2020-07-22 11:42:59 +00:00
|
|
|
# cython: infer_types=True, profile=True, binding=True
|
|
|
|
from typing import Optional, Iterable
|
2020-07-30 21:30:54 +00:00
|
|
|
from thinc.api import Model, Config
|
2020-07-22 11:42:59 +00:00
|
|
|
|
2020-07-30 21:30:54 +00:00
|
|
|
from .transition_parser cimport Parser
|
|
|
|
from ._parser_internals.arc_eager cimport ArcEager
|
2020-07-22 11:42:59 +00:00
|
|
|
|
|
|
|
from .functions import merge_subtokens
|
|
|
|
from ..language import Language
|
2020-07-30 21:30:54 +00:00
|
|
|
from ._parser_internals import nonproj
|
Refactor the Scorer to improve flexibility (#5731)
* Refactor the Scorer to improve flexibility
Refactor the `Scorer` to improve flexibility for arbitrary pipeline
components.
* Individual pipeline components provide their own `evaluate` methods
that score a list of `Example`s and return a dictionary of scores
* `Scorer` is initialized either:
* with a provided pipeline containing components to be scored
* with a default pipeline containing the built-in statistical
components (senter, tagger, morphologizer, parser, ner)
* `Scorer.score` evaluates a list of `Example`s and returns a dictionary
of scores referring to the scores provided by the components in the
pipeline
Significant differences:
* `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc`
and the new `morph_acc`, `pos_acc`, and `lemma_acc`
* Scoring is no longer cumulative: `Scorer.score` scores a list of
examples rather than a single example and does not retain any state
about previously scored examples
* PRF values in the returned scores are no longer multiplied by 100
* Add kwargs to Morphologizer.evaluate
* Create generalized scoring methods in Scorer
* Generalized static scoring methods are added to `Scorer`
* Methods require an attribute (either on Token or Doc) that is
used to key the returned scores
Naming differences:
* `uas`, `las`, and `las_per_type` in the scores dict are renamed to
`dep_uas`, `dep_las`, and `dep_las_per_type`
Scoring differences:
* `Doc.sents` is now scored as spans rather than on sentence-initial
token positions so that `Doc.sents` and `Doc.ents` can be scored with
the same method (this lowers scores since a single incorrect sentence
start results in two incorrect spans)
* Simplify / extend hasattr check for eval method
* Add hasattr check to tokenizer scoring
* Simplify to hasattr check for component scoring
* Reset Example alignment if docs are set
Reset the Example alignment if either doc is set in case the
tokenization has changed.
* Add PRF tokenization scoring for tokens as spans
Add PRF scores for tokens as character spans. The scores are:
* token_acc: # correct tokens / # gold tokens
* token_p/r/f: PRF for (token.idx, token.idx + len(token))
* Add docstring to Scorer.score_tokenization
* Rename component.evaluate() to component.score()
* Update Scorer API docs
* Update scoring for positive_label in textcat
* Fix TextCategorizer.score kwargs
* Update Language.evaluate docs
* Update score names in default config
2020-07-25 10:53:02 +00:00
|
|
|
from ..scorer import Scorer
|
2020-09-09 08:31:03 +00:00
|
|
|
from ..training import validate_examples
|
2020-07-22 11:42:59 +00:00
|
|
|
|
|
|
|
|
|
|
|
default_model_config = """
|
|
|
|
[model]
|
|
|
|
@architectures = "spacy.TransitionBasedParser.v1"
|
2020-09-23 14:53:49 +00:00
|
|
|
state_type = "parser"
|
2020-09-23 11:35:09 +00:00
|
|
|
extra_state_tokens = false
|
2020-07-22 11:42:59 +00:00
|
|
|
hidden_width = 64
|
|
|
|
maxout_pieces = 2
|
|
|
|
|
|
|
|
[model.tok2vec]
|
|
|
|
@architectures = "spacy.HashEmbedCNN.v1"
|
|
|
|
pretrained_vectors = null
|
|
|
|
width = 96
|
|
|
|
depth = 4
|
|
|
|
embed_size = 2000
|
|
|
|
window_size = 1
|
|
|
|
maxout_pieces = 3
|
|
|
|
subword_features = true
|
|
|
|
"""
|
|
|
|
DEFAULT_PARSER_MODEL = Config().from_str(default_model_config)["model"]
|
|
|
|
|
|
|
|
|
|
|
|
@Language.factory(
|
|
|
|
"parser",
|
2020-07-31 08:55:44 +00:00
|
|
|
assigns=["token.dep", "token.head", "token.is_sent_start", "doc.sents"],
|
2020-07-22 11:42:59 +00:00
|
|
|
default_config={
|
|
|
|
"moves": None,
|
|
|
|
"update_with_oracle_cut_size": 100,
|
|
|
|
"learn_tokens": False,
|
|
|
|
"min_action_freq": 30,
|
|
|
|
"model": DEFAULT_PARSER_MODEL,
|
2020-07-26 11:18:43 +00:00
|
|
|
},
|
2020-09-24 08:27:33 +00:00
|
|
|
default_score_weights={
|
|
|
|
"dep_uas": 0.5,
|
|
|
|
"dep_las": 0.5,
|
|
|
|
"dep_las_per_type": None,
|
|
|
|
"sents_p": None,
|
|
|
|
"sents_r": None,
|
|
|
|
"sents_f": 0.0,
|
|
|
|
},
|
2020-07-22 11:42:59 +00:00
|
|
|
)
|
|
|
|
def make_parser(
|
|
|
|
nlp: Language,
|
|
|
|
name: str,
|
|
|
|
model: Model,
|
|
|
|
moves: Optional[list],
|
|
|
|
update_with_oracle_cut_size: int,
|
|
|
|
learn_tokens: bool,
|
|
|
|
min_action_freq: int
|
|
|
|
):
|
2020-08-09 12:46:13 +00:00
|
|
|
"""Create a transition-based DependencyParser component. The dependency parser
|
|
|
|
jointly learns sentence segmentation and labelled dependency parsing, and can
|
|
|
|
optionally learn to merge tokens that had been over-segmented by the tokenizer.
|
|
|
|
|
|
|
|
The parser uses a variant of the non-monotonic arc-eager transition-system
|
|
|
|
described by Honnibal and Johnson (2014), with the addition of a "break"
|
|
|
|
transition to perform the sentence segmentation. Nivre's pseudo-projective
|
|
|
|
dependency transformation is used to allow the parser to predict
|
|
|
|
non-projective parses.
|
|
|
|
|
|
|
|
The parser is trained using an imitation learning objective. The parser follows
|
|
|
|
the actions predicted by the current weights, and at each state, determines
|
|
|
|
which actions are compatible with the optimal parse that could be reached
|
|
|
|
from the current state. The weights such that the scores assigned to the
|
|
|
|
set of optimal actions is increased, while scores assigned to other
|
|
|
|
actions are decreased. Note that more than one action may be optimal for
|
|
|
|
a given state.
|
|
|
|
|
2020-08-09 14:10:48 +00:00
|
|
|
model (Model): The model for the transition-based parser. The model needs
|
|
|
|
to have a specific substructure of named components --- see the
|
|
|
|
spacy.ml.tb_framework.TransitionModel for details.
|
|
|
|
moves (List[str]): A list of transition names. Inferred from the data if not
|
|
|
|
provided.
|
2020-08-09 12:46:13 +00:00
|
|
|
update_with_oracle_cut_size (int):
|
|
|
|
During training, cut long sequences into shorter segments by creating
|
|
|
|
intermediate states based on the gold-standard history. The model is
|
|
|
|
not very sensitive to this parameter, so you usually won't need to change
|
|
|
|
it. 100 is a good default.
|
|
|
|
learn_tokens (bool): Whether to learn to merge subtokens that are split
|
|
|
|
relative to the gold standard. Experimental.
|
|
|
|
min_action_freq (int): The minimum frequency of labelled actions to retain.
|
|
|
|
Rarer labelled actions have their label backed-off to "dep". While this
|
|
|
|
primarily affects the label accuracy, it can also affect the attachment
|
|
|
|
structure, as the labels are used to represent the pseudo-projectivity
|
|
|
|
transformation.
|
|
|
|
"""
|
2020-07-22 11:42:59 +00:00
|
|
|
return DependencyParser(
|
|
|
|
nlp.vocab,
|
|
|
|
model,
|
|
|
|
name,
|
|
|
|
moves=moves,
|
|
|
|
update_with_oracle_cut_size=update_with_oracle_cut_size,
|
2020-08-09 12:46:13 +00:00
|
|
|
multitasks=[],
|
2020-07-22 11:42:59 +00:00
|
|
|
learn_tokens=learn_tokens,
|
|
|
|
min_action_freq=min_action_freq
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
|
|
cdef class DependencyParser(Parser):
|
|
|
|
"""Pipeline component for dependency parsing.
|
|
|
|
|
2020-09-04 10:58:50 +00:00
|
|
|
DOCS: https://nightly.spacy.io/api/dependencyparser
|
2020-07-22 11:42:59 +00:00
|
|
|
"""
|
|
|
|
TransitionSystem = ArcEager
|
|
|
|
|
|
|
|
@property
|
|
|
|
def postprocesses(self):
|
|
|
|
output = [nonproj.deprojectivize]
|
|
|
|
if self.cfg.get("learn_tokens") is True:
|
|
|
|
output.append(merge_subtokens)
|
|
|
|
return tuple(output)
|
|
|
|
|
|
|
|
def add_multitask_objective(self, mt_component):
|
|
|
|
self._multitasks.append(mt_component)
|
|
|
|
|
2020-09-29 16:33:16 +00:00
|
|
|
def init_multitask_objectives(self, get_examples, nlp=None, **cfg):
|
2020-07-22 11:42:59 +00:00
|
|
|
# TODO: transfer self.model.get_ref("tok2vec") to the multitask's model ?
|
|
|
|
for labeller in self._multitasks:
|
|
|
|
labeller.model.set_dim("nO", len(self.labels))
|
|
|
|
if labeller.model.has_ref("output_layer"):
|
|
|
|
labeller.model.get_ref("output_layer").set_dim("nO", len(self.labels))
|
2020-09-29 16:33:16 +00:00
|
|
|
labeller.initialize(get_examples, nlp=nlp)
|
2020-07-22 11:42:59 +00:00
|
|
|
|
|
|
|
@property
|
|
|
|
def labels(self):
|
|
|
|
labels = set()
|
|
|
|
# Get the labels from the model by looking at the available moves
|
|
|
|
for move in self.move_names:
|
|
|
|
if "-" in move:
|
|
|
|
label = move.split("-")[1]
|
|
|
|
if "||" in label:
|
|
|
|
label = label.split("||")[1]
|
|
|
|
labels.add(label)
|
|
|
|
return tuple(sorted(labels))
|
Refactor the Scorer to improve flexibility (#5731)
* Refactor the Scorer to improve flexibility
Refactor the `Scorer` to improve flexibility for arbitrary pipeline
components.
* Individual pipeline components provide their own `evaluate` methods
that score a list of `Example`s and return a dictionary of scores
* `Scorer` is initialized either:
* with a provided pipeline containing components to be scored
* with a default pipeline containing the built-in statistical
components (senter, tagger, morphologizer, parser, ner)
* `Scorer.score` evaluates a list of `Example`s and returns a dictionary
of scores referring to the scores provided by the components in the
pipeline
Significant differences:
* `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc`
and the new `morph_acc`, `pos_acc`, and `lemma_acc`
* Scoring is no longer cumulative: `Scorer.score` scores a list of
examples rather than a single example and does not retain any state
about previously scored examples
* PRF values in the returned scores are no longer multiplied by 100
* Add kwargs to Morphologizer.evaluate
* Create generalized scoring methods in Scorer
* Generalized static scoring methods are added to `Scorer`
* Methods require an attribute (either on Token or Doc) that is
used to key the returned scores
Naming differences:
* `uas`, `las`, and `las_per_type` in the scores dict are renamed to
`dep_uas`, `dep_las`, and `dep_las_per_type`
Scoring differences:
* `Doc.sents` is now scored as spans rather than on sentence-initial
token positions so that `Doc.sents` and `Doc.ents` can be scored with
the same method (this lowers scores since a single incorrect sentence
start results in two incorrect spans)
* Simplify / extend hasattr check for eval method
* Add hasattr check to tokenizer scoring
* Simplify to hasattr check for component scoring
* Reset Example alignment if docs are set
Reset the Example alignment if either doc is set in case the
tokenization has changed.
* Add PRF tokenization scoring for tokens as spans
Add PRF scores for tokens as character spans. The scores are:
* token_acc: # correct tokens / # gold tokens
* token_p/r/f: PRF for (token.idx, token.idx + len(token))
* Add docstring to Scorer.score_tokenization
* Rename component.evaluate() to component.score()
* Update Scorer API docs
* Update scoring for positive_label in textcat
* Fix TextCategorizer.score kwargs
* Update Language.evaluate docs
* Update score names in default config
2020-07-25 10:53:02 +00:00
|
|
|
|
|
|
|
def score(self, examples, **kwargs):
|
2020-07-27 16:11:45 +00:00
|
|
|
"""Score a batch of examples.
|
|
|
|
|
|
|
|
examples (Iterable[Example]): The examples to score.
|
|
|
|
RETURNS (Dict[str, Any]): The scores, produced by Scorer.score_spans
|
|
|
|
and Scorer.score_deps.
|
|
|
|
|
2020-09-04 10:58:50 +00:00
|
|
|
DOCS: https://nightly.spacy.io/api/dependencyparser#score
|
2020-07-27 16:11:45 +00:00
|
|
|
"""
|
2020-08-11 21:29:31 +00:00
|
|
|
validate_examples(examples, "DependencyParser.score")
|
Refactor the Scorer to improve flexibility (#5731)
* Refactor the Scorer to improve flexibility
Refactor the `Scorer` to improve flexibility for arbitrary pipeline
components.
* Individual pipeline components provide their own `evaluate` methods
that score a list of `Example`s and return a dictionary of scores
* `Scorer` is initialized either:
* with a provided pipeline containing components to be scored
* with a default pipeline containing the built-in statistical
components (senter, tagger, morphologizer, parser, ner)
* `Scorer.score` evaluates a list of `Example`s and returns a dictionary
of scores referring to the scores provided by the components in the
pipeline
Significant differences:
* `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc`
and the new `morph_acc`, `pos_acc`, and `lemma_acc`
* Scoring is no longer cumulative: `Scorer.score` scores a list of
examples rather than a single example and does not retain any state
about previously scored examples
* PRF values in the returned scores are no longer multiplied by 100
* Add kwargs to Morphologizer.evaluate
* Create generalized scoring methods in Scorer
* Generalized static scoring methods are added to `Scorer`
* Methods require an attribute (either on Token or Doc) that is
used to key the returned scores
Naming differences:
* `uas`, `las`, and `las_per_type` in the scores dict are renamed to
`dep_uas`, `dep_las`, and `dep_las_per_type`
Scoring differences:
* `Doc.sents` is now scored as spans rather than on sentence-initial
token positions so that `Doc.sents` and `Doc.ents` can be scored with
the same method (this lowers scores since a single incorrect sentence
start results in two incorrect spans)
* Simplify / extend hasattr check for eval method
* Add hasattr check to tokenizer scoring
* Simplify to hasattr check for component scoring
* Reset Example alignment if docs are set
Reset the Example alignment if either doc is set in case the
tokenization has changed.
* Add PRF tokenization scoring for tokens as spans
Add PRF scores for tokens as character spans. The scores are:
* token_acc: # correct tokens / # gold tokens
* token_p/r/f: PRF for (token.idx, token.idx + len(token))
* Add docstring to Scorer.score_tokenization
* Rename component.evaluate() to component.score()
* Update Scorer API docs
* Update scoring for positive_label in textcat
* Fix TextCategorizer.score kwargs
* Update Language.evaluate docs
* Update score names in default config
2020-07-25 10:53:02 +00:00
|
|
|
def dep_getter(token, attr):
|
|
|
|
dep = getattr(token, attr)
|
|
|
|
dep = token.vocab.strings.as_string(dep).lower()
|
|
|
|
return dep
|
|
|
|
results = {}
|
|
|
|
results.update(Scorer.score_spans(examples, "sents", **kwargs))
|
2020-07-31 08:55:44 +00:00
|
|
|
kwargs.setdefault("getter", dep_getter)
|
2020-09-05 12:08:34 +00:00
|
|
|
kwargs.setdefault("ignore_labels", ("p", "punct"))
|
2020-07-31 08:55:44 +00:00
|
|
|
results.update(Scorer.score_deps(examples, "dep", **kwargs))
|
2020-07-27 10:27:40 +00:00
|
|
|
del results["sents_per_type"]
|
Refactor the Scorer to improve flexibility (#5731)
* Refactor the Scorer to improve flexibility
Refactor the `Scorer` to improve flexibility for arbitrary pipeline
components.
* Individual pipeline components provide their own `evaluate` methods
that score a list of `Example`s and return a dictionary of scores
* `Scorer` is initialized either:
* with a provided pipeline containing components to be scored
* with a default pipeline containing the built-in statistical
components (senter, tagger, morphologizer, parser, ner)
* `Scorer.score` evaluates a list of `Example`s and returns a dictionary
of scores referring to the scores provided by the components in the
pipeline
Significant differences:
* `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc`
and the new `morph_acc`, `pos_acc`, and `lemma_acc`
* Scoring is no longer cumulative: `Scorer.score` scores a list of
examples rather than a single example and does not retain any state
about previously scored examples
* PRF values in the returned scores are no longer multiplied by 100
* Add kwargs to Morphologizer.evaluate
* Create generalized scoring methods in Scorer
* Generalized static scoring methods are added to `Scorer`
* Methods require an attribute (either on Token or Doc) that is
used to key the returned scores
Naming differences:
* `uas`, `las`, and `las_per_type` in the scores dict are renamed to
`dep_uas`, `dep_las`, and `dep_las_per_type`
Scoring differences:
* `Doc.sents` is now scored as spans rather than on sentence-initial
token positions so that `Doc.sents` and `Doc.ents` can be scored with
the same method (this lowers scores since a single incorrect sentence
start results in two incorrect spans)
* Simplify / extend hasattr check for eval method
* Add hasattr check to tokenizer scoring
* Simplify to hasattr check for component scoring
* Reset Example alignment if docs are set
Reset the Example alignment if either doc is set in case the
tokenization has changed.
* Add PRF tokenization scoring for tokens as spans
Add PRF scores for tokens as character spans. The scores are:
* token_acc: # correct tokens / # gold tokens
* token_p/r/f: PRF for (token.idx, token.idx + len(token))
* Add docstring to Scorer.score_tokenization
* Rename component.evaluate() to component.score()
* Update Scorer API docs
* Update scoring for positive_label in textcat
* Fix TextCategorizer.score kwargs
* Update Language.evaluate docs
* Update score names in default config
2020-07-25 10:53:02 +00:00
|
|
|
return results
|