spaCy/website/docs/usage/v2-2.md

title	teaser	menu
What's New in v2.2	New features, backwards incompatibilities and migration guide	New Features features Backwards Incompatibilities incompat

New Features

spaCy v2.2 features improved statistical models, new pretrained models for Norwegian and Lithuanian, better Dutch NER, as well as a new mechanism for storing language data that makes the installation about 15× smaller on disk. We've also added a new API for entity linking, a new class to efficiently serialize annotations, an improved and 10× faster phrase matching engine, built-in scoring and CLI training for text classification and a new command to analyze and debug training data. For the full changelog, see the release notes on GitHub.

Better pretrained models and more languages

Example

python -m spacy download nl_core_news_sm
python -m spacy download nb_core_news_sm
python -m spacy download lt_core_news_sm

The new version also features new and re-trained models for all languages and resolves a number of data bugs. The Dutch model has been retrained with a new and custom-labelled NER corpus using the same extended label scheme as the English models. It should now produce significantly better NER results overall. We've also added new core models for Norwegian (MIT) and Lithuanian (CC BY-SA).

Usage: Models directory **Benchmarks: ** Release notes

Entity linking API

Example

nlp = spacy.load("my_custom_wikidata_model")
doc = nlp("Ada Lovelace was born in London")
print([(e.text, e.label_, e.kb_id_) for e in doc.ents])
# [('Ada Lovelace', 'PERSON', 'Q7259'), ('London', 'GPE', 'Q84')]

Entity linking lets you ground named entities into the "real world". We're excited to now provide a built-in API for training entity linking models and resolving textual entities to unique identifiers from a knowledge base. The annotated KB identifier is accessible as either a hash value or as a string from a Span or Token object. For more details on entity linking in spaCy, check out Sofie's talk at spaCy IRL 2019.

API: EntityLinker, KnowledgeBase **Code: ** bin/wiki_entity_linking **Usage: ** Entity linking, Training an entity linking model

Serializable lookup table and dictionary API

Example

data = {"foo": "bar"}
nlp.vocab.lookups.add_table("my_dict", data)

def custom_component(doc):
   table = doc.vocab.lookups.get_table("my_dict")
   print(table.get("foo"))  # look something up
   return doc

The new Lookups API lets you add large dictionaries and lookup tables to the Vocab and access them from the tokenizer or custom components and extension attributes. Internally, the tables use Bloom filters for efficient lookup checks. They're also fully serializable out-of-the-box. All large data resources included with spaCy now use this API and are additionally compressed at build time. This allowed us to make the installed library roughly 15 times smaller on disk.

API: Lookups **Usage: ** Adding languages: Lemmatizer

Text classification scores and CLI training

Example

$ python -m spacy train en /output /train /dev \\
--pipeline textcat --textcat-arch simple_cnn \\
--textcat-multilabel

When training your models using the spacy train command, you can now also include text categories in the JSON-formatted training data. The Scorer and nlp.evaluate now report the text classification scores, calculated as the F-score on positive label for binary exclusive tasks, the macro-averaged F-score for 3+ exclusive labels or the macro-averaged AUC ROC score for multilabel classification.

API: spacy train, Scorer, Language.evaluate

New DocBin class to efficiently serialize Doc collections

Example

from spacy.tokens import DocBin
doc_bin = DocBin(attrs=["LEMMA", "ENT_IOB", "ENT_TYPE"], store_user_data=True)
for doc in nlp.pipe(texts):
    doc_bin.add(doc)
bytes_data = doc_bin.to_bytes()
# Deserialize later, e.g. in a new process
nlp = spacy.blank("en")
doc_bin = DocBin().from_bytes(bytes_data)
docs = list(doc_bin.get_docs(nlp.vocab))

If you're working with lots of data, you'll probably need to pass analyses between machines, either to use something like Dask or Spark, or even just to save out work to disk. Often it's sufficient to use the Doc.to_array functionality for this, and just serialize the numpy arrays – but other times you want a more general way to save and restore Doc objects.

The new DocBin class makes it easy to serialize and deserialize a collection of Doc objects together, and is much more efficient than calling Doc.to_bytes on each individual Doc object. You can also control what data gets saved, and you can merge pallets together for easy map/reduce-style processing.

API: DocBin **Usage: ** Serializing Doc objects

CLI command to debug and validate training data

Example

$ python -m spacy debug-data en train.json dev.json

The new debug-data command lets you analyze and validate your training and development data, get useful stats, and find problems like invalid entity annotations, cyclic dependencies, low data labels and more. If you're training a model with spacy train and the results seem surprising or confusing, debug-data may help you track down the problems and improve your training data.

=========================== Data format validation ===========================
✔ Corpus is loadable

=============================== Training stats ===============================
Training pipeline: tagger, parser, ner
Starting with blank model 'en'
18127 training docs
2939 evaluation docs
⚠ 34 training examples also in evaluation data

============================== Vocab & Vectors ==============================
ℹ 2083156 total words in the data (56962 unique)
⚠ 13020 misaligned tokens in the training data
⚠ 2423 misaligned tokens in the dev data
10 most common words: 'the' (98429), ',' (91756), '.' (87073), 'to' (50058),
'of' (49559), 'and' (44416), 'a' (34010), 'in' (31424), 'that' (22792), 'is'
(18952)
ℹ No word vectors present in the model

========================== Named Entity Recognition ==========================
ℹ 18 new labels, 0 existing labels
528978 missing values (tokens with '-' label)
New: 'ORG' (23860), 'PERSON' (21395), 'GPE' (21193), 'DATE' (18080), 'CARDINAL'
(10490), 'NORP' (9033), 'MONEY' (5164), 'PERCENT' (3761), 'ORDINAL' (2122),
'LOC' (2113), 'TIME' (1616), 'WORK_OF_ART' (1229), 'QUANTITY' (1150), 'FAC'
(1134), 'EVENT' (974), 'PRODUCT' (935), 'LAW' (444), 'LANGUAGE' (338)
✔ Good amount of examples for all labels
✔ Examples without occurences available for all labels
✔ No entities consisting of or starting/ending with whitespace

=========================== Part-of-speech Tagging ===========================
ℹ 49 labels in data (57 labels in tag map)
'NN' (266331), 'IN' (227365), 'DT' (185600), 'NNP' (164404), 'JJ' (119830),
'NNS' (110957), '.' (101482), ',' (92476), 'RB' (90090), 'PRP' (90081), 'VB'
(74538), 'VBD' (68199), 'CC' (62862), 'VBZ' (50712), 'VBP' (43420), 'VBN'
(42193), 'CD' (40326), 'VBG' (34764), 'TO' (31085), 'MD' (25863), 'PRP$'
(23335), 'HYPH' (13833), 'POS' (13427), 'UH' (13322), 'WP' (10423), 'WDT'
(9850), 'RP' (8230), 'WRB' (8201), ':' (8168), '''' (7392), '``' (6984), 'NNPS'
(5817), 'JJR' (5689), '$' (3710), 'EX' (3465), 'JJS' (3118), 'RBR' (2872),
'-RRB-' (2825), '-LRB-' (2788), 'PDT' (2078), 'XX' (1316), 'RBS' (1142), 'FW'
(794), 'NFP' (557), 'SYM' (440), 'WP$' (294), 'LS' (293), 'ADD' (191), 'AFX'
(24)
✔ All labels present in tag map for language 'en'

============================= Dependency Parsing =============================
ℹ Found 111703 sentences with an average length of 18.6 words.
ℹ Found 2251 nonprojective train sentences
ℹ Found 303 nonprojective dev sentences
ℹ 47 labels in train data
ℹ 211 labels in projectivized train data
'punct' (236796), 'prep' (188853), 'pobj' (182533), 'det' (172674), 'nsubj'
(169481), 'compound' (116142), 'ROOT' (111697), 'amod' (107945), 'dobj' (93540),
'aux' (86802), 'advmod' (86197), 'cc' (62679), 'conj' (59575), 'poss' (36449),
'ccomp' (36343), 'advcl' (29017), 'mark' (27990), 'nummod' (24582), 'relcl'
(21359), 'xcomp' (21081), 'attr' (18347), 'npadvmod' (17740), 'acomp' (17204),
'auxpass' (15639), 'appos' (15368), 'neg' (15266), 'nsubjpass' (13922), 'case'
(13408), 'acl' (12574), 'pcomp' (10340), 'nmod' (9736), 'intj' (9285), 'prt'
(8196), 'quantmod' (7403), 'dep' (4300), 'dative' (4091), 'agent' (3908), 'expl'
(3456), 'parataxis' (3099), 'oprd' (2326), 'predet' (1946), 'csubj' (1494),
'subtok' (1147), 'preconj' (692), 'meta' (469), 'csubjpass' (64), 'iobj' (1)
⚠ Low number of examples for label 'iobj' (1)
⚠ Low number of examples for 130 labels in the projectivized dependency
trees used for training. You may want to projectivize labels such as punct
before training in order to improve parser performance.
⚠ Projectivized labels with low numbers of examples: appos||attr: 12
advmod||dobj: 13 prep||ccomp: 12 nsubjpass||ccomp: 15 pcomp||prep: 14
amod||dobj: 9 attr||xcomp: 14 nmod||nsubj: 17 prep||advcl: 2 prep||prep: 5
nsubj||conj: 12 advcl||advmod: 18 ccomp||advmod: 11 ccomp||pcomp: 5 acl||pobj:
10 npadvmod||acomp: 7 dobj||pcomp: 14 nsubjpass||pcomp: 1 nmod||pobj: 8
amod||attr: 6 nmod||dobj: 12 aux||conj: 1 neg||conj: 1 dative||xcomp: 11
pobj||dative: 3 xcomp||acomp: 19 advcl||pobj: 2 nsubj||advcl: 2 csubj||ccomp: 1
advcl||acl: 1 relcl||nmod: 2 dobj||advcl: 10 advmod||advcl: 3 nmod||nsubjpass: 6
amod||pobj: 5 cc||neg: 1 attr||ccomp: 16 advcl||xcomp: 3 nmod||attr: 4
advcl||nsubjpass: 5 advcl||ccomp: 4 ccomp||conj: 1 punct||acl: 1 meta||acl: 1
parataxis||acl: 1 prep||acl: 1 amod||nsubj: 7 ccomp||ccomp: 3 acomp||xcomp: 5
dobj||acl: 5 prep||oprd: 6 advmod||acl: 2 dative||advcl: 1 pobj||agent: 5
xcomp||amod: 1 dep||advcl: 1 prep||amod: 8 relcl||compound: 1 advcl||csubj: 3
npadvmod||conj: 2 npadvmod||xcomp: 4 advmod||nsubj: 3 ccomp||amod: 7
advcl||conj: 1 nmod||conj: 2 advmod||nsubjpass: 2 dep||xcomp: 2 appos||ccomp: 1
advmod||dep: 1 advmod||advmod: 5 aux||xcomp: 8 dep||advmod: 1 dative||ccomp: 2
prep||dep: 1 conj||conj: 1 dep||ccomp: 4 cc||ROOT: 1 prep||ROOT: 1 nsubj||pcomp:
3 advmod||prep: 2 relcl||dative: 1 acl||conj: 1 advcl||attr: 4 prep||npadvmod: 1
nsubjpass||xcomp: 1 neg||advmod: 1 xcomp||oprd: 1 advcl||advcl: 1 dobj||dep: 3
nsubjpass||parataxis: 1 attr||pcomp: 1 ccomp||parataxis: 1 advmod||attr: 1
nmod||oprd: 1 appos||nmod: 2 advmod||relcl: 1 appos||npadvmod: 1 appos||conj: 1
prep||expl: 1 nsubjpass||conj: 1 punct||pobj: 1 cc||pobj: 1 conj||pobj: 1
punct||conj: 1 ccomp||dep: 1 oprd||xcomp: 3 ccomp||xcomp: 1 ccomp||nsubj: 1
nmod||dep: 1 xcomp||ccomp: 1 acomp||advcl: 1 intj||advmod: 1 advmod||acomp: 2
relcl||oprd: 1 advmod||prt: 1 advmod||pobj: 1 appos||nummod: 1 relcl||npadvmod:
3 mark||advcl: 1 aux||ccomp: 1 amod||nsubjpass: 1 npadvmod||advmod: 1 conj||dep:
1 nummod||pobj: 1 amod||npadvmod: 1 intj||pobj: 1 nummod||npadvmod: 1
xcomp||xcomp: 1 aux||dep: 1 advcl||relcl: 1
⚠ The following labels were found only in the train data: xcomp||amod,
advcl||relcl, prep||nsubjpass, acl||nsubj, nsubjpass||conj, xcomp||oprd,
advmod||conj, advmod||advmod, iobj, advmod||nsubjpass, dobj||conj, ccomp||amod,
meta||acl, xcomp||xcomp, prep||attr, prep||ccomp, advcl||acomp, acl||dobj,
advcl||advcl, pobj||agent, prep||advcl, nsubjpass||xcomp, prep||dep,
acomp||xcomp, aux||ccomp, ccomp||dep, conj||dep, relcl||compound,
nsubjpass||ccomp, nmod||dobj, advmod||advcl, advmod||acl, dobj||advcl,
dative||xcomp, prep||nsubj, ccomp||ccomp, nsubj||ccomp, xcomp||acomp,
prep||acomp, dep||advmod, acl||pobj, appos||dobj, npadvmod||acomp, cc||ROOT,
relcl||nsubj, nmod||pobj, acl||nsubjpass, ccomp||advmod, pcomp||prep,
amod||dobj, advmod||attr, advcl||csubj, appos||attr, dobj||pcomp, prep||ROOT,
relcl||pobj, advmod||pobj, amod||nsubj, ccomp||xcomp, prep||oprd,
npadvmod||advmod, appos||nummod, advcl||pobj, neg||advmod, acl||attr,
appos||nsubjpass, csubj||ccomp, amod||nsubjpass, intj||pobj, dep||advcl,
cc||neg, xcomp||ccomp, dative||ccomp, nmod||oprd, pobj||dative, prep||dobj,
dep||ccomp, relcl||attr, ccomp||nsubj, advcl||xcomp, nmod||dep, advcl||advmod,
ccomp||conj, pobj||prep, advmod||acomp, advmod||relcl, attr||pcomp,
ccomp||parataxis, oprd||xcomp, intj||advmod, nmod||nsubjpass, prep||npadvmod,
parataxis||acl, prep||pobj, advcl||dobj, amod||pobj, prep||acl, conj||pobj,
advmod||dep, punct||pobj, ccomp||acomp, acomp||advcl, nummod||npadvmod,
dobj||dep, npadvmod||xcomp, advcl||conj, relcl||npadvmod, punct||acl,
relcl||dobj, dobj||xcomp, nsubjpass||parataxis, dative||advcl, relcl||nmod,
advcl||ccomp, appos||npadvmod, ccomp||pcomp, prep||amod, mark||advcl,
prep||advmod, prep||xcomp, appos||nsubj, attr||ccomp, advmod||prt, dobj||ccomp,
aux||conj, advcl||nsubj, conj||conj, advmod||ccomp, advcl||nsubjpass,
attr||xcomp, nmod||conj, npadvmod||conj, relcl||dative, prep||expl,
nsubjpass||pcomp, advmod||xcomp, advmod||dobj, appos||pobj, nsubj||conj,
relcl||nsubjpass, advcl||attr, appos||ccomp, advmod||prep, prep||conj,
nmod||attr, punct||conj, neg||conj, dep||xcomp, aux||xcomp, dobj||acl,
nummod||pobj, amod||npadvmod, nsubj||pcomp, advcl||acl, appos||nmod,
relcl||oprd, prep||prep, cc||pobj, nmod||nsubj, amod||attr, aux||dep,
appos||conj, advmod||nsubj, nsubj||advcl, acl||conj
To train a parser, your data should include at least 20 instances of each label.
⚠ Multiple root labels (ROOT, nsubj, aux, npadvmod, prep) found in
training data. spaCy's parser uses a single root label ROOT so this distinction
will not be available.

================================== Summary ==================================
✔ 5 checks passed
⚠ 8 warnings

API: spacy debug-data

Backwards incompatibilities

If you've been training your own models, you'll need to retrain them with the new version. Also don't forget to upgrade all models to the latest versions. Models for v2.0 or v2.1 aren't compatible with models for v2.2. To check if all of your models are up to date, you can run the spacy validate command.

• The Dutch model has been trained on a new NER corpus (custom labelled UD instead of WikiNER), so their predictions may be very different compared to the previous version. The results should be significantly better and more generalizable, though.
• The spacy download command does not set the --no-deps pip argument anymore by default, meaning that model package dependencies (if available) will now be also downloaded and installed. If spaCy (which is also a model dependency) is not installed in the current environment, e.g. if a user has built from source, --no-deps is added back automatically to prevent spaCy from being downloaded and installed again from pip.
• The built-in biluo_tags_from_offsets converter is now stricter and will raise an error if entities are overlapping (instead of silently skipping them). If your data contains invalid entity annotations, make sure to clean it and resolve conflicts. You can now also use the new debug-data command to find problems in your data.
• Pipeline components can now overwrite IOB tags of tokens that are not yet part of an entity. Once a token has an ent_iob value set, it won't be reset to an "unset" state and will always have at least O assigned. list(doc.ents) now actually keeps the annotations on the token level consistent, instead of resetting O to an empty string.
• The default punctuation in the Sentencizer has been extended and now includes more characters common in various languages. This also means that the results it produces may change, depending on your text. If you want the previous behaviour with limited characters, set punct_chars=[".", "!", "?"] on initialization.
• The PhraseMatcher algorithm was rewritten from scratch and it's now 10× faster. The rewrite also resolved a few subtle bugs with very large terminology lists. So if you were matching large lists, you may see slightly different results – however, the results should now be fully correct. See this PR for more details.
• Lemmatization tables (rules, exceptions, index and lookups) are now part of the Vocab and serialized with it. This means that serialized objects (nlp, pipeline components, vocab) will now include additional data, and models written to disk will include additional files.
• The Serbian language class (introduced in v2.1.8) incorrectly used the language code rs instead of sr. This has now been fixed, so Serbian is now available via spacy.lang.sr.
• The "sources" in the meta.json have changed from a list of strings to a list of dicts. This is mostly internals, but if your code used nlp.meta["sources"], you might have to update it.