Skip to content

v2.1.0: New models, ULMFit/BERT/Elmo-like pretraining, faster tokenization, better Matcher, bug fixes & more

Choose a tag to compare
@ines ines released this 18 Mar 15:07
· 6262 commits to master since this release

鈿狅笍 This version of spaCy requires downloading new models. You can use the spacy validate command to find out which models need updating, and print update instructions. If you've been training your own models, you'll need to retrain them with the new version.

鉁 New features and improvements

Tagger, Parser, NER and Text Categorizer

  • NEW: Experimental ULMFit/BERT/Elmo-like pretraining (see #2931) via the new spacy pretrain command. This pre-trains the CNN using BERT's cloze task. A new trick we're calling Language Modelling with Approximate Outputs is used to apply the pre-training to smaller models. The pre-training outputs CNN and embedding weights that can be used in spacy train, using the new -t2v argument.
  • NEW: Allow parser to do joint word segmentation and parsing. If you pass in data where the tokenizer over-segments, the parser now learns to merge the tokens.
  • Make parser, tagger and NER faster, through better hyperparameters.
  • Add simpler, GPU-friendly option to TextCategorizer, and allow setting exclusive_classes and architecture arguments on initialization.
  • Add EntityRecognizer.labels property.
  • Remove document length limit during training, by implementing faster Levenshtein alignment.
  • Use Thinc v7.0, which defaults to single-thread with fast blis kernel for matrix multiplication. Parallelisation should be performed at the task level, e.g. by running more containers.

Models & Language Data

  • NEW: 2-3 times faster tokenization across all languages at the same accuracy!
  • NEW: Small accuracy improvements for parsing, tagging and NER for 6+ languages.
  • NEW: The English and German models are now available under the MIT license.
  • NEW: Statistical models for Greek.
  • NEW: Alpha support for Tamil, Ukrainian and Kannada, and base language classes for Afrikaans, Bulgarian, Czech, Icelandic, Lithuanian, Latvian, Slovak, Slovenian and Albanian.
  • Improve loading time of French by ~30%.
  • Add Vocab.writing_system (populated via the language data) to expose settings like writing direction.


  • NEW: pretrain command for ULMFit/BERT/Elmo-like pretraining (see #2931).
  • NEW: New ud-train command, to train and evaluate using the CoNLL 2017 shared task data.
  • Check if model is already installed before downloading it via spacy download.
  • Pass additional arguments of download command to pip to customise installation.
  • Improve train command by letting GoldCorpus stream data, instead of loading into memory.
  • Improve init-model command, including support for lexical attributes and word-vectors, using a variety of formats. This replaces the spacy vocab command, which is now deprecated.
  • Add support for multi-task objectives to train command.
  • Add support for data-augmentation to train command.


  • NEW: Enhanced pattern API for rule-based Matcher (see #1971).
  • NEW: Doc.retokenize context manager for merging and splitting tokens more efficiently.
  • NEW: Add support for custom pipeline component factories via entry points (#2348).
  • NEW: Implement fastText vectors with subword features.
  • NEW: Built-in rule-based NER component to add entities based on match patterns (see #2513).
  • NEW: Allow PhraseMatcher to match on token attributes other than ORTH, e.g. LOWER (for case-insensitive matching) or even POS or TAG.
  • NEW: Replace ujson, msgpack, msgpack-numpy, pickle, cloudpickle and dill with our own package srsly to centralise dependencies and allow binary wheels.
  • NEW: Doc.to_json() method which outputs data in spaCy's training format. This will be the only place where the format is hard-coded (see #2932).
  • NEW: Built-in EntityRuler component to make it easier to build rule-based NER and combinations of statistical and rule-based systems.
  • NEW: gold.spans_from_biluo_tags helper that returns Span objects, e.g. to overwrite the doc.ents.
  • Add warnings if .similarity method is called with empty vectors or without word vectors.
  • Improve rule-based Matcher and add return_matches keyword argument to Matcher.pipe to yield (doc, matches) tuples instead of only Doc objects, and as_tuples to add context to the Doc objects.
  • Make stop words via Token.is_stop and Lexeme.is_stop case-insensitive.
  • Accept "TEXT" as an alternative to "ORTH" in Matcher patterns.
  • Use black for auto-formatting .py source and optimse codebase using flake8. You can now run flake8 spacy and it should return no errors or warnings. See for details.

馃敶 Bug fixes

  • Fix issue #795: Fix behaviour of Token.conjuncts.
  • Fix issue #1487: Add Doc.retokenize() context manager.
  • Fix issue #1537: Make Span.as_doc return a copy, not a view.
  • Fix issue #1574: Make sure stop words are available in medium and large English models.
  • Fix issue #1585: Prevent parser from predicting unseen classes.
  • Fix issue #1642: Replace regex with re and speed up tokenization.
  • Fix issue #1665: Correct typos in symbol Animacy_inan and add Animacy_nhum.
  • Fix issue #1748, #1798, #2756, #2934: Add simpler GPU-friendly option to TextCategorizer.
  • Fix issue #1773: Prevent tokenizer exceptions from setting POS but not TAG.
  • Fix issue #1782, #2343: Fix training on GPU.
  • Fix issue #1816: Allow custom Language subclasses via entry points.
  • Fix issue #1865: Correct licensing of it_core_news_sm model.
  • Fix issue #1889: Make stop words case-insensitive.
  • Fix issue #1903: Add relcl dependency label to symbols.
  • Fix issue #1963: Resize Doc.tensor when merging spans.
  • Fix issue #1971: Update Matcher engine to support regex, extension attributes and rich comparison.
  • Fix issue #2014: Make Token.pos_ writeable.
  • Fix issue #2091: Fix displacy support for RTL languages.
  • Fix issue #2203, #3268: Prevent bad interaction of lemmatizer and tokenizer exceptions.
  • Fix issue #2329: Correct TextCategorizer and GoldParse API docs.
  • Fix issue #2369: Respect pre-defined warning filters.
  • Fix issue #2390: Support setting lexical attributes during retokenization.
  • Fix issue #2396: Fix Doc.get_lca_matrix.
  • Fix issue #2464, #3009: Fix behaviour of Matcher's ? quantifier.
  • Fix issue #2482: Fix serialization when parser model is empty.
  • Fix issue #2512, #2153: Fix issue with deserialization into non-empty vocab.
  • Fix issue #2603: Improve handling of missing NER tags.
  • Fix issue #2644: Add table explaining training metrics to docs.
  • Fix issue #2648: Fix KeyError in Vectors.most_similar.
  • Fix issue #2671, #2675: Fix incorrect match ID on some patterns.
  • Fix issue #2693: Only use 'sentencizer' as built-in sentence boundary component name.
  • Fix issue #2728: Fix HTML escaping in displacy NER visualization and correct API docs.
  • Fix issue #2740: Add ability to pass additional arguments to pipeline components.
  • Fix issue #2754, #3028: Make NORM a Token attribute instead of a Lexeme attribute to allow setting context-specific norms in tokenizer exceptions.
  • Fix issue #2769: Fix issue that'd cause segmentation fault when calling EntityRecognizer.add_label.
  • Fix issue #2772: Fix bug in sentence starts for non-projective parses.
  • Fix issue #2779: Fix handling of pre-set entities.
  • Fix issue #2782: Make like_num work with prefixed numbers.
  • Fix issue #2833: Raise better error if Token or Span are pickled.
  • Fix issue #2838: Add Retokenizer.split method to split one token into several.
  • Fix issue #2869: Make doc[0].is_sent_start == True.
  • Fix issue #2870: Make it illegal for the entity recognizer to predict whitespace tokens as B, L or U.
  • Fix issue #2871: Fix vectors for reserved words.
  • Fix issue #2901: Fix issue with first call of nlp in Japanese (MeCab).
  • Fix issue #2924: Make IDs of displaCy arcs more unique to avoid clashes.
  • Fix issue #3012: Fix clobber of Doc.is_tagged in Doc.from_array.
  • Fix issue #3027: Allow Span to take unicode value for label argument.
  • Fix issue #3036: Support mutable default arguments in extension attributes.
  • Fix issue #3048: Raise better errors for uninitialized pipeline components.
  • Fix issue #3064: Allow single string attributes in Doc.to_array.
  • Fix issue #3093, #3067: Set correctly when exporting model via CLI.
  • Fix issue #3112: Make sure entity types are added correctly on GPU.
  • Fix issue #3191: Fix pickling of Japanese.
  • Fix issue #3122: Correct docs of Token.subtree and Span.subtree.
  • Fix issue #3128: Improve error handling in converters.
  • Fix issue #3248: Fix PhraseMatcher pickling and make __len__ consistent.
  • Fix issue #3274: Make Token.sent work as expected without the parser.
  • Fix issue #3277: Add en/em dash to tokenizer prefixes and suffixes.
  • Fix issue #3346: Expose Japanese stop words in language class.
  • Fix issue #3357: Update displaCy examples in docs to correctly show Token.pos_.
  • Fix issue #3345: Fix NER when preset entities cross-sentence boundaries.
  • Fix issue #3348: Don't use numpy directly for similarity.
  • Fix issue #3366: Improve converters, training data formats and docs.
  • Fix issue #3369: Fix #egg fragments in direct downloads.
  • Fix issue #3382: Make Doc.from_array consistent with Doc.to_array.
  • Fix issue #3398: Don't set extension attributes in language classes.
  • Fix issue #3373: Merge and improve conllu converters.
  • Fix serialization of custom tokenizer if not all functions are defined.
  • Fix bugs in beam-search training objective.
  • Fix problems with model pickling.

鈿狅笍 Backwards incompatibilities

  • This version of spaCy requires downloading new models. You can use the spacy validate command to find out which models need updating, and print update instructions.
  • If you've been training your own models, you'll need to retrain them with the new version.
  • Due to difficulties linking our new blis for faster platform-independent matrix multiplication, v2.1.x currently doesn't work on Python 2.7 on Windows. We expect this to be corrected in the future.
  • While the Matcher API is fully backwards compatible, its algorithm has changed to fix a number of bugs and performance issues. This means that the Matcher in v2.1.x may produce different results compared to the Matcher in v2.0.x.
  • The deprecated Doc.merge and Span.merge methods still work, but you may notice that they now run slower when merging many objects in a row. That's because the merging engine was rewritten to be more reliable and to support more efficient merging in bulk. To take advantage of this, you should rewrite your logic to use the Doc.retokenize context manager and perform as many merges as possible together in the with block.
- doc[1:5].merge()
- doc[6:8].merge()
+ with doc.retokenize() as retokenizer:
+     retokenizer.merge(doc[1:5])
+     retokenizer.merge(doc[6:8])
  • The serialization methods to_disk, from_disk, to_bytes and from_bytes now support a single exclude argument to provide a list of string names to exclude. The docs have been updated to list the available serialization fields for each class. The disable argument on the Language serialization methods has been renamed to exclude for consistency.
- nlp.to_disk("/path", disable=["parser", "ner"])
+ nlp.to_disk("/path", exclude=["parser", "ner"])
- data = nlp.tokenizer.to_bytes(vocab=False)
+ data = nlp.tokenizer.to_bytes(exclude=["vocab"])
  • The .pos value for several common English words has changed, due to corrections to long-standing mistakes in the English tag map (see #593, #3311).
  • For better compatibility with the Universal Dependencies data, the lemmatizer now preserves capitalization, e.g. for proper nouns (see #3256).
  • The keyword argument n_threads on the .pipe methods is now deprecated, as the v2.x models cannot release the global interpreter lock. (Future versions may introduce a n_process argument for parallel inference via multiprocessing.)
  • The Doc.print_tree method is not deprecated in favour of a unified Doc.to_json method, which outputs data in the same format as the expected JSON training data.
  • The built-in rule-based sentence boundary detector is now only called 'sentencizer' 鈥 the name 'sbd' is deprecated.
- sentence_splitter = nlp.create_pipe('sbd')
+ sentence_splitter = nlp.create_pipe('sentencizer')
  • The is_sent_start attribute of the first token in a Doc now correctly defaults to True. It previously defaulted to None.
  • The spacy train command now lets you specify a comma-separated list of pipeline component names, instead of separate flags like --no-parser to disable components. This is more flexible and also handles custom components out-of-the-box.
- $ spacy train en /output train_data.json dev_data.json --no-parser
+ $ spacy train en /output train_data.json dev_data.json --pipeline tagger,ner
  • The spacy init-model command now uses a --jsonl-loc argument to pass in a a newline-delimited JSON (JSONL) file containing one lexical entry per line instead of a separate --freqs-loc and --clusters-loc.
- $ spacy init-model en ./model --freqs-loc ./freqs.txt --clusters-loc ./clusters.txt
+ $ spacy init-model en ./model --jsonl-loc ./vocab.jsonl
  • Also note that some of the model licenses have changed: it_core_news_sm is now correctly licensed under CC BY-NC-SA 3.0, and all English and German models are now published under the MIT license.

馃搱 Benchmarks

Model Language Version UAS LAS POS NER F Vec Size
en_core_web_sm English 2.1.0 91.5 89.7 96.8 85.9 饜剛 10 MB
en_core_web_md English 2.1.0 91.8 90.0 96.9 86.6 90 MB
en_core_web_lg English 2.1.0 91.8 90.1 97.0 86.6 788 MB
de_core_news_sm German 2.1.0 90.7 88.6 96.3 83.1 饜剛 10 MB
de_core_news_md German 2.1.0 91.2 89.4 96.6 83.8 210 MB
es_core_news_sm Spanish 2.1.0 90.4 87.3 96.9 89.5 饜剛 10 MB
es_core_news_md Spanish 2.1.0 91.0 88.2 97.2 89.7 69 MB
pt_core_news_sm Portuguese 2.1.0 89.1 85.9 80.4 88.9 饜剛 12 MB
fr_core_news_sm French 2.1.0 87.6 84.7 94.5 82.6 饜剛 14 MB
fr_core_news_md French 2.1.0 89.1 86.4 95.3 83.1 82 MB
it_core_news_sm Italian 2.1.0 91.0 87.3 95.8 86.1 饜剛 10 MB
nl_core_news_sm Dutch 2.1.0 83.7 77.6 91.6 87.0 饜剛 10 MB
el_core_news_sm Greek 2.1.0 84.4 80.6 94.6 71.6 饜剛 10 MB
el_core_news_md Greek 2.1.0 88.3 85.0 96.6 81.1 126 MB
xx_ent_wiki_sm Multi 2.1.0 - - - 81.3 饜剛 3 MB

馃挰 UAS: Unlabelled dependencies (parser). LAS: Labelled dependencies (parser). POS: Part-of-speech tags (fine-grained tags, i.e. Token.tag_). NER F: Named entities (F-score). Vec: Model contains word vectors. Size: Model file size (zipped archive).

馃摉 Documentation and examples

Although it looks pretty much the same, we've rebuilt the entire documentation using Gatsby and MDX. It's now an even faster progressive web app and allows us to write all content entirely in Markdown, without having to compromise on easy-to-use custom UI components. We're hoping that the Markdown source will make it even easier to contribute to the documentation. For more details, check out the styleguide and source.

While converting the pages to Markdown, we've also fixed a bunch of typos, improved the existing pages and added some new content:

  • Usage Guide: Rule-based Matching. How to use the Matcher, PhraseMatcher and the new EntityRuler, and write powerful components to combine statistical models and rules.
  • Usage Guide: Saving and Loading. Everything you need to know about serialization, and how to save and load pipeline components, package your spaCy models as Python modules and use entry points.
  • Usage Guide: Merging and Splitting. How to retokenize a Doc using the new retokenize context manager and merge spans into single tokens and split single tokens into multiple.
  • Universe: Videos and Podcasts
  • API: EntityRuler
  • API: SentenceSegmenter
  • API: Pipeline functions

馃懃 Contributors

Thanks to @DuyguA, @giannisdaras, @mgogoulos, @louridas, @skrcode, @gavrieltal, @svlandeg, @jarib, @alvaroabascar, @kbulygin, @moreymat, @mirfan899, @ozcankasal, @willprice, @alvations, @amperinet, @retnuh, @loghijiaha, @DeNeutoy, @gavrieltal, @boena, @BramVanroy, @pganssle, @foufaster, @adrianeboyd, @maknotavailable, @pierremonico, @lauraBaakman, @juliamakogon, @Gizzio, @Abhijit-2592, @akki2825, @grivaz, @roshni-b, @mpuig, @mikelibg, @danielkingai2, @adrienball and @Poluglottos for the pull requests and contributions.