Update morphologizer (#5766)

* update `Morphologizer.begin_training` for use with `Example`

* make init and begin_training more consistent

* add `Morphology.normalize_features` to normalize outside of
`Morphology.add`

* make sure `get_loss` doesn't create unknown labels when the POS and
morph alignments differ

spacy/morphology.pyx

@@ -58,7 +58,7 @@ cdef class Morphology:
     FEATURE_SEP = "|"
     FIELD_SEP = "="
     VALUE_SEP = ","
-    EMPTY_MORPH = "_"
+    EMPTY_MORPH = "_" # not an empty string so that the PreshMap key is not 0
 
     def __init__(self, StringStore strings, tag_map, lemmatizer, exc=None):
         self.mem = Pool()
@@ -117,13 +117,7 @@ cdef class Morphology:
         if not isinstance(features, dict):
             warnings.warn(Warnings.W100.format(feature=features))
             features = {}
-        features = _normalize_props(features)
         string_features = {self.strings.as_string(field): self.strings.as_string(values) for field, values in features.items()}
-        # normalized UFEATS string with sorted fields and values
-        norm_feats_string = self.FEATURE_SEP.join(sorted([
-                self.FIELD_SEP.join([field, values])
-            for field, values in string_features.items()
-        ]))
         # intified ("Field", "Field=Value") pairs
         field_feature_pairs = []
         for field in sorted(string_features):
@@ -137,6 +131,7 @@ cdef class Morphology:
         # the hash key for the tag is either the hash of the normalized UFEATS
         # string or the hash of an empty placeholder (using the empty string
         # would give a hash key of 0, which is not good for PreshMap)
+        norm_feats_string = self.normalize_features(features)
         if norm_feats_string:
             tag.key = self.strings.add(norm_feats_string)
         else:
@@ -144,6 +139,26 @@ cdef class Morphology:
         self.insert(tag)
         return tag.key
 
+    def normalize_features(self, features):
+        """Create a normalized UFEATS string from a features string or dict.
+
+        features (Union[dict, str]): Features as dict or UFEATS string.
+        RETURNS (str): Features as normalized UFEATS string.
+        """
+        if isinstance(features, str):
+            features = self.feats_to_dict(features)
+        if not isinstance(features, dict):
+            warnings.warn(Warnings.W100.format(feature=features))
+            features = {}
+        features = _normalize_props(features)
+        string_features = {self.strings.as_string(field): self.strings.as_string(values) for field, values in features.items()}
+        # normalized UFEATS string with sorted fields and values
+        norm_feats_string = self.FEATURE_SEP.join(sorted([
+                self.FIELD_SEP.join([field, values])
+            for field, values in string_features.items()
+        ]))
+        return norm_feats_string or self.EMPTY_MORPH
+
     cdef MorphAnalysisC create_morph_tag(self, field_feature_pairs) except *:
         """Creates a MorphAnalysisC from a list of intified
         ("Field", "Field=Value") tuples where fields with multiple values have

spacy/pipeline/morphologizer.pyx

@@ -23,29 +23,45 @@ from .defaults import default_morphologizer
 @component("morphologizer", assigns=["token.morph", "token.pos"], default_model=default_morphologizer)
 class Morphologizer(Tagger):
 
+    POS_FEAT = "POS"
+
     def __init__(self, vocab, model, **cfg):
         self.vocab = vocab
         self.model = model
         self._rehearsal_model = None
         self.cfg = dict(sorted(cfg.items()))
-        self.cfg.setdefault("labels", {})
+        # to be able to set annotations without string operations on labels,
-        self.cfg.setdefault("morph_pos", {})
+        # store mappings from morph+POS labels to token-level annotations:
+        # 1) labels_morph stores a mapping from morph+POS->morph
+        self.cfg.setdefault("labels_morph", {})
+        # 2) labels_pos stores a mapping from morph+POS->POS
+        self.cfg.setdefault("labels_pos", {})
+        # add mappings for empty morph
+        self.cfg["labels_morph"][Morphology.EMPTY_MORPH] = Morphology.EMPTY_MORPH
+        self.cfg["labels_pos"][Morphology.EMPTY_MORPH] = POS_IDS[""]
 
     @property
     def labels(self):
-        return tuple(self.cfg["labels"].keys())
+        return tuple(self.cfg["labels_morph"].keys())
 
     def add_label(self, label):
         if not isinstance(label, str):
             raise ValueError(Errors.E187)
         if label in self.labels:
             return 0
-        morph = Morphology.feats_to_dict(label)
+        # normalize label
-        norm_morph_pos = self.vocab.strings[self.vocab.morphology.add(morph)]
+        norm_label = self.vocab.morphology.normalize_features(label)
-        pos = morph.get("POS", "")
+        # extract separate POS and morph tags
-        if norm_morph_pos not in self.cfg["labels"]:
+        label_dict = Morphology.feats_to_dict(label)
-            self.cfg["labels"][norm_morph_pos] = norm_morph_pos
+        pos = label_dict.get(self.POS_FEAT, "")
-            self.cfg["morph_pos"][norm_morph_pos] = POS_IDS[pos]
+        if self.POS_FEAT in label_dict:
+            label_dict.pop(self.POS_FEAT)
+        # normalize morph string and add to morphology table
+        norm_morph = self.vocab.strings[self.vocab.morphology.add(label_dict)]
+        # add label mappings
+        if norm_label not in self.cfg["labels_morph"]:
+            self.cfg["labels_morph"][norm_label] = norm_morph
+            self.cfg["labels_pos"][norm_label] = POS_IDS[pos]
         return 1
 
     def begin_training(self, get_examples=lambda: [], pipeline=None, sgd=None,
@@ -53,14 +69,16 @@ class Morphologizer(Tagger):
         for example in get_examples():
             for i, token in enumerate(example.reference):
                 pos = token.pos_
-                morph = token.morph
+                morph = token.morph_
-                norm_morph = self.vocab.strings[self.vocab.morphology.add(morph)]
+                # create and add the combined morph+POS label
+                morph_dict = Morphology.feats_to_dict(morph)
                 if pos:
-                    morph["POS"] = pos
+                    morph_dict[self.POS_FEAT] = pos
-                norm_morph_pos = self.vocab.strings[self.vocab.morphology.add(morph)]
+                norm_label = self.vocab.strings[self.vocab.morphology.add(morph_dict)]
-                if norm_morph_pos not in self.cfg["labels"]:
+                # add label->morph and label->POS mappings
-                    self.cfg["labels"][norm_morph_pos] = norm_morph
+                if norm_label not in self.cfg["labels_morph"]:
-                    self.cfg["morph_pos"][norm_morph_pos] = POS_IDS[pos]
+                    self.cfg["labels_morph"][norm_label] = morph
+                    self.cfg["labels_pos"][norm_label] = POS_IDS[pos]
         self.set_output(len(self.labels))
         self.model.initialize()
         link_vectors_to_models(self.vocab)
@@ -79,8 +97,8 @@ class Morphologizer(Tagger):
                 doc_tag_ids = doc_tag_ids.get()
             for j, tag_id in enumerate(doc_tag_ids):
                 morph = self.labels[tag_id]
-                doc.c[j].morph = self.vocab.morphology.add(self.cfg["labels"][morph])
+                doc.c[j].morph = self.vocab.morphology.add(self.cfg["labels_morph"][morph])
-                doc.c[j].pos = self.cfg["morph_pos"][morph]
+                doc.c[j].pos = self.cfg["labels_pos"][morph]
 
             doc.is_morphed = True
 
@@ -94,14 +112,17 @@ class Morphologizer(Tagger):
             for i in range(len(morphs)):
                 pos = pos_tags[i]
                 morph = morphs[i]
-                feats = Morphology.feats_to_dict(morph)
+                # POS may align (same value for multiple tokens) when morph
+                # doesn't, so if either is None, treat both as None here so that
+                # truths doesn't end up with an unknown morph+POS combination
+                if pos is None or morph is None:
+                    pos = None
+                    morph = None
+                label_dict = Morphology.feats_to_dict(morph)
                 if pos:
-                    feats["POS"] = pos
+                    label_dict[self.POS_FEAT] = pos
-                if len(feats) > 0:
+                label = self.vocab.strings[self.vocab.morphology.add(label_dict)]
-                    morph = self.vocab.strings[self.vocab.morphology.add(feats)]
+                eg_truths.append(label)
-                if morph == "":
-                    morph = Morphology.EMPTY_MORPH
-                eg_truths.append(morph)
             truths.append(eg_truths)
         d_scores, loss = loss_func(scores, truths)
         if self.model.ops.xp.isnan(loss):

spacy/tests/pipeline/test_morphologizer.py

@@ -5,6 +5,7 @@ from spacy.gold import Example
 from spacy.lang.en import English
 from spacy.language import Language
 from spacy.tests.util import make_tempdir
+from spacy.morphology import Morphology
 
 
 def test_label_types():
@@ -23,9 +24,10 @@ TRAIN_DATA = [
             "pos": ["NOUN", "VERB", "ADJ", "NOUN"],
         },
     ),
+    # test combinations of morph+POS
     (
         "Eat blue ham",
-        {"morphs": ["Feat=V", "Feat=J", "Feat=N"], "pos": ["VERB", "ADJ", "NOUN"]},
+        {"morphs": ["Feat=V", "", ""], "pos": ["", "ADJ", ""]},
     ),
 ]
 
@@ -38,7 +40,12 @@ def test_overfitting_IO():
     for inst in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(inst[0]), inst[1]))
         for morph, pos in zip(inst[1]["morphs"], inst[1]["pos"]):
-            morphologizer.add_label(morph + "|POS=" + pos)
+            if morph and pos:
+                morphologizer.add_label(morph + Morphology.FEATURE_SEP + "POS" + Morphology.FIELD_SEP + pos)
+            elif pos:
+                morphologizer.add_label("POS" + Morphology.FIELD_SEP + pos)
+            elif morph:
+                morphologizer.add_label(morph)
     nlp.add_pipe(morphologizer)
     optimizer = nlp.begin_training()
 
@@ -48,19 +55,27 @@ def test_overfitting_IO():
     assert losses["morphologizer"] < 0.00001
 
     # test the trained model
-    test_text = "I like blue eggs"
+    test_text = "I like blue ham"
     doc = nlp(test_text)
     gold_morphs = [
-        "Feat=N|POS=NOUN",
+        "Feat=N",
-        "Feat=V|POS=VERB",
+        "Feat=V",
-        "Feat=J|POS=ADJ",
+        "",
-        "Feat=N|POS=NOUN",
+        "",
+    ]
+    gold_pos_tags = [
+        "NOUN",
+        "VERB",
+        "ADJ",
+        "",
     ]
     assert [t.morph_ for t in doc] == gold_morphs
+    assert [t.pos_ for t in doc] == gold_pos_tags
 
     # Also test the results are still the same after IO
     with make_tempdir() as tmp_dir:
         nlp.to_disk(tmp_dir)
         nlp2 = util.load_model_from_path(tmp_dir)
         doc2 = nlp2(test_text)
-        assert gold_morphs == [t.morph_ for t in doc2]
+        assert [t.morph_ for t in doc2] == gold_morphs
+        assert [t.pos_ for t in doc2] == gold_pos_tags