explosion/spaCy: v3.3.0: Improved speed, new trainable lemmatizer, and pipelines for Finnish, Korean and Swedish

✨ New features and improvements

• Improved speeds for many components, see speed benchmarks for trained pipelines:
  • Speed up parser and NER by using constant-time head lookups (#10048).
  • Support unnormalized softmax probabilities in spacy.Tagger.v2 to speed up inference for the tagger, morphologizer, senter and trainable lemmatizer (#10197).
  • Speed up parser projectivization functions (#10241).
  • Replace Ragged with faster AlignmentArray in Example for training (#10319).
  • Improve Matcher speed (#10659).
  • Improve serialization speed for empty Doc.spans (#10250).
• NEW: A trainable lemmatizer component that uses edit trees to transform tokens to lemmas. Add it to your config with spacy init config -p trainable_lemmatizer or using the quickstart.
• Language updates:
  • Initial support for Lower Sorbian and Upper Sorbian.
  • New noun chunks for Finnish.
  • Updated noun chunks for French, Italian and Spanish.
  • Additional updates for English, French, Italian, Japanese, Korean, Norwegian, Russian, Slovenian, Spanish, Turkish, Ukrainian and Vietnamese.
• Big endian support with thinc v8.0.14+ and thinc-bigendian-ops.
• Config comparisons with spacy debug diff-config.
• displaCy support for overlapping span annotation and multiple labeled arcs between the same tokens.
• SpanCategorizer.set_candidates for debugging span suggesters.
• The quickstart now supports adding spancat and trainable_lemmatizer components.

📦 Trained pipelines

v3.3 introduces trained pipelines for Finnish, Korean and Swedish which feature the trainable lemmatizer and floret vectors. Due to the use Bloom embeddings and subwords, the pipelines have compact vectors with no out-of-vocabulary words.

Package	Language	UPOS	Parser LAS	NER F
fi_core_news_sm	Finnish	92.5	71.9	75.9
fi_core_news_md	Finnish	95.9	78.6	80.6
fi_core_news_lg	Finnish	96.2	79.4	82.4
ko_core_news_sm	Korean	86.1	65.6	71.3
ko_core_news_md	Korean	94.7	80.9	83.1
ko_core_news_lg	Korean	94.7	81.3	85.3
sv_core_news_sm	Swedish	95.0	75.9	74.7
sv_core_news_md	Swedish	96.3	78.5	79.3
sv_core_news_lg	Swedish	96.3	79.1	81.1

🙏 Special thanks to @aajanki, @thiippal (Finnish) and Elena Fano (Swedish) for their help with the new pipelines!

The new trainable lemmatizer is used for Danish, Dutch, Finnish, German, Greek, Italian, Korean, Lithuanian, Norwegian, Polish, Portuguese, Romanian and Swedish.

Model	v3.2 Lemma Acc	v3.3 Lemma Acc
da_core_news_md	84.9	94.8
de_core_news_md	73.4	97.7
el_core_news_md	56.5	88.9
fi_core_news_md	-	86.2
it_core_news_md	86.6	97.2
ko_core_news_md	-	90.0
lt_core_news_md	71.1	84.8
nb_core_news_md	76.7	97.1
nl_core_news_md	81.5	94.0
pl_core_news_md	87.1	93.7
pt_core_news_md	76.7	96.9
ro_core_news_md	81.8	95.5
sv_core_news_md	-	95.5

🔴 Bug fixes

• Fix issue #5447: Avoid overlapping arcs when using displaCy in manual mode.
• Fix issue #9443: Fix Scorer.score_cats for missing labels.
• Fix issue #9669: Fix entity linker batching.
• Fix issue #9903: Handle _ value for UPOS in CoNLL-U converter.
• Fix issue #9904: Fix textcat loss scaling.
• Fix issue #9956: Compare all Span attributes consistently.
• Fix issue #10073: Add "spans" to the output of doc.to_json.
• Fix issue #10086: Add tokenizer option to allow Matcher handling for all special cases.
• Fix issue #10189: Allow Example to align whitespace annotation.
• Fix issue #10302: Fix check for NER annotation in MISC in CoNLL-U converter.
• Fix issue #10324: Fix Tok2Vec for empty batches.
• Fix issue #10347: Update basic functionality for rehearse.
• Fix issue #10394: Fix Vectors.n_keys for floret vectors.
• Fix issue #10400: Use meta in util.load_model_from_config.
• Fix issue #10451: Fix Example.get_matching_ents.
• Fix issue #10460: Fix initial special cases for Tokenizer.explain.
• Fix issue #10521: Stream large assets on download in spaCy projects.
• Fix issue #10536: Handle unknown tags in KoreanTokenizer tag map.
• Fix issue #10551: Add automatic vector deduplication for init vectors.

🚀 Notes about upgrading from v3.2

• To see the speed improvements for the Tagger architecture, edit your configs to switch from spacy.Tagger.v1 to spacy.Tagger.v2 and then run init fill-config.
• Span comparisons involving ordering (<, <=, >, >=) now take all span attributes into account (start, end, label, and KB ID) so spans may be sorted in a slightly different order (#9956).
• Annotation on whitespace tokens is handled in the same way as annotation on non-whitespace tokens during training in order to allow custom whitespace annotation (#10189).
• Doc.from_docs now includes Doc.tensor by default and supports excludes with an exclude argument in the same format as Doc.to_bytes. The supported exclude fields are spans, tensor and user_data.

📖 Documentation and examples

• spaCy universe additions:
  • classy-classification: A Python library for classy few-shot and zero-shot classification within spaCy.
  • Concise Concepts: Concise Concepts uses few-shot NER based on word embedding similarity.
  • Crosslingual Coreference: Crosslingual coreference with an English coreference model plus crosslingual embeddings.
  • EDS-NLP: spaCy components to extract information from clinical notes written in French.
  • HuSpaCy: Industrial-strength Hungarian natural language processing.
  • Klayers: spaCy as a AWS Lambda Layer.
  • Named Entity Recognition (NER) using spaCy (video).
  • Scrubadub: Remove personally identifiable information from text using spaCy.
  • spacy-setfit-textcat: Experiments with SetFit & Few-Shot Classification.
  • tmtoolkit: Text mining and topic modeling toolkit.

👥 Contributors

@aajanki, @adrianeboyd, @apjanco, @bdura, @BramVanroy, @danieldk, @danmysak, @davidberenstein1957, @DuyguA, @fonfonx, @gremur, @HaakonME, @harmbuisman, @honnibal, @ines, @internaut, @jfainberg, @jnphilipp, @jsnfly, @kadarakos, @koaning, @ljvmiranda921, @martinjack, @mgrojo, @nrodnova, @ofirnk, @orglce, @pepemedigu, @philipvollet, @pmbaumgartner, @polm, @richardpaulhudson, @ryndaniels, @SamEdwardes, @Schero1994, @shadeMe, @single-fingal, @svlandeg, @thebugcreator, @thomashacker, @umaxfun, @y961996