Sense2vec Versions Save

• Fix issue #155: Pin vectors to the CPU after deserialization (https://github.com/explosion/sense2vec/pull/157).

• In the sense2vec.teach prodigy recipe: only fail if no seeds are available.
• Extend support for wasabi to v1.1.x.

• Update component and internals for spaCy v3.

• Various small fixes and improvements.
• Improve training scripts.
• Fix issue #102: split binary .spacy files.
• Fix issue #118: Fix typo in s2v_other_senses.

Thanks to @ahalterman, @dshefman1 and @Anxo06 for the pull requests!

🔴 Bug fixes

• Add defaults for config if attributes are not included in saved model.
• Fix serialization and deserialization of string store in component.

🔴 Bug fixes

• Fix bug that'd cause the scores to not be read correctly from precomputed most_similar caches.

✨ New features and improvements

• Completely rewrite package from scratch.
• Replace built-in vector storage with spaCy's Vectors, making this package a pure Python package and allowing easy out-of-the-box serialization of vectors.
• Add fully serializable spaCy pipeline component and extension attributes.
• Add new methods get_best_sense and get_other_senses and improve most_similar.
• Add script for precomputing index of nearest neighbors for super fast "most similar" queries.
• Add annotation recipes for Prodigy to easily create word lists and match patterns from similar phrases using sense2vec vectors (like the terms.teach recipe, just with multi-word expressions).
• New and more efficient training and preprocessing scripts using GloVe and fastText.

⚠️ Backwards incompatibilities

• The sense2vec.load method has been removed. Use Sense2Vec.from_disk instead.
• The previous VectorMap and VectorStorage have been removed.
• This package now requires Python 3.6+.
• This update requires a new vectors format (see attached files).

📖 Documentation and examples

• Rewrite README from scratch and include full API docs.

👥 Contributors

Thanks to @kabirkhan for contributing the initial Prodigy recipes!