def full_model_accuracy():
# train a classifier on the entire training data
train_parses = depparse.read_conllu(
"UD_English-EWT/en_ewt-ud-train.conllu")
classifier = depparse.Classifier(train_parses)
# test the classifier on the development set
correct = 0
total = 0
for deps in depparse.read_conllu("UD_English-EWT/en_ewt-ud-dev.conllu"):
total += len(deps)
# clear out all the head information
orig_heads = clear_heads(deps)
# parse using the classifier to predict actions
depparse.parse(deps, classifier)
# count how many of the heads have been correctly restored
for dep, orig_head in zip(deps, orig_heads):
if dep.head == orig_head:
correct += 1
# return the accuracy
return correct / total
def test_parse():
# consider a specific sentence from the training data
# # sent_id = weblog-blogspot.com_alaindewitt_20040929103700_ENG_20040929_103700-0026
# # text = The future president joined the Guard in May 1968.
# 1 The the DET DT Definite=Def|PronType=Art 3 det 3:det _
# 2 future future ADJ JJ Degree=Pos 3 amod 3:amod _
# 3 president president NOUN NN Number=Sing 4 nsubj 4:nsubj _
# 4 joined join VERB VBD Mood=Ind|Tense=Past|VerbForm=Fin 0 root 0:root _
# 5 the the DET DT Definite=Def|PronType=Art 6 det 6:det _
# 6 Guard Guard PROPN NNP Number=Sing 4 obj 4:obj _
# 7 in in ADP IN _ 8 case 8:case _
# 8 May May PROPN NNP Number=Sing 4 obl 4:obl:in _
# 9 1968 1968 NUM CD NumType=Card 8 nummod 8:nummod SpaceAfter=No
# 10 . . PUNCT . _ 4 punct 4:punct _
parses = depparse.read_conllu("UD_English-EWT/en_ewt-ud-train.conllu")
[deps] = itertools.islice(parses, 352, 353)
# clear out all the head information
orig_heads = clear_heads(deps)
# run the parser with the oracle list of actions
depparse.parse(
deps,
IterActions([
Action.SHIFT,
Action.SHIFT,
Action.SHIFT,
Action.LEFT_ARC,
Action.LEFT_ARC,
Action.SHIFT,
Action.LEFT_ARC,
Action.SHIFT,
Action.SHIFT,
Action.LEFT_ARC,
Action.RIGHT_ARC,
Action.SHIFT,
Action.SHIFT,
Action.LEFT_ARC,
Action.SHIFT,
Action.RIGHT_ARC,
Action.RIGHT_ARC,
Action.SHIFT,
Action.RIGHT_ARC,
]))
# make sure that the original heads have been restored by the parser
assert [dep.head for dep in deps] == orig_heads
def test_oracle():
# consider a specific sentence from the training data
# # sent_id = answers-20111108085734AATXy0E_ans-0004
# # text = Plaster of Paris does two things
# 1 Plaster plaster NOUN NN Number=Sing 4 nsubj 4:nsubj _
# 2 of of ADP IN _ 3 case 3:case _
# 3 Paris Paris PROPN NNP Number=Sing 1 nmod 1:nmod:of _
# 4 does do VERB VBZ Mood=Ind|Number=Sing|Person=3|Tense=Pres|VerbForm=Fin 0 root 0:root _
# 5 two two NUM CD NumType=Card 6 nummod 6:nummod _
# 6 things thing NOUN NNS Number=Plur 4 obj 4:obj _
parses = depparse.read_conllu("UD_English-EWT/en_ewt-ud-train.conllu")
[deps] = itertools.islice(parses, 7475, 7476)
# create an oracle for the sentence and try a few actions
oracle = depparse.Oracle(deps)
# shift on an empty stack
assert oracle([], deps) == Action.SHIFT
# shift on a stack with only one entry
assert oracle(deps[:1], deps[1:]) == Action.SHIFT
# shift because "Plaster" and "of" are not in a head-dependent relation
assert oracle(deps[:2], deps[2:]) == Action.SHIFT
# left-arc because "Paris" is the head of "of"
assert oracle(deps[:3], deps[3:]) == Action.LEFT_ARC
# right-arc because "Plaster" is the head of "Paris"
assert oracle(deps[:1] + deps[2:3], deps[3:]) == Action.RIGHT_ARC
# create a new oracle for the same sentence and extract all the actions
oracle = depparse.Oracle(deps)
depparse.parse(deps, oracle)
assert oracle.actions == [
Action.SHIFT, #1
Action.SHIFT, #2
Action.SHIFT, #3
Action.LEFT_ARC, #4
Action.RIGHT_ARC, #5
Action.SHIFT, #6
Action.LEFT_ARC, #7
Action.SHIFT, #8
Action.SHIFT, #9
Action.LEFT_ARC, #10
Action.RIGHT_ARC, #11
]
'''print([(dep.form, dep.head, dep.id) for dep in deps])
def test_oracle_round_trip():
# take the first 50 parses from the training data
parses = depparse.read_conllu("UD_English-EWT/en_ewt-ud-train.conllu")
for i, deps in enumerate(itertools.islice(parses, 50)):
# skip the non-projective parses
if i in {4, 21, 25, 31}:
continue
# collect the head for each word
orig_heads = [dep.head for dep in deps]
# run the oracle to determine the sequence of actions
oracle = depparse.Oracle(deps)
depparse.parse(deps, oracle)
# clear out all the head information
clear_heads(deps)
# feed the oracle-identified actions in, one at a time
depparse.parse(deps, IterActions(oracle.actions))
# make sure that the original heads have been restored by the parser
assert [dep.head for dep in deps] == orig_heads