spacy-experimental: Cutting-edge experimental spaCy components and features
This package includes experimental components and features for spaCy v3.x, for example model architectures, pipeline components and utilities.
Installation
Install with pip:
python -m pip install -U pip setuptools wheel
python -m pip install spacy-experimental
Using spacy-experimental
Components and features may be modified or removed in any release, so always specify the exact version as a package requirement if you're experimenting with a particular component, e.g.:
spacy-experimental==0.147.0
Then you can add the experimental components to your config or import from spacy_experimental:
[components.experimental_edit_tree_lemmatizer]
factory = "experimental_edit_tree_lemmatizer"
Components
Edit tree lemmatizer
[components.experimental_edit_tree_lemmatizer]
factory = "experimental_edit_tree_lemmatizer"
# token attr to use as backoff with the predicted trees are not applicable; null to leave unset
backoff = "orth"
# prune trees that are applied less than this frequency in the training data
min_tree_freq = 2
# whether to overwrite existing lemma annotation
overwrite = false
scorer = {"@scorers":"spacy.lemmatizer_scorer.v1"}
# try to apply at most the k most probable edit trees
top_k = 1
Trainable character-based tokenizers
Two trainable tokenizers represent tokenization as a sequence tagging problem over individual characters and use the existing spaCy tagger and NER architectures to perform the tagging.
In the spaCy pipeline, a simple "pretokenizer" is applied as the pipeline tokenizer to split each doc into individual characters and the trainable tokenizer is a pipeline component that retokenizes the doc. The pretokenizer needs to be configured manually in the config or with spacy.blank():
nlp = spacy.blank(
"en",
config={
"nlp": {
"tokenizer": {"@tokenizers": "spacy-experimental.char_pretokenizer.v1"}
}
},
)
The two tokenizers currently reset any existing tag or entity annotation respectively in the process of retokenizing.
Character-based tagger tokenizer
In the tagger version experimental_char_tagger_tokenizer, the tagging problem is represented internally with character-level tags for token start (T), token internal (I), and outside a token (O). This representation comes from Elephant: Sequence Labeling for Word and Sentence Segmentation (Evang et al., 2013).
This is a sentence.
TIIIOTIOTOTIIIIIIIT
With the option annotate_sents, S replaces T for the first token in each sentence and the component predicts both token and sentence boundaries.
This is a sentence.
SIIIOTIOTOTIIIIIIIT
A config excerpt for experimental_char_tagger_tokenizer:
[nlp]
pipeline = ["experimental_char_tagger_tokenizer"]
tokenizer = {"@tokenizers":"spacy-experimental.char_pretokenizer.v1"}
[components]
[components.experimental_char_tagger_tokenizer]
factory = "experimental_char_tagger_tokenizer"
annotate_sents = true
scorer = {"@scorers":"spacy-experimental.tokenizer_senter_scorer.v1"}
[components.experimental_char_tagger_tokenizer.model]
@architectures = "spacy.Tagger.v1"
nO = null
[components.experimental_char_tagger_tokenizer.model.tok2vec]
@architectures = "spacy.Tok2Vec.v2"
[components.experimental_char_tagger_tokenizer.model.tok2vec.embed]
@architectures = "spacy.MultiHashEmbed.v2"
width = 128
attrs = ["ORTH","LOWER","IS_DIGIT","IS_ALPHA","IS_SPACE","IS_PUNCT"]
rows = [1000,500,50,50,50,50]
include_static_vectors = false
[components.experimental_char_tagger_tokenizer.model.tok2vec.encode]
@architectures = "spacy.MaxoutWindowEncoder.v2"
width = 128
depth = 4
window_size = 4
maxout_pieces = 2
Character-based NER tokenizer
In the NER version, each character in a token is part of an entity:
T B-TOKEN
h I-TOKEN
i I-TOKEN
s I-TOKEN
O
i B-TOKEN
s I-TOKEN
O
a B-TOKEN
O
s B-TOKEN
e I-TOKEN
n I-TOKEN
t I-TOKEN
e I-TOKEN
n I-TOKEN
c I-TOKEN
e I-TOKEN
. B-TOKEN
A config excerpt for experimental_char_ner_tokenizer:
[nlp]
pipeline = ["experimental_char_ner_tokenizer"]
tokenizer = {"@tokenizers":"spacy-experimental.char_pretokenizer.v1"}
[components]
[components.experimental_char_ner_tokenizer]
factory = "experimental_char_ner_tokenizer"
scorer = {"@scorers":"spacy-experimental.tokenizer_scorer.v1"}
[components.experimental_char_ner_tokenizer.model]
@architectures = "spacy.TransitionBasedParser.v2"
state_type = "ner"
extra_state_tokens = false
hidden_width = 64
maxout_pieces = 2
use_upper = true
nO = null
[components.experimental_char_ner_tokenizer.model.tok2vec]
@architectures = "spacy.Tok2Vec.v2"
[components.experimental_char_ner_tokenizer.model.tok2vec.embed]
@architectures = "spacy.MultiHashEmbed.v2"
width = 128
attrs = ["ORTH","LOWER","IS_DIGIT","IS_ALPHA","IS_SPACE","IS_PUNCT"]
rows = [1000,500,50,50,50,50]
include_static_vectors = false
[components.experimental_char_ner_tokenizer.model.tok2vec.encode]
@architectures = "spacy.MaxoutWindowEncoder.v2"
width = 128
depth = 4
window_size = 4
maxout_pieces = 2
The NER version does not currently support sentence boundaries, but it would be easy to extend using a B-SENT entity type.
Biaffine parser
A biaffine dependency parser, similar to that proposed in [Deep Biaffine Attention for Neural Dependency Parsing](Deep Biaffine Attention for Neural Dependency Parsing) (Dozat & Manning, 2016). The parser consists of two parts: an edge predicter and an edge labeler. For example:
[components.experimental_arc_predicter]
factory = "experimental_arc_predicter"
[components.experimental_arc_labeler]
factory = "experimental_arc_labeler"
The arc predicter requires that a previous component (such as senter) sets sentence boundaries during training. Therefore, such a component must be added to annotating_components:
[training]
annotating_components = ["senter"]
The biaffine parser sample project provides an example biaffine parser pipeline.
Architectures
None currently.
Other
Tokenizers
spacy-experimental.char_pretokenizer.v1: Tokenize a text into individual characters.
Scorers
spacy-experimental.tokenizer_scorer.v1: Score tokenization.spacy-experimental.tokenizer_senter_scorer.v1: Score tokenization and sentence segmentation.
Bug reports and issues
Please report bugs in the spaCy issue tracker or open a new thread on the discussion board for other issues.
Older documentation
See the READMEs in earlier tagged versions for details about components in earlier releases.
1 Jan 13, 2022
0 Feb 08, 2022
2 Dec 18, 2022
2 Dec 18, 2021
247 Jan 05, 2023
241 Jan 04, 2023
1 Jan 15, 2022
1.5k Jan 04, 2023
2 Oct 26, 2021
62 Dec 20, 2022
11.4k Jan 01, 2023
289 Jan 06, 2023
2 May 17, 2022
2.9k Dec 31, 2022
606 Dec 28, 2022
88 Dec 29, 2022
286 Dec 06, 2022
28 May 25, 2021
1 Mar 11, 2022
26 May 27, 2022