Add AttributeRuler for token attribute exceptions (#5842)

* Add AttributeRuler for token attribute exceptions

Add the `AttributeRuler` to handle exceptions for token-level
attributes. The `AttributeRuler` uses `Matcher` patterns to identify
target spans and applies the specified attributes to the token at the
provided index in the matched span. A negative index can be used to
index from the end of the matched span. The retokenizer is used to
"merge" the individual tokens and assign them the provided attributes.

Helper functions can import existing tag maps and morph rules to the
corresponding `Matcher` patterns.

There is an additional minor bug fix for `MORPH` attributes in the
retokenizer to correctly normalize the values and to handle `MORPH`
alongside `_` in an attrs dict.

* Fix default name

* Update name in error message

* Extend AttributeRuler functionality

* Add option to initialize with a dict of AttributeRuler patterns

* Instead of silently discarding overlapping matches (the default
behavior for the retokenizer if only the attrs differ), split the
matches into disjoint sets and retokenize each set separately. This
allows, for instance, one pattern to set the POS and another pattern to
set the lemma. (If two matches modify the same attribute, it looks like
the attrs are applied in the order they were added, but it may not be
deterministic?)

* Improve types

* Sort spans before processing

* Fix index boundaries in Span

* Refactor retokenizer to separate attrs methods

Add top-level `normalize_token_attrs` and `set_token_attrs` methods.

* Update AttributeRuler to use refactored methods

Update `AttributeRuler` to replace use of full retokenizer with only the
relevant methods for normalizing and setting attributes for a single
token.

* Update spacy/pipeline/attributeruler.py

Co-authored-by: Ines Montani <ines@ines.io>

* Make API more similar to EntityRuler

* Add `AttributeRuler.add_patterns` to add patterns from a list of dicts
* Return list of dicts as property `AttributeRuler.patterns`

* Make attrs_unnormed private

* Add test loading patterns from assets

* Revert "Fix index boundaries in Span"

This reverts commit 8f8a5c3386.

* Add Span index boundary checks (#5861)

* Add Span index boundary checks

* Return Span-specific IndexError in all cases

* Simplify and fix if/else

Co-authored-by: Ines Montani <ines@ines.io>

spacy/errors.py

@@ -610,6 +610,9 @@ class Errors:
              "initializing the pipeline:\n"
              'cfg = {"tokenizer": {"segmenter": "pkuseg", "pkuseg_model": name_or_path}}\n'
              'nlp = Chinese(config=cfg)')
+    E1001 = ("Target token outside of matched span for match with tokens "
+             "'{span}' and offset '{index}' matched by patterns '{patterns}'.")
+    E1002 = ("Span index out of range.")
 
 
 @add_codes

spacy/pipeline/__init__.py

@@ -1,3 +1,4 @@
+from .attributeruler import AttributeRuler
 from .dep_parser import DependencyParser
 from .entity_linker import EntityLinker
 from .ner import EntityRecognizer
@@ -13,6 +14,7 @@ from .tok2vec import Tok2Vec
 from .functions import merge_entities, merge_noun_chunks, merge_subtokens
 
 __all__ = [
+    "AttributeRuler",
     "DependencyParser",
     "EntityLinker",
     "EntityRecognizer",

spacy/pipeline/attributeruler.py

@@ -0,0 +1,268 @@
+import srsly
+from typing import List, Dict, Union, Iterable, Any, Optional
+from pathlib import Path
+
+from .pipe import Pipe
+from ..errors import Errors
+from ..language import Language
+from ..matcher import Matcher
+from ..symbols import IDS
+from ..tokens import Doc, Span
+from ..tokens._retokenize import normalize_token_attrs, set_token_attrs
+from ..vocab import Vocab
+from .. import util
+
+
+MatcherPatternType = List[Dict[Union[int, str], Any]]
+AttributeRulerPatternType = Dict[str, Union[MatcherPatternType, Dict, int]]
+
+
+@Language.factory(
+    "attribute_ruler",
+)
+def make_attribute_ruler(
+    nlp: Language,
+    name: str,
+    pattern_dicts: Optional[Iterable[AttributeRulerPatternType]] = None,
+):
+    return AttributeRuler(nlp.vocab, name, pattern_dicts=pattern_dicts)
+
+
+class AttributeRuler(Pipe):
+    """Set token-level attributes for tokens matched by Matcher patterns.
+    Additionally supports importing patterns from tag maps and morph rules.
+
+    DOCS: https://spacy.io/api/attributeruler
+    """
+
+    def __init__(
+        self,
+        vocab: Vocab,
+        name: str = "attribute_ruler",
+        *,
+        pattern_dicts: Optional[Iterable[AttributeRulerPatternType]] = None,
+    ) -> None:
+        """Initialize the AttributeRuler.
+
+        vocab (Vocab): The vocab.
+        name (str): The pipe name. Defaults to "attribute_ruler".
+        pattern_dicts (Iterable[Dict]): A list of pattern dicts with the keys as
+        the arguments to AttributeRuler.add (`patterns`/`attrs`/`index`) to add
+        as patterns.
+
+        RETURNS (AttributeRuler): The AttributeRuler component.
+
+        DOCS: https://spacy.io/api/attributeruler#init
+        """
+        self.name = name
+        self.vocab = vocab
+        self.matcher = Matcher(self.vocab)
+        self.attrs = []
+        self._attrs_unnormed = [] # store for reference
+        self.indices = []
+
+        if pattern_dicts:
+            self.add_patterns(pattern_dicts)
+
+    def __call__(self, doc: Doc) -> Doc:
+        """Apply the attributeruler to a Doc and set all attribute exceptions.
+
+        doc (Doc): The document to process.
+        RETURNS (Doc): The processed Doc.
+
+        DOCS: https://spacy.io/api/attributeruler#call
+        """
+        matches = self.matcher(doc)
+
+        for match_id, start, end in matches:
+            span = Span(doc, start, end, label=match_id)
+            attrs = self.attrs[span.label]
+            index = self.indices[span.label]
+            try:
+                token = span[index]
+            except IndexError:
+                raise ValueError(
+                    Errors.E1001.format(
+                        patterns=self.matcher.get(span.label),
+                        span=[t.text for t in span],
+                        index=index,
+                    )
+                )
+            set_token_attrs(token, attrs)
+        return doc
+
+    def load_from_tag_map(
+        self, tag_map: Dict[str, Dict[Union[int, str], Union[int, str]]]
+    ) -> None:
+        for tag, attrs in tag_map.items():
+            pattern = [{"TAG": tag}]
+            attrs, morph_attrs = _split_morph_attrs(attrs)
+            morph = self.vocab.morphology.add(morph_attrs)
+            attrs["MORPH"] = self.vocab.strings[morph]
+            self.add([pattern], attrs)
+
+    def load_from_morph_rules(
+        self, morph_rules: Dict[str, Dict[str, Dict[Union[int, str], Union[int, str]]]]
+    ) -> None:
+        for tag in morph_rules:
+            for word in morph_rules[tag]:
+                pattern = [{"ORTH": word, "TAG": tag}]
+                attrs = morph_rules[tag][word]
+                attrs, morph_attrs = _split_morph_attrs(attrs)
+                morph = self.vocab.morphology.add(morph_attrs)
+                attrs["MORPH"] = self.vocab.strings[morph]
+                self.add([pattern], attrs)
+
+    def add(
+        self, patterns: Iterable[MatcherPatternType], attrs: Dict, index: int = 0
+    ) -> None:
+        """Add Matcher patterns for tokens that should be modified with the
+        provided attributes. The token at the specified index within the
+        matched span will be assigned the attributes.
+
+        patterns (Iterable[List[Dict]]): A list of Matcher patterns.
+        attrs (Dict): The attributes to assign to the target token in the
+            matched span.
+        index (int): The index of the token in the matched span to modify. May
+            be negative to index from the end of the span. Defaults to 0.
+
+        DOCS: https://spacy.io/api/attributeruler#add
+        """
+        self.matcher.add(len(self.attrs), patterns)
+        self._attrs_unnormed.append(attrs)
+        attrs = normalize_token_attrs(self.vocab, attrs)
+        self.attrs.append(attrs)
+        self.indices.append(index)
+
+    def add_patterns(self, pattern_dicts: Iterable[AttributeRulerPatternType]) -> None:
+        for p in pattern_dicts:
+            self.add(**p)
+
+    @property
+    def patterns(self) -> List[AttributeRulerPatternType]:
+        all_patterns = []
+        for i in range(len(self.attrs)):
+            p = {}
+            p["patterns"] = self.matcher.get(i)[1]
+            p["attrs"] = self._attrs_unnormed[i]
+            p["index"] = self.indices[i]
+            all_patterns.append(p)
+        return all_patterns
+
+    def to_bytes(self, exclude: Iterable[str] = tuple()) -> bytes:
+        """Serialize the attributeruler to a bytestring.
+
+        exclude (Iterable[str]): String names of serialization fields to exclude.
+        RETURNS (bytes): The serialized object.
+
+        DOCS: https://spacy.io/api/attributeruler#to_bytes
+        """
+        serialize = {}
+        serialize["vocab"] = self.vocab.to_bytes
+        patterns = {k: self.matcher.get(k)[1] for k in range(len(self.attrs))}
+        serialize["patterns"] = lambda: srsly.msgpack_dumps(patterns)
+        serialize["attrs"] = lambda: srsly.msgpack_dumps(self.attrs)
+        serialize["indices"] = lambda: srsly.msgpack_dumps(self.indices)
+        return util.to_bytes(serialize, exclude)
+
+    def from_bytes(self, bytes_data: bytes, exclude: Iterable[str] = tuple()):
+        """Load the attributeruler from a bytestring.
+
+        bytes_data (bytes): The data to load.
+        exclude (Iterable[str]): String names of serialization fields to exclude.
+        returns (AttributeRuler): The loaded object.
+
+        DOCS: https://spacy.io/api/attributeruler#from_bytes
+        """
+        data = {"patterns": b""}
+
+        def load_patterns(b):
+            data["patterns"] = srsly.msgpack_loads(b)
+
+        def load_attrs(b):
+            self.attrs = srsly.msgpack_loads(b)
+
+        def load_indices(b):
+            self.indices = srsly.msgpack_loads(b)
+
+        deserialize = {
+            "vocab": lambda b: self.vocab.from_bytes(b),
+            "patterns": load_patterns,
+            "attrs": load_attrs,
+            "indices": load_indices,
+        }
+        util.from_bytes(bytes_data, deserialize, exclude)
+
+        if data["patterns"]:
+            for key, pattern in data["patterns"].items():
+                self.matcher.add(key, pattern)
+            assert len(self.attrs) == len(data["patterns"])
+            assert len(self.indices) == len(data["patterns"])
+
+        return self
+
+    def to_disk(self, path: Union[Path, str], exclude: Iterable[str] = tuple()) -> None:
+        """Serialize the attributeruler to disk.
+
+        path (Union[Path, str]): A path to a directory.
+        exclude (Iterable[str]): String names of serialization fields to exclude.
+        DOCS: https://spacy.io/api/attributeruler#to_disk
+        """
+        patterns = {k: self.matcher.get(k)[1] for k in range(len(self.attrs))}
+        serialize = {
+            "vocab": lambda p: self.vocab.to_disk(p),
+            "patterns": lambda p: srsly.write_msgpack(p, patterns),
+            "attrs": lambda p: srsly.write_msgpack(p, self.attrs),
+            "indices": lambda p: srsly.write_msgpack(p, self.indices),
+        }
+        util.to_disk(path, serialize, exclude)
+
+    def from_disk(
+        self, path: Union[Path, str], exclude: Iterable[str] = tuple()
+    ) -> None:
+        """Load the attributeruler from disk.
+
+        path (Union[Path, str]): A path to a directory.
+        exclude (Iterable[str]): String names of serialization fields to exclude.
+        DOCS: https://spacy.io/api/attributeruler#from_disk
+        """
+        data = {"patterns": b""}
+
+        def load_patterns(p):
+            data["patterns"] = srsly.read_msgpack(p)
+
+        def load_attrs(p):
+            self.attrs = srsly.read_msgpack(p)
+
+        def load_indices(p):
+            self.indices = srsly.read_msgpack(p)
+
+        deserialize = {
+            "vocab": lambda p: self.vocab.from_disk(p),
+            "patterns": load_patterns,
+            "attrs": load_attrs,
+            "indices": load_indices,
+        }
+        util.from_disk(path, deserialize, exclude)
+
+        if data["patterns"]:
+            for key, pattern in data["patterns"].items():
+                self.matcher.add(key, pattern)
+            assert len(self.attrs) == len(data["patterns"])
+            assert len(self.indices) == len(data["patterns"])
+
+        return self
+
+
+def _split_morph_attrs(attrs):
+    """Split entries from a tag map or morph rules dict into to two dicts, one
+    with the token-level features (POS, LEMMA) and one with the remaining
+    features, which are presumed to be individual MORPH features."""
+    other_attrs = {}
+    morph_attrs = {}
+    for k, v in attrs.items():
+        if k in "_" or k in IDS.keys() or k in IDS.values():
+            other_attrs[k] = v
+        else:
+            morph_attrs[k] = v
+    return other_attrs, morph_attrs

spacy/tests/doc/test_span.py

@@ -282,3 +282,15 @@ def test_span_eq_hash(doc, doc_not_parsed):
     assert hash(doc[0:2]) == hash(doc[0:2])
     assert hash(doc[0:2]) != hash(doc[1:3])
     assert hash(doc[0:2]) != hash(doc_not_parsed[0:2])
+
+
+def test_span_boundaries(doc):
+    start = 1
+    end = 5
+    span = doc[start:end]
+    for i in range(start, end):
+        assert span[i - start] == doc[i]
+    with pytest.raises(IndexError):
+        _ = span[-5]
+    with pytest.raises(IndexError):
+        _ = span[5]

spacy/tests/pipeline/test_attributeruler.py

@@ -0,0 +1,208 @@
+import pytest
+import numpy
+from spacy.lang.en import English
+from spacy.pipeline import AttributeRuler
+from spacy import util, registry
+
+from ..util import get_doc, make_tempdir
+
+
+@pytest.fixture
+def nlp():
+    return English()
+
+
+@pytest.fixture
+def pattern_dicts():
+    return [
+        {
+            "patterns": [[{"ORTH": "a"}], [{"ORTH": "irrelevant"}]],
+            "attrs": {"LEMMA": "the", "MORPH": "Case=Nom|Number=Plur"},
+        },
+        # one pattern sets the lemma
+        {"patterns": [[{"ORTH": "test"}]], "attrs": {"LEMMA": "cat"}},
+        # another pattern sets the morphology
+        {
+            "patterns": [[{"ORTH": "test"}]],
+            "attrs": {"MORPH": "Case=Nom|Number=Sing"},
+            "index": 0,
+        },
+    ]
+
+
+@registry.assets("attribute_ruler_patterns")
+def attribute_ruler_patterns():
+    return [
+        {
+            "patterns": [[{"ORTH": "a"}], [{"ORTH": "irrelevant"}]],
+            "attrs": {"LEMMA": "the", "MORPH": "Case=Nom|Number=Plur"},
+        },
+        # one pattern sets the lemma
+        {"patterns": [[{"ORTH": "test"}]], "attrs": {"LEMMA": "cat"}},
+        # another pattern sets the morphology
+        {
+            "patterns": [[{"ORTH": "test"}]],
+            "attrs": {"MORPH": "Case=Nom|Number=Sing"},
+            "index": 0,
+        },
+    ]
+
+
+@pytest.fixture
+def tag_map():
+    return {
+        ".": {"POS": "PUNCT", "PunctType": "peri"},
+        ",": {"POS": "PUNCT", "PunctType": "comm"},
+    }
+
+
+@pytest.fixture
+def morph_rules():
+    return {"DT": {"the": {"POS": "DET", "LEMMA": "a", "Case": "Nom"}}}
+
+
+def test_attributeruler_init(nlp, pattern_dicts):
+    a = nlp.add_pipe("attribute_ruler")
+    for p in pattern_dicts:
+        a.add(**p)
+
+    doc = nlp("This is a test.")
+    assert doc[2].lemma_ == "the"
+    assert doc[2].morph_ == "Case=Nom|Number=Plur"
+    assert doc[3].lemma_ == "cat"
+    assert doc[3].morph_ == "Case=Nom|Number=Sing"
+
+
+def test_attributeruler_init_patterns(nlp, pattern_dicts):
+    # initialize with patterns
+    a = nlp.add_pipe("attribute_ruler", config={"pattern_dicts": pattern_dicts})
+
+    doc = nlp("This is a test.")
+    assert doc[2].lemma_ == "the"
+    assert doc[2].morph_ == "Case=Nom|Number=Plur"
+    assert doc[3].lemma_ == "cat"
+    assert doc[3].morph_ == "Case=Nom|Number=Sing"
+
+    nlp.remove_pipe("attribute_ruler")
+
+    # initialize with patterns from asset
+    a = nlp.add_pipe("attribute_ruler", config={"pattern_dicts": {"@assets": "attribute_ruler_patterns"}})
+
+    doc = nlp("This is a test.")
+    assert doc[2].lemma_ == "the"
+    assert doc[2].morph_ == "Case=Nom|Number=Plur"
+    assert doc[3].lemma_ == "cat"
+    assert doc[3].morph_ == "Case=Nom|Number=Sing"
+
+
+def test_attributeruler_tag_map(nlp, tag_map):
+    a = AttributeRuler(nlp.vocab)
+    a.load_from_tag_map(tag_map)
+    doc = get_doc(
+        nlp.vocab,
+        words=["This", "is", "a", "test", "."],
+        tags=["DT", "VBZ", "DT", "NN", "."],
+    )
+    doc = a(doc)
+
+    for i in range(len(doc)):
+        if i == 4:
+            assert doc[i].pos_ == "PUNCT"
+            assert doc[i].morph_ == "PunctType=peri"
+        else:
+            assert doc[i].pos_ == ""
+            assert doc[i].morph_ == ""
+
+
+def test_attributeruler_morph_rules(nlp, morph_rules):
+    a = AttributeRuler(nlp.vocab)
+    a.load_from_morph_rules(morph_rules)
+    doc = get_doc(
+        nlp.vocab,
+        words=["This", "is", "the", "test", "."],
+        tags=["DT", "VBZ", "DT", "NN", "."],
+    )
+    doc = a(doc)
+
+    for i in range(len(doc)):
+        if i != 2:
+            assert doc[i].pos_ == ""
+            assert doc[i].morph_ == ""
+        else:
+            assert doc[2].pos_ == "DET"
+            assert doc[2].lemma_ == "a"
+            assert doc[2].morph_ == "Case=Nom"
+
+
+def test_attributeruler_indices(nlp):
+    a = nlp.add_pipe("attribute_ruler")
+    a.add(
+        [[{"ORTH": "a"}, {"ORTH": "test"}]],
+        {"LEMMA": "the", "MORPH": "Case=Nom|Number=Plur"},
+        index=0,
+    )
+    a.add(
+        [[{"ORTH": "This"}, {"ORTH": "is"}]],
+        {"LEMMA": "was", "MORPH": "Case=Nom|Number=Sing"},
+        index=1,
+    )
+    a.add([[{"ORTH": "a"}, {"ORTH": "test"}]], {"LEMMA": "cat"}, index=-1)
+
+    text = "This is a test."
+    doc = nlp(text)
+
+    for i in range(len(doc)):
+        if i == 1:
+            assert doc[i].lemma_ == "was"
+            assert doc[i].morph_ == "Case=Nom|Number=Sing"
+        elif i == 2:
+            assert doc[i].lemma_ == "the"
+            assert doc[i].morph_ == "Case=Nom|Number=Plur"
+        elif i == 3:
+            assert doc[i].lemma_ == "cat"
+        else:
+            assert doc[i].morph_ == ""
+
+    # raises an error when trying to modify a token outside of the match
+    a.add([[{"ORTH": "a"}, {"ORTH": "test"}]], {"LEMMA": "cat"}, index=2)
+    with pytest.raises(ValueError):
+        doc = nlp(text)
+
+    # raises an error when trying to modify a token outside of the match
+    a.add([[{"ORTH": "a"}, {"ORTH": "test"}]], {"LEMMA": "cat"}, index=10)
+    with pytest.raises(ValueError):
+        doc = nlp(text)
+
+
+def test_attributeruler_patterns_prop(nlp, pattern_dicts):
+    a = nlp.add_pipe("attribute_ruler")
+    a.add_patterns(pattern_dicts)
+
+    for p1, p2 in zip(pattern_dicts, a.patterns):
+        assert p1["patterns"] == p2["patterns"]
+        assert p1["attrs"] == p2["attrs"]
+        if p1.get("index"):
+            assert p1["index"] == p2["index"]
+
+
+def test_attributeruler_serialize(nlp, pattern_dicts):
+    a = nlp.add_pipe("attribute_ruler")
+    a.add_patterns(pattern_dicts)
+
+    text = "This is a test."
+    attrs = ["ORTH", "LEMMA", "MORPH"]
+    doc = nlp(text)
+
+    # bytes roundtrip
+    a_reloaded = AttributeRuler(nlp.vocab).from_bytes(a.to_bytes())
+    assert a.to_bytes() == a_reloaded.to_bytes()
+    doc1 = a_reloaded(nlp.make_doc(text))
+    numpy.array_equal(doc.to_array(attrs), doc1.to_array(attrs))
+
+    # disk roundtrip
+    with make_tempdir() as tmp_dir:
+        nlp.to_disk(tmp_dir)
+        nlp2 = util.load_model_from_path(tmp_dir)
+        doc2 = nlp2(text)
+        assert nlp2.get_pipe("attribute_ruler").to_bytes() == a.to_bytes()
+        assert numpy.array_equal(doc.to_array(attrs), doc2.to_array(attrs))

spacy/tokens/_retokenize.pyx

@@ -12,6 +12,7 @@ from .token cimport Token
 from ..lexeme cimport Lexeme, EMPTY_LEXEME
 from ..structs cimport LexemeC, TokenC
 from ..attrs cimport TAG, MORPH
+from ..vocab cimport Vocab
 
 from .underscore import is_writable_attr
 from ..attrs import intify_attrs
@@ -57,16 +58,7 @@ cdef class Retokenizer:
                 raise ValueError(Errors.E102.format(token=repr(token)))
             self.tokens_to_merge.add(token.i)
         self._spans_to_merge.append((span.start, span.end))
-        if "_" in attrs:  # Extension attributes
+        attrs = normalize_token_attrs(self.doc.vocab, attrs)
-            extensions = attrs["_"]
-            _validate_extensions(extensions)
-            attrs = {key: value for key, value in attrs.items() if key != "_"}
-            attrs = intify_attrs(attrs, strings_map=self.doc.vocab.strings)
-            attrs["_"] = extensions
-        else:
-            attrs = intify_attrs(attrs, strings_map=self.doc.vocab.strings)
-            if MORPH in attrs:
-                self.doc.vocab.morphology.add(self.doc.vocab.strings.as_string(attrs[MORPH]))
         self.merges.append((span, attrs))
 
     def split(self, Token token, orths, heads, attrs=SimpleFrozenDict()):
@@ -98,9 +90,11 @@ cdef class Retokenizer:
             # NB: Since we support {"KEY": [value, value]} syntax here, this
             # will only "intify" the keys, not the values
             attrs = intify_attrs(attrs, strings_map=self.doc.vocab.strings)
-            if MORPH in attrs:
+        if MORPH in attrs:
-                for morph in attrs[MORPH]:
+            for i, morph in enumerate(attrs[MORPH]):
-                    self.doc.vocab.morphology.add(self.doc.vocab.strings.as_string(morph))
+                # add and set to normalized value
+                morph = self.doc.vocab.morphology.add(self.doc.vocab.strings.as_string(morph))
+                attrs[MORPH][i] = morph
         head_offsets = []
         for head in heads:
             if isinstance(head, Token):
@@ -224,21 +218,7 @@ def _merge(Doc doc, merges):
         token.lex = lex
         # We set trailing space here too
         token.spacy = doc.c[spans[token_index].end-1].spacy
-        py_token = span[0]
+        set_token_attrs(span[0], attributes)
-        # Assign attributes
-        for attr_name, attr_value in attributes.items():
-            if attr_name == "_":  # Set extension attributes
-                for ext_attr_key, ext_attr_value in attr_value.items():
-                    py_token._.set(ext_attr_key, ext_attr_value)
-            elif attr_name == TAG:
-                doc.vocab.morphology.assign_tag(token, attr_value)
-            else:
-                # Set attributes on both token and lexeme to take care of token
-                # attribute vs. lexical attribute without having to enumerate
-                # them. If an attribute name is not valid, set_struct_attr will
-                # ignore it.
-                Token.set_struct_attr(token, attr_name, attr_value)
-                Lexeme.set_struct_attr(<LexemeC*>lex, attr_name, attr_value)
     # Begin by setting all the head indices to absolute token positions
     # This is easier to work with for now than the offsets
     # Before thinking of something simpler, beware the case where a
@@ -423,3 +403,40 @@ cdef make_iob_consistent(TokenC* tokens, int length):
     for i in range(1, length):
         if tokens[i].ent_iob == 1 and tokens[i - 1].ent_type != tokens[i].ent_type:
             tokens[i].ent_iob = 3
+
+
+def normalize_token_attrs(Vocab vocab, attrs):
+    if "_" in attrs:  # Extension attributes
+        extensions = attrs["_"]
+        print("EXTENSIONS", extensions)
+        _validate_extensions(extensions)
+        attrs = {key: value for key, value in attrs.items() if key != "_"}
+        attrs = intify_attrs(attrs, strings_map=vocab.strings)
+        attrs["_"] = extensions
+    else:
+        attrs = intify_attrs(attrs, strings_map=vocab.strings)
+    if MORPH in attrs:
+        # add and set to normalized value
+        morph = vocab.morphology.add(vocab.strings.as_string(attrs[MORPH]))
+        attrs[MORPH] = morph
+    return attrs
+
+
+def set_token_attrs(Token py_token, attrs):
+    cdef TokenC* token = py_token.c
+    cdef const LexemeC* lex = token.lex
+    cdef Doc doc = py_token.doc
+    # Assign attributes
+    for attr_name, attr_value in attrs.items():
+        if attr_name == "_":  # Set extension attributes
+            for ext_attr_key, ext_attr_value in attr_value.items():
+                py_token._.set(ext_attr_key, ext_attr_value)
+        elif attr_name == TAG:
+            doc.vocab.morphology.assign_tag(token, attr_value)
+        else:
+            # Set attributes on both token and lexeme to take care of token
+            # attribute vs. lexical attribute without having to enumerate
+            # them. If an attribute name is not valid, set_struct_attr will
+            # ignore it.
+            Token.set_struct_attr(token, attr_name, attr_value)
+            Lexeme.set_struct_attr(<LexemeC*>lex, attr_name, attr_value)

spacy/tokens/span.pyx

@@ -176,9 +176,13 @@ cdef class Span:
             return Span(self.doc, start + self.start, end + self.start)
         else:
             if i < 0:
-                return self.doc[self.end + i]
+                token_i = self.end + i
             else:
-                return self.doc[self.start + i]
+                token_i = self.start + i
+            if self.start <= token_i < self.end:
+                return self.doc[token_i]
+            else:
+                raise IndexError(Errors.E1002)
 
     def __iter__(self):
         """Iterate over `Token` objects.