Introducing spaCy v2.3 · Explosion

spaCy now speaks Chinese, Japanese, Danish, Polish and
Romanian! Version 2.3 of the spaCy Natural Language Processing library adds
models for five new languages. We’ve also updated all 15 model families with
word vectors and improved accuracy, while also decreasing model size and loading
times for models with vectors.

This is the last major release of v2, by the way. We’ve been working hard on
spaCy v3, which comes with a lot of cool improvements, especially for training,
configuration and custom modeling. We’ll start making prereleases on
spacy-nightly soon, so stay tuned.

spaCy v2.3 provides new model families for five languages:
Chinese, Danish,
Japanese, Polish and
Romanian. The Chinese and Japanese language models
are the first provided models that use external libraries for word segmentation
rather than spaCy’s tokenizer.

Chinese

The new Chinese models use pkuseg
for word segmentation and ship with a custom model trained on OntoNotes with a
token accuracy of 94.6%. Users can initialize the tokenizer with both pkuseg
and custom models and customize the user dictionary. Details can be found in the
Chinese docs. The Chinese tokenizer
continues to support jieba as the default word segmenter along with
character-based segmentation as in v2.2.

Japanese

The updated Japanese language class switches to
SudachiPy for word
segmentation and part-of-speech tagging. Using sudachipy greatly simplifies
installing spaCy for Japanese, which is now possible with a single command:
pip install spacy[ja]. More details are in the
Japanese docs.

Model Performance

Following our usual convention, the sm, md and lg models differ in their
word vectors. The lg models include one word vector for most words in the
training data, while the md model prunes the vectors table to only include
entries for the 20,000 most common words, mapping less frequent words to the
most similar vector in the reduced table. The sm models do not use pretrained
vectors.

Language	Model	Size	TAG	UAS	LAS	ENTS F
Chinese	zh_core_web_sm	45 MB	89.63	68.55	63.21	66.57
	zh_core_web_md	75 MB	90.23	69.39	64.43	68.52
	zh_core_web_lg	575 MB	90.55	69.77	64.99	69.33
Danish	da_core_news_sm	16 MB	92.79	80.48	75.65	72.79
	da_core_news_md	46 MB	94.13	82.71	78.98	81.45
	da_core_news_lg	546 MB	94.95	82.53	78.99	82.73
Japanese	ja_core_news_sm	7 MB	97.30	88.68	86.87	59.93
	ja_core_news_md	37 MB	97.30	89.26	87.76	67.68
	ja_core_news_lg	526 MB	97.30	88.94	87.55	70.48
Polish	pl_core_news_sm	46 MB	98.03	85.61	78.09	81.32
	pl_core_news_md	76 MB	98.28	90.41	84.47	84.68
	pl_core_news_lg	576 MB	98.45	90.80	85.52	85.67
Romanian	ro_core_news_sm	13 MB	95.65	87.20	79.79	71.05
	ro_core_news_md	43 MB	96.32	88.69	81.77	75.42
	ro_core_news_lg	545 MB	96.78	88.87	82.05	76.71

The training data for Danish, Japanese and Romanian is relatively small, so the
pretrained word vectors improve accuracy quite a lot, in particular for NER. The
Chinese model uses a larger training corpus, but word segmentation errors may
make the word vectors less effective. Word segmentation accuracy also explains
some of the lower scores for Chinese, as the model has to get the word
segmentation correct before it can be scored as accurate on any of the
subsequent tasks.

All model families now include medium and large models with 20k and 500k unique
vectors respectively. For most languages, spaCy v2.3 introduces custom word
vectors trained using spaCy’s language-specific tokenizers on data from
OSCAR and Wikipedia.
The vectors are trained using FastText with the same settings as
FastText’s word vectors (CBOW,
300 dimensions, character n-grams of length 5).

In particular for languages with smaller training corpora, the addition of word
vectors greatly improves the model accuracy. For example, the Lithuanian tagger
increases from 81.7% for the small model (no vectors) to 89.3% for the large
model. The parser increases by a similar margin and the NER F-score increases
from 66.0% to 70.1%. For German, updating the word vectors increases the scores
for the medium model for all components by 1.5 percentage points across the
board.

Model compatibility

Remember that models trained with v2.2 will be incompatible with the new
version. To find out if you need to update your models, you can run
python -m spacy validate. If you’re using your own custom models, you’ll need
to retrain them with the new version.

All spaCy training corpora based on
Universal Dependencies corpora have been
updated to UD v2.5 (v2.6 for Japanese, v2.3 for Polish). The updated data
improves the quality and size of the training corpora, increasing the tagger and
parser accuracy for all provided models. For example, the Dutch training data is
extended to include both UD Dutch Alpino and LassySmall, which improves the
tagger and parser scores for the small models by 3%, and the addition of the new
word vectors improve the scores further by 3-5%.

Fine-grained POS tags

As a result of the updates, many of the fine-grained part-of-speech tag sets
will differ from v2.2 models. The coarse-grained tag-set remains the same,
although there are some minor differences in how they are calculated from the
fine-grained tags.

For French, Italian, Portuguese and Spanish, the fine-grained part-of-speech tag
sets contain new merged tags related to contracted forms, such as ADP_DET for
French "au", which maps to UPOS ADP based on the head "à". This increases
the accuracy of the models by improving the alignment between spaCy’s
tokenization and Universal Dependencies multi-word tokens used for contractions.

The medium model packages with 20k vectors are at least 2× smaller than in
v2.2, the large English model is 120M smaller, and the loading times are
2-4× faster for all models with vectors. To achieve this, models no longer
store derivable lexeme attributes such as lower and is_alpha and the
remaining lexeme attributes (norm, cluster and prob) have been moved to
spacy-lookups-data.

If you’re training new models, you’ll probably want to install
spacy-lookups-data for normalization and lemmatization tables! The provided
models include the norm lookup tables for use with the core pipeline
components, but the optional cluster and prob features are now only
available through spacy-lookups-data.

We’re also proud to announce updates and translations of our online course,
“Advanced NLP with spaCy”. We’ve made a few small
updates to the English version, including new videos to go with the interactive
exercises. It’s really the translations we’re excited about though. We have
translations into Japanese, German and Spanish, with Chinese, French
and Russian soon to come.

Speaking of videos, you should also check out Sofie’s tutorial on training a
custom entity linking model with spaCy.
You can find the code and data in our growing
projects repository.

Another cool video to check out is the new episode of Vincent Warmerdam’s
“Intro to NLP with spaCy”
. The series lets you sit beside Vincent as he works through an example data
science project using spaCy. In episode 5,
“Rules vs. Machine Learning”,
Vincent uses spaCy’s rule-based matcher to probe the decisions of the NER model
he trained previously, using the rules to understand the model’s behavior and
figure out how to improve the training data to get better results.

spaCy v2.3 is the last big release of v2. We’ve been working hard on v3,
which we expect to start publishing prereleases of in the next few weeks. spaCy
v3 comes with a lot of cool improvements, especially for training,
configuration and custom modeling. The training and data formats are the main
thing we’ve taken the opportunity to fix, so v3 will have some breaking changes,
but don’t worry — it’s nothing like the big transformations seen in libraries
like TensorFlow or Angular. It should be pretty easy to upgrade, but we’ve still
tried to backport as much as possible into v2.3, so you can use it right away.
We’ll also continue to make maintenance releases of v2.3 with bug fixes as they
come in.

We also have a big release of our annotation tool Prodigy
pretty much ready to go. In addition to the spaCy v2.3 update (giving you all
the new models), Prodigy v1.10 comes with a new annotation interface for tasks
like relation extraction and coreference resolution, full-featured
audio and video annotation (including recipes using
pyannote.audio models in the
loop), a new and improved manual image UI, more options for NER annotation, new
recipe callbacks, and lots more. To get notified when it’s ready, follow us
on Twitter!

Resources