What's New in v3.6

New features {id="features",hidden="true"}

spaCy v3.6 adds the new SpanFinder component to the core spaCy library and new trained pipelines for Slovenian.

SpanFinder {id="spanfinder"}

The SpanFinder component identifies potentially overlapping, unlabeled spans by identifying span start and end tokens. It is intended for use in combination with a component like SpanCategorizer that may further filter or label the spans. See our Spancat blog post for a more detailed introduction to the span finder.

To train a pipeline with span_finder + spancat, remember to add span_finder (and its tok2vec or transformer if required) to [training.annotating_components] so that the spancat component can be trained directly from its predictions:

[nlp]
pipeline = ["tok2vec","span_finder","spancat"]

[training]
annotating_components = ["tok2vec","span_finder"]

In practice it can be helpful to initially train the span_finder separately before sourcing it (along with its tok2vec) into the spancat pipeline for further training. Otherwise the memory usage can spike for spancat in the first few training steps if the span_finder makes a large number of predictions.

Additional features and improvements {id="additional-features-and-improvements"}

• Language updates:
  • Add initial support for Malay.
  • Update Latin defaults to support noun chunks, update lexical/tokenizer settings and add example sentences.
• Support spancat_singlelabel in spacy debug data CLI.
• Add doc.spans rendering to spacy evaluate CLI displaCy output.
• Support custom token/lexeme attribute for vectors.
• Add option to return scores separately keyed by component name with spacy evaluate --per-component, Language.evaluate(per_component=True) and Scorer.score(per_component=True). This is useful when the pipeline contains more than one of the same component like textcat that may have overlapping scores keys.
• Typing updates for PhraseMatcher and SpanGroup.

Trained pipelines {id="pipelines"}

New trained pipelines {id="new-pipelines"}

v3.6 introduces new pipelines for Slovenian, which use the trainable lemmatizer and floret vectors.

Pipeline updates {id="pipeline-updates"}

The English pipelines have been updated to improve handling of contractions with various apostrophes and to lemmatize "get" as a passive auxiliary.

The Danish pipeline da_core_news_trf has been updated to use vesteinn/DanskBERT with performance improvements across the board.

Notes about upgrading from v3.5 {id="upgrading"}

SpanGroup spans are now required to be from the same doc {id="spangroup-spans"}

When initializing a SpanGroup, there is a new check to verify that all added spans refer to the current doc. Without this check, it was possible to run into string store or other errors.

One place this may crop up is when creating Example objects for training with custom spans:

     doc = Doc(nlp.vocab, words=tokens)  # predicted doc
     example = Example.from_dict(doc, {"ner": iob_tags})
     # use the reference doc when creating reference spans
-    span = Span(doc, 0, 5, "ORG")
+    span = Span(example.reference, 0, 5, "ORG")
     example.reference.spans[spans_key] = [span]

Pipeline package version compatibility {id="version-compat"}

Using legacy implementations

In spaCy v3, you'll still be able to load and reference legacy implementations via spacy-legacy, even if the components or architectures change and newer versions are available in the core library.

When you're loading a pipeline package trained with an earlier version of spaCy v3, you will see a warning telling you that the pipeline may be incompatible. This doesn't necessarily have to be true, but we recommend running your pipelines against your test suite or evaluation data to make sure there are no unexpected results.

If you're using one of the trained pipelines we provide, you should run spacy download to update to the latest version. To see an overview of all installed packages and their compatibility, you can run spacy validate.

If you've trained your own custom pipeline and you've confirmed that it's still working as expected, you can update the spaCy version requirements in the meta.json:

- "spacy_version": ">=3.5.0,<3.6.0",
+ "spacy_version": ">=3.5.0,<3.7.0",

Updating v3.5 configs

To update a config from spaCy v3.5 with the new v3.6 settings, run init fill-config:

$ python -m spacy init fill-config config-v3.5.cfg config-v3.6.cfg

In many cases (spacy train, spacy.load), the new defaults will be filled in automatically, but you'll need to fill in the new settings to run debug config and debug data.