negative tag annotation (#8731)

* unit test to unlearn tag via negative annotation

* bump thinc to 8.0.8

pyproject.toml

@@ -5,7 +5,7 @@ requires = [
     "cymem>=2.0.2,<2.1.0",
     "preshed>=3.0.2,<3.1.0",
     "murmurhash>=0.28.0,<1.1.0",
-    "thinc>=8.0.7,<8.1.0",
+    "thinc>=8.0.8,<8.1.0",
     "blis>=0.4.0,<0.8.0",
     "pathy",
     "numpy>=1.15.0",

requirements.txt

@@ -2,7 +2,7 @@
 spacy-legacy>=3.0.7,<3.1.0
 cymem>=2.0.2,<2.1.0
 preshed>=3.0.2,<3.1.0
-thinc>=8.0.7,<8.1.0
+thinc>=8.0.8,<8.1.0
 blis>=0.4.0,<0.8.0
 ml_datasets>=0.2.0,<0.3.0
 murmurhash>=0.28.0,<1.1.0

setup.cfg

@@ -37,14 +37,14 @@ setup_requires =
     cymem>=2.0.2,<2.1.0
     preshed>=3.0.2,<3.1.0
     murmurhash>=0.28.0,<1.1.0
-    thinc>=8.0.7,<8.1.0
+    thinc>=8.0.8,<8.1.0
 install_requires =
     # Our libraries
     spacy-legacy>=3.0.7,<3.1.0
     murmurhash>=0.28.0,<1.1.0
     cymem>=2.0.2,<2.1.0
     preshed>=3.0.2,<3.1.0
-    thinc>=8.0.7,<8.1.0
+    thinc>=8.0.8,<8.1.0
     blis>=0.4.0,<0.8.0
     wasabi>=0.8.1,<1.1.0
     srsly>=2.4.1,<3.0.0

spacy/pipeline/tagger.pyx

@@ -222,7 +222,7 @@ class Tagger(TrainablePipe):
         DOCS: https://spacy.io/api/tagger#get_loss
         """
         validate_examples(examples, "Tagger.get_loss")
-        loss_func = SequenceCategoricalCrossentropy(names=self.labels, normalize=False)
+        loss_func = SequenceCategoricalCrossentropy(names=self.labels, normalize=False, neg_prefix="!")
         # Convert empty tag "" to missing value None so that both misaligned
         # tokens and tokens with missing annotation have the default missing
         # value None.

spacy/tests/pipeline/test_tagger.py

@@ -182,6 +182,17 @@ def test_overfitting_IO():
     assert_equal(batch_deps_1, batch_deps_2)
     assert_equal(batch_deps_1, no_batch_deps)
 
+    # Try to unlearn the first 'N' tag with negative annotation
+    neg_ex = Example.from_dict(nlp.make_doc(test_text), {"tags": ["!N", "V", "J", "N"]})
+
+    for i in range(20):
+        losses = {}
+        nlp.update([neg_ex], sgd=optimizer, losses=losses)
+
+    # test the "untrained" tag
+    doc3 = nlp(test_text)
+    assert doc3[0].tag_ != "N"
+
 
 def test_tagger_requires_labels():
     nlp = English()