SpaCy Pattern Builder

Use training examples to build and refine patterns for use with SpaCy's DependencyMatcher.

Motivation

Generating patterns programmatically from training data is more efficient than creating them manually.

Installation

With pip:

pip install spacy-pattern-builder

Usage

# Import a SpaCy model, parse a string to create a Doc object
import en_core_web_sm

text = 'We introduce efficient methods for fitting Boolean models to molecular data.'
nlp = en_core_web_sm.load()
doc = nlp(text)

from spacy_pattern_builder import build_dependency_pattern

# Provide a list of tokens we want to match.
match_tokens = [doc[i] for i in [0, 1, 3]]  # [We, introduce, methods]

''' Note that these tokens must be fully connected. That is,
all tokens must have a path to all other tokens in the list,
without needing to traverse tokens outside of the list.
Otherwise, spacy-pattern-builder will raise a TokensNotFullyConnectedError.
You can get a connected set that includes your tokens with the following: '''
from spacy_pattern_builder import util
connected_tokens = util.smallest_connected_subgraph(match_tokens, doc)
assert match_tokens == connected_tokens  # In this case, the tokens we provided are already fully connected

# Specify the token attributes / features to use
feature_dict = {  # This is equal to the default feature_dict
    'DEP': 'dep_',
    'TAG': 'tag_'
}

# Build the pattern
pattern = build_dependency_pattern(doc, match_tokens, feature_dict=feature_dict)

from pprint import pprint
pprint(pattern)  # In the format consumed by SpaCy's DependencyMatcher:
'''
[{'PATTERN': {'DEP': 'ROOT', 'TAG': 'VBP'}, 'SPEC': {'NODE_NAME': 'node1'}},
 {'PATTERN': {'DEP': 'nsubj', 'TAG': 'PRP'},
  'SPEC': {'NBOR_NAME': 'node1', 'NBOR_RELOP': '>', 'NODE_NAME': 'node0'}},
 {'PATTERN': {'DEP': 'dobj', 'TAG': 'NNS'},
  'SPEC': {'NBOR_NAME': 'node1', 'NBOR_RELOP': '>', 'NODE_NAME': 'node3'}}]
'''

# Create a matcher and add the newly generated pattern
from spacy.matcher import DependencyMatcher

matcher = DependencyTreeMatcher(doc.vocab)
matcher.add('pattern', None, pattern)

# And get matches
matches = matcher(doc)
for match_id, token_idxs in matches:
    tokens = [doc[i] for i in token_idxs]
    tokens = sorted(tokens, key=lambda w: w.i)  # Make sure tokens are in their original order
    print(tokens)  # [We, introduce, methods]

Acknowledgements

Uses:

• SpaCy
• networkx