(en_tokenizer):
+    doc = en_tokenizer("WKRO played songs by the beach boys all night")
+    # test both string and integer attributes and values
+    attrs = {LEMMA: "boys", "ENT_TYPE": doc.vocab.strings["ORG"]}
+    with doc.retokenize() as retokenizer:
+        retokenizer.merge(doc[4:7], attrs=attrs)
+    assert len(doc) == 7
+    assert doc[4].text == "the beach boys"
+    assert doc[4].lemma_ == "boys"
+    assert doc[4].ent_type_ == "ORG"
+
+
+@pytest.mark.xfail
+def test_doc_retokenize_lex_attrs(en_tokenizer):
+    """Test that lexical attributes can be changed (see #2390)."""
+    doc = en_tokenizer("WKRO played beach boys songs")
+    assert not any(token.is_stop for token in doc)
+    with doc.retokenize() as retokenizer:
+        retokenizer.merge(doc[2:4], attrs={"LEMMA": "boys", "IS_STOP": True})
+    assert doc[2].text == "beach boys"
+    assert doc[2].lemma_ == "boys"
+    assert doc[2].is_stop
+    new_doc = Doc(doc.vocab, words=["beach boys"])
+    assert new_doc[0].is_stop
+
+
+def test_doc_retokenize_spans_merge_tokens(en_tokenizer):
     text = "Los Angeles start."
     heads = [1, 1, 0, -1]
     tokens = en_tokenizer(text)
@@ -25,7 +100,7 @@ def test_spans_merge_tokens(en_tokenizer):
     assert doc[0].ent_type_ == "GPE"
 
 
-def test_spans_merge_heads(en_tokenizer):
+def test_doc_retokenize_spans_merge_heads(en_tokenizer):
     text = "I found a pilates class near work."
     heads = [1, 0, 2, 1, -3, -1, -1, -6]
     tokens = en_tokenizer(text)
@@ -43,7 +118,7 @@ def test_spans_merge_heads(en_tokenizer):
     assert doc[5].head.i == 4
 
 
-def test_spans_merge_non_disjoint(en_tokenizer):
+def test_doc_retokenize_spans_merge_non_disjoint(en_tokenizer):
     text = "Los Angeles start."
     doc = en_tokenizer(text)
     with pytest.raises(ValueError):
@@ -58,7 +133,7 @@ def test_spans_merge_non_disjoint(en_tokenizer):
             )
 
 
-def test_span_np_merges(en_tokenizer):
+def test_doc_retokenize_span_np_merges(en_tokenizer):
     text = "displaCy is a parse tool built with Javascript"
     heads = [1, 0, 2, 1, -3, -1, -1, -1]
     tokens = en_tokenizer(text)
@@ -87,7 +162,7 @@ def test_span_np_merges(en_tokenizer):
             retokenizer.merge(ent)
 
 
-def test_spans_entity_merge(en_tokenizer):
+def test_doc_retokenize_spans_entity_merge(en_tokenizer):
     # fmt: off
     text = "Stewart Lee is a stand up comedian who lives in England and loves Joe Pasquale.\n"
     heads = [1, 1, 0, 1, 2, -1, -4, 1, -2, -1, -1, -3, -10, 1, -2, -13, -1]
@@ -108,7 +183,7 @@ def test_spans_entity_merge(en_tokenizer):
     assert len(doc) == 15
 
 
-def test_spans_entity_merge_iob():
+def test_doc_retokenize_spans_entity_merge_iob():
     # Test entity IOB stays consistent after merging
     words = ["a", "b", "c", "d", "e"]
     doc = Doc(Vocab(), words=words)
@@ -147,7 +222,7 @@ def test_spans_entity_merge_iob():
     assert doc[4].ent_iob_ == "I"
 
 
-def test_spans_sentence_update_after_merge(en_tokenizer):
+def test_doc_retokenize_spans_sentence_update_after_merge(en_tokenizer):
     # fmt: off
     text = "Stewart Lee is a stand up comedian. He lives in England and loves Joe Pasquale."
     heads = [1, 1, 0, 1, 2, -1, -4, -5, 1, 0, -1, -1, -3, -4, 1, -2, -7]
@@ -155,7 +230,6 @@ def test_spans_sentence_update_after_merge(en_tokenizer):
             'punct', 'nsubj', 'ROOT', 'prep', 'pobj', 'cc', 'conj',
             'compound', 'dobj', 'punct']
     # fmt: on
-
     tokens = en_tokenizer(text)
     doc = get_doc(tokens.vocab, words=[t.text for t in tokens], heads=heads, deps=deps)
     sent1, sent2 = list(doc.sents)
@@ -169,7 +243,7 @@ def test_spans_sentence_update_after_merge(en_tokenizer):
     assert len(sent2) == init_len2 - 1
 
 
-def test_spans_subtree_size_check(en_tokenizer):
+def test_doc_retokenize_spans_subtree_size_check(en_tokenizer):
     # fmt: off
     text = "Stewart Lee is a stand up comedian who lives in England and loves Joe Pasquale"
     heads = [1, 1, 0, 1, 2, -1, -4, 1, -2, -1, -1, -3, -10, 1, -2]
@@ -177,7 +251,6 @@ def test_spans_subtree_size_check(en_tokenizer):
             "nsubj", "relcl", "prep", "pobj", "cc", "conj", "compound",
             "dobj"]
     # fmt: on
-
     tokens = en_tokenizer(text)
     doc = get_doc(tokens.vocab, words=[t.text for t in tokens], heads=heads, deps=deps)
     sent1 = list(doc.sents)[0]

spacy/tests/doc/test_retokenize_split.py

@@ -8,7 +8,7 @@ from spacy.tokens import Doc
 from ..util import get_doc
 
 
-def test_doc_split(en_vocab):
+def test_doc_retokenize_split(en_vocab):
     words = ["LosAngeles", "start", "."]
     heads = [1, 1, 0]
     doc = get_doc(en_vocab, words=words, heads=heads)
@@ -41,7 +41,7 @@ def test_doc_split(en_vocab):
     assert len(str(doc)) == 19
 
 
-def test_split_dependencies(en_vocab):
+def test_doc_retokenize_split_dependencies(en_vocab):
     doc = Doc(en_vocab, words=["LosAngeles", "start", "."])
     dep1 = doc.vocab.strings.add("amod")
     dep2 = doc.vocab.strings.add("subject")
@@ -56,7 +56,7 @@ def test_split_dependencies(en_vocab):
     assert doc[1].dep == dep2
 
 
-def test_split_heads_error(en_vocab):
+def test_doc_retokenize_split_heads_error(en_vocab):
     doc = Doc(en_vocab, words=["LosAngeles", "start", "."])
     # Not enough heads
     with pytest.raises(ValueError):
@@ -69,7 +69,7 @@ def test_split_heads_error(en_vocab):
             retokenizer.split(doc[0], ["Los", "Angeles"], [doc[1], doc[1], doc[1]])
 
 
-def test_spans_entity_merge_iob():
+def test_doc_retokenize_spans_entity_split_iob():
     # Test entity IOB stays consistent after merging
     words = ["abc", "d", "e"]
     doc = Doc(Vocab(), words=words)
@@ -84,7 +84,7 @@ def test_spans_entity_merge_iob():
     assert doc[3].ent_iob_ == "I"
 
 
-def test_spans_sentence_update_after_merge(en_vocab):
+def test_doc_retokenize_spans_sentence_update_after_split(en_vocab):
     # fmt: off
     words = ["StewartLee", "is", "a", "stand", "up", "comedian", ".", "He",
              "lives", "in", "England", "and", "loves", "JoePasquale", "."]
@@ -114,7 +114,7 @@ def test_spans_sentence_update_after_merge(en_vocab):
     assert len(sent2) == init_len2 + 1
 
 
-def test_split_orths_mismatch(en_vocab):
+def test_doc_retokenize_split_orths_mismatch(en_vocab):
     """Test that the regular retokenizer.split raises an error if the orths
     don't match the original token text. There might still be a method that
     allows this, but for the default use cases, merging and splitting should

spacy/tests/regression/test_issue1971.py

@@ -1,7 +1,6 @@
 # coding: utf8
 from __future__ import unicode_literals
 
-import pytest
 from spacy.matcher import Matcher
 from spacy.tokens import Token, Doc
 
@@ -28,7 +27,7 @@ def test_issue1971(en_vocab):
 def test_issue_1971_2(en_vocab):
     matcher = Matcher(en_vocab)
     pattern1 = [{"ORTH": "EUR", "LOWER": {"IN": ["eur"]}}, {"LIKE_NUM": True}]
-    pattern2 = [{"LIKE_NUM": True}, {"ORTH": "EUR"}] #{"IN": ["EUR"]}}]
+    pattern2 = [{"LIKE_NUM": True}, {"ORTH": "EUR"}]  # {"IN": ["EUR"]}}]
     doc = Doc(en_vocab, words=["EUR", "10", "is", "10", "EUR"])
     matcher.add("TEST1", None, pattern1, pattern2)
     matches = matcher(doc)
@@ -59,6 +58,5 @@ def test_issue_1971_4(en_vocab):
     pattern = [{"_": {"ext_a": "str_a", "ext_b": "str_b"}}] * 3
     matcher.add("TEST", None, pattern)
     matches = matcher(doc)
-    # Interesting: uncommenting this causes a segmentation fault, so there's
+    # Uncommenting this caused a segmentation fault
-    # definitely something going on here
+    assert len(matches) == 1
-    # assert len(matches) == 1

spacy/tests/regression/test_issue3288.py

@@ -1,7 +1,6 @@
 # coding: utf-8
 from __future__ import unicode_literals
 
-import pytest
 import numpy
 from spacy import displacy
 

spacy/tests/regression/test_issue3289.py

@@ -1,7 +1,6 @@
 # coding: utf-8
 from __future__ import unicode_literals
 
-import pytest
 from spacy.lang.en import English
 
 

spacy/util.py

@@ -315,6 +315,11 @@ def read_regex(path):
 
 
 def compile_prefix_regex(entries):
+    """Compile a list of prefix rules into a regex object.
+
+    entries (tuple): The prefix rules, e.g. spacy.lang.punctuation.TOKENIZER_PREFIXES.
+    RETURNS (regex object): The regex object. to be used for Tokenizer.prefix_search.
+    """
     if "(" in entries:
         # Handle deprecated data
         expression = "|".join(
@@ -327,11 +332,21 @@ def compile_prefix_regex(entries):
 
 
 def compile_suffix_regex(entries):
+    """Compile a list of suffix rules into a regex object.
+
+    entries (tuple): The suffix rules, e.g. spacy.lang.punctuation.TOKENIZER_SUFFIXES.
+    RETURNS (regex object): The regex object. to be used for Tokenizer.suffix_search.
+    """
     expression = "|".join([piece + "$" for piece in entries if piece.strip()])
     return re.compile(expression)
 
 
 def compile_infix_regex(entries):
+    """Compile a list of infix rules into a regex object.
+
+    entries (tuple): The infix rules, e.g. spacy.lang.punctuation.TOKENIZER_INFIXES.
+    RETURNS (regex object): The regex object. to be used for Tokenizer.infix_finditer.
+    """
     expression = "|".join([piece for piece in entries if piece.strip()])
     return re.compile(expression)
 

website/docs/api/top-level.md

@@ -504,6 +504,57 @@ an error if key doesn't match `ORTH` values.
 | `*addition_dicts` | dicts | Exception dictionaries to add to the base exceptions, in order. |
 | **RETURNS**       | dict  | Combined tokenizer exceptions.                                  |
 
+### util.compile_prefix_regex {#util.compile_prefix_regex tag="function"}
+
+Compile a sequence of prefix rules into a regex object.
+
+> #### Example
+>
+> ```python
+> prefixes = ("§", "%", "=", r"\+")
+> prefix_regex = util.compile_prefix_regex(prefixes)
+> nlp.tokenizer.prefix_search = prefix_regex.search
+> ```
+
+| Name        | Type                                                          | Description                                                                                                                               |
+| ----------- | ------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------- |
+| `entries`   | tuple                                                         | The prefix rules, e.g. [`lang.punctuation.TOKENIZER_PREFIXES`](https://github.com/explosion/spaCy/tree/master/spacy/lang/punctuation.py). |
+| **RETURNS** | [regex](https://docs.python.org/3/library/re.html#re-objects) | The regex object. to be used for [`Tokenizer.prefix_search`](/api/tokenizer#attributes).                                                  |
+
+### util.compile_suffix_regex {#util.compile_suffix_regex tag="function"}
+
+Compile a sequence of suffix rules into a regex object.
+
+> #### Example
+>
+> ```python
+> suffixes = ("'s", "'S", r"(?<=[0-9])\+")
+> suffix_regex = util.compile_suffix_regex(suffixes)
+> nlp.tokenizer.suffix_search = suffix_regex.search
+> ```
+
+| Name        | Type                                                          | Description                                                                                                                               |
+| ----------- | ------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------- |
+| `entries`   | tuple                                                         | The suffix rules, e.g. [`lang.punctuation.TOKENIZER_SUFFIXES`](https://github.com/explosion/spaCy/tree/master/spacy/lang/punctuation.py). |
+| **RETURNS** | [regex](https://docs.python.org/3/library/re.html#re-objects) | The regex object. to be used for [`Tokenizer.suffix_search`](/api/tokenizer#attributes).                                                  |
+
+### util.compile_infix_regex {#util.compile_infix_regex tag="function"}
+
+Compile a sequence of infix rules into a regex object.
+
+> #### Example
+>
+> ```python
+> infixes = ("…", "-", "—", r"(?<=[0-9])[+\-\*^](?=[0-9-])")
+> infix_regex = util.compile_infix_regex(infixes)
+> nlp.tokenizer.infix_finditer = infix_regex.finditer
+> ```
+
+| Name        | Type                                                          | Description                                                                                                                             |
+| ----------- | ------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
+| `entries`   | tuple                                                         | The infix rules, e.g. [`lang.punctuation.TOKENIZER_INFIXES`](https://github.com/explosion/spaCy/tree/master/spacy/lang/punctuation.py). |
+| **RETURNS** | [regex](https://docs.python.org/3/library/re.html#re-objects) | The regex object. to be used for [`Tokenizer.infix_finditer`](/api/tokenizer#attributes).                                               |
+
 ### util.minibatch {#util.minibatch tag="function" new="2"}
 
 Iterate over batches of items. `size` may be an iterator, so that batch-size can

website/docs/usage/linguistic-features.md

@@ -812,6 +812,40 @@ only be applied at the **end of a token**, so your expression should end with a
 
 </Infobox>
 
+#### Adding to existing rule sets {#native-tokenizer-additions}
+
+In many situations, you don't necessarily need entirely custom rules. Sometimes
+you just want to add another character to the prefixes, suffixes or infixes. The
+default prefix, suffix and infix rules are available via the `nlp` object's
+`Defaults` and the [`Tokenizer.suffix_search`](/api/tokenizer#attributes)
+attribute is writable, so you can overwrite it with a compiled regular
+expression object using of the modified default rules. spaCy ships with utility
+functions to help you compile the regular expressions – for example,
+[`compile_suffix_regex`](/api/top-level#util.compile_suffix_regex):
+
+```python
+suffixes = nlp.Defaults.suffixes + (r'''-+$''',)
+suffix_regex = spacy.util.compile_suffix_regex(suffixes)
+nlp.tokenizer.suffix_search = suffix_regex.search
+```
+
+For an overview of the default regular expressions, see
+[`lang/punctuation.py`](https://github.com/explosion/spaCy/blob/master/spacy/lang/punctuation.py).
+The `Tokenizer.suffix_search` attribute should be a function which takes a
+unicode string and returns a **regex match object** or `None`. Usually we use
+the `.search` attribute of a compiled regex object, but you can use some other
+function that behaves the same way.
+
+<Infobox title="Important note" variant="warning">
+
+If you're using a statistical model, writing to the `nlp.Defaults` or
+`English.Defaults` directly won't work, since the regular expressions are read
+from the model and will be compiled when you load it. You'll only see the effect
+if you call [`spacy.blank`](/api/top-level#spacy.blank) or
+`Defaults.create_tokenizer()`.
+
+</Infobox>
+
 ### Hooking an arbitrary tokenizer into the pipeline {#custom-tokenizer}
 
 The tokenizer is the first component of the processing pipeline and the only one