It’s been great to see the adoption of the new spaCy v3, which introduced transformer-based pipelines, a new config and training system for reproducible experiments, projects for end-to-end workflows, and many other features. Version 3.1 adds more on top of it, including the ability to use predicted annotations during training, a new SpanCategorizer component for predicting arbitrary and potentially overlapping spans, support for partial incorrect annotations in the entity recognizer, new trained pipelines for Catalan and Danish, as well as many bug fixes and improvements.

Using predicted annotations during training

By default, components are updated in isolation during training, which means that they don’t see the predictions of any earlier components in the pipeline. The new [training.annotating_components] config setting lets you specify pipeline components that should set annotations on the predicted docs during training. This makes it easy to use the predictions of a previous component in the pipeline as features for a subsequent component, e.g. the dependency labels in the tagger:

config.cfg (excerpt)

SpanCategorizer for predicting arbitrary and overlapping spans experimental

A common task in applied NLP is extracting spans of texts from documents, including longer phrases or nested expressions. Named entity recognition isn’t the right tool for this problem, since an entity recognizer typically predicts single token-based tags that are very sensitive to boundaries. This is effective for proper nouns and self-contained expressions, but less useful for other types of phrases or overlapping spans. The new SpanCategorizer component and SpanCategorizer architecture let you label arbitrary and potentially overlapping spans of texts. A span categorizer consists of two parts: a suggester function that proposes candidate spans, which may or may not overlap, and a labeler model that predicts zero or more labels for each candidate. The predicted spans are available via the Doc.spans container.

Update the entity recognizer with partial incorrect annotations

The EntityRecognizer can now be updated with known incorrect annotations, which lets you take advantage of partial and sparse data. For example, you’ll be able to use the information that certain spans of text are definitely not PERSON entities, without having to provide the complete gold-standard annotations for the given example. The incorrect span annotations can be added via the Doc.spans in the training data under the key defined as incorrect_spans_key in the component config.

New pipeline packages for Catalan and Danish

spaCy v3.1 adds 5 new pipeline packages, including a new core family for Catalan and a new transformer-based pipeline for Danish using the danish-bert-botxo weights. See the models directory for an overview of all available trained pipelines and the training guide for details on how to train your own.

Resizable text classification architectures

Previously, the TextCategorizer architectures could not be resized, meaning that you couldn’t add new labels to an already trained model. In spaCy v3.1, the TextCatCNN and TextCatBOW architectures are now resizable, while ensuring that the predictions for the old labels remain the same.

CLI command to assemble pipeline from config

The spacy assemble command lets you assemble a pipeline from a config file without additional training. It can be especially useful for creating a blank pipeline with a custom tokenizer, rule-based components or word vectors.

Pretty pipeline package READMEs

The spacy package command now auto-generates a pretty README.md based on the pipeline information defined in the meta.json. This includes a table with a general overview, as well as the label scheme and accuracy figures, if available. For an example, see the model releases.

Support for streaming large or infinite corpora

The training process now supports streaming large or infinite corpora out-of-the-box, which can be controlled via the [training.max_epochs] config setting. Setting it to -1 means that the train corpus should be streamed rather than loaded into memory with no shuffling within the training loop. For details on how to implement a custom corpus loader, e.g. to stream in data from a remote storage, see the usage guide on custom data reading.

When streaming a corpus, only the first 100 examples will be used for initialization. This is no problem if you’re training a component like the text classifier with data that specifies all available labels in every example. If necessary, you can use the init labels command to pre-generate the labels for your components using a representative sample so the model can be initialized correctly before training.

New lemmatizers for Catalan and Italian

The trained pipelines for Catalan and Italian now include lemmatizers that use the predicted part-of-speech tags as part of the lookup lemmatization for higher lemmatization accuracy. If you’re training your own pipelines for these languages and you want to include a lemmatizer, make sure you have the spacy-lookups-data package installed, which provides the relevant tables.

Upload your pipelines to the Hugging Face Hub

The Hugging Face Hub lets you upload models and share them with others, and it now supports spaCy pipelines out-of-the-box. The new spacy-huggingface-hub package automatically adds the huggingface-hub command to your spacy CLI. It lets you upload any pipelines packaged with spacy package and --build wheel and takes care of auto-generating all required meta information.

After uploading, you’ll get a live URL for your model page that includes all details, files and interactive visualizers, as well as a direct URL to the wheel file that you can install via pip install. For examples, check out the spaCy pipelines we’ve uploaded.

You can also integrate the upload command into your project template to automatically upload your packaged pipelines after training.

Notes about upgrading from v3.0

Pipeline package version compatibility

When you’re loading a pipeline package trained with spaCy v3.0, 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:

Updating v3.0 configs

To update a config from spaCy v3.0 with the new v3.1 settings, run init fill-config:

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.

Sourcing pipeline components with vectors

If you’re sourcing a pipeline component that requires static vectors (for example, a tagger or parser from an md or lg pretrained pipeline), be sure to include the source model’s vectors in the setting [initialize.vectors]. In spaCy v3.0, a bug allowed vectors to be loaded implicitly through source, however in v3.1 this setting must be provided explicitly as [initialize.vectors]:

config.cfg (excerpt)

spacy train and spacy assemble will provide warnings if the source and target pipelines don’t contain the same vectors. If you are sourcing a rule-based component like an entity ruler or lemmatizer that does not use the vectors as a model feature, then this warning can be safely ignored.

Warnings

Logger warnings have been converted to Python warnings. Use warnings.filterwarnings or the new helper method spacy.errors.filter_warning(action, error_msg='') to manage warnings.