def parse(text):
"""
Primary function to run syntaxnet and PredPatt over input sentences.
"""
parse_tree, trace = annotate_text(text)
conll_parsed = parse_to_conll(parse_tree)
conll_pp = [ud_parse for sent_id, ud_parse in load_conllu(conll_parsed)][0]
#PredPatt options. Modify as needed.
resolve_relcl = True # relative clauses
resolve_appos = True # appositional modifiers
resolve_amod = True # adjectival modifiers
resolve_conj = True # conjuction
resolve_poss = True # possessives
ud = dep_v2.VERSION # the version of UD
opts = PredPattOpts(resolve_relcl=resolve_relcl,
resolve_appos=resolve_appos,
resolve_amod=resolve_amod,
resolve_conj=resolve_conj,
resolve_poss=resolve_poss,
ud=ud)
ppatt = PredPatt(conll_pp, opts=opts)
predicate_deps, arg_deps = get_ud_fragments(ppatt)
#NOTE:
#This returns the pretty print formatted string from PredPatt. This is done
#largely as a place holder for JSON compatability within the REST API.
return {'predpatt': {'predicate_deps': predicate_deps,
'arg_deps': arg_deps},
'conll': conll_parsed,
'original': text}
def extract_predpatt_text(row, eid_num:int):
'''
Given a pandas dataframe of TB data
and eid_num (1 or 2)
output predpatt predicate text
(adds copula fillers in text)
'''
tokenid = getattr(row, f'eid{eid_num}_token_id')
conllu_string = getattr(row, f'eid{eid_num}_sent_conllu')
parsed_tb = [PredPatt(ud_parse, opts=options) for sent_id, ud_parse in load_conllu(conllu_string)]
pred_objects = parsed_tb[0].instances
curr_text = getattr(row, f'eid{eid_num}_text')
pred_match = False
#print(f"{(row['docid'], row['eventInstanceID'], row['relatedToEventInstance'])}")
if pred_objects:
for pred in pred_objects:
if int(pred.root.position)==int(tokenid):
pred_match = True
pred_object = pred
break
else:
pred_match=False
if pred_match:
pred_text, _, _, _ = predicate_info(pred_object)
return pred_text
else:
return curr_text
else:
return getattr(row, f'eid{eid_num}_text')
def test(data):
from predpatt import PredPatt, load_conllu
def fail(g, t):
if len(g) != len(t):
return True
else:
for i in g:
if i not in t:
return True
no_color = lambda x, _: x
count, failed = 0, 0
ret = ""
for sent_id, ud_parse in load_conllu(data):
count += 1
pp = PredPatt(ud_parse)
sent = ' '.join(t.text for t in pp.tokens)
linearized_pp = linearize(pp)
gold_preds = [
predicate.format(C=no_color, track_rule=False)
for predicate in pp.instances if likely_to_be_pred(predicate)
]
test_preds = pprint_preds(
construct_pred_from_flat(linearized_pp.split()))
if fail(gold_preds, test_preds):
failed += 1
ret += (
"Sent: %s\nLinearized PredPatt:\n\t%s\nGold:\n%s\nYours:\n%s\n\n"
% (sent, linearized_pp, "\n".join(gold_preds),
"\n".join(test_preds)))
print(ret)
print("You have test %d instances, and %d failed the test." %
(count, failed))
def from_conll(cls,
corpus: Union[str, TextIO],
name: str = 'ewt',
options: Optional[PredPattOpts] = None) -> 'PredPattCorpus':
"""Load a CoNLL dependency corpus and apply predpatt
Parameters
----------
corpus
(path to) a .conllu file
name
the name of the corpus; used in constructing treeids
options
options for predpatt extraction
"""
options = DEFAULT_PREDPATT_OPTIONS if options is None else options
corp_is_str = isinstance(corpus, str)
if corp_is_str and splitext(basename(corpus))[1] == '.conllu':
with open(corpus) as infile:
data = infile.read()
elif corp_is_str:
data = corpus
else:
data = corpus.read()
# load the CoNLL dependency parses as graphs
ud_corp = {name+'-'+str(i+1): [line.split()
for line in block.split('\n')
if len(line) > 0
if line[0] != '#']
for i, block in enumerate(data.split('\n\n'))}
ud_corp = CoNLLDependencyTreeCorpus(ud_corp)
# extract the predpatt for those dependency parses
try:
predpatt = {name+'-'+sid.split('_')[1]: PredPatt(ud_parse,
opts=options)
for sid, ud_parse in load_conllu(data)}
except ValueError:
errmsg = 'PredPatt was unable to parse the CoNLL you provided.' +\
' This is likely due to using a version of UD that is' +\
' incompatible with PredPatt. Use of version 1.2 is' +\
' suggested.'
raise ValueError(errmsg)
return cls({n: (pp, ud_corp[n])
for n, pp in predpatt.items()})
def setup_graph():
ud = DependencyGraphBuilder.from_conll(listtree, 'tree1')
pp = PredPatt(next(load_conllu(rawtree))[1],
opts=PredPattOpts(resolve_relcl=True,
borrow_arg_for_relcl=True,
resolve_conj=False,
cut=True))
graph = PredPattGraphBuilder.from_predpatt(pp, ud, 'tree1')
return pp, graph
def generate_predicates(
abstract_text:str,
pred_patt_opts=None
)->Iterable[Tuple[str, str, str]]:
"Requires that pred_util:nlp and pred_util:stopwords be initialized"
nlp = dpg.get("pred_util:nlp")
parser = Spacy2ConllParser(nlp=nlp)
stopwords = dpg.get("pred_util:stopwords")
doc = nlp(abstract_text)
for sent in doc.sents:
# if the sentence is very long
if len(sent) >= 20:
word_count = defaultdict(int)
for tok in sent:
word_count[str(tok)] += 1
# if one word dominates the long sentence
if max(word_count.values()) >= len(sent)*0.2:
continue # we likely generated the same word over-and-over
conllu = "".join(list(parser.parse(input_str=str(sent))))
for _, pred_patt_parse in load_conllu(conllu):
predicates = PredPatt(
pred_patt_parse,
opts=pred_patt_opts
).instances
for predicate in predicates:
# We only care about 2-entity predicates
if len(predicate.arguments) == 2:
a_ents, b_ents = [
# Get the set of entities
filter(
# Not in the stopword list
lambda x: x not in stopwords,
[str(e).strip() for e in nlp(args.phrase()).ents]
)
# For each argument
for args in predicate.arguments
]
# Slight cleaning needed to better match the predicate phrase
# Note, that PredPatt predicates use ?a and ?b placeholders
predicate_stmt = (
re.match(
r".*\?a(.*)\?b.*", # get text between placeholders
predicate.phrase()
)
.group(1) # get the group matched between the placeholders
.strip()
)
if len(predicate_stmt) > 0:
# We're going to iterate all predicates
for a, b in product(a_ents, b_ents):
if a != b:
yield (a, predicate_stmt, b)
def extract_predpatt(path='../../data/corpora/ud/UD_English-EWT-r1.2/'):
'''
Extract PredPatt objects from CONLLU files
'''
options = PredPattOpts(resolve_relcl=True,
borrow_arg_for_relcl=True,
resolve_conj=False,
cut=True) # Resolve relative clause
patt = {}
for file in os.listdir(path):
if file.endswith('.conllu'):
with open(path + file, 'r') as infile:
for sent_id, ud_parse in load_conllu(infile.read()):
patt[file + " " + sent_id] = \
PredPatt(ud_parse, opts=options)
return patt
def extract(self, sentence: str) -> List[Dict[str, Any]]:
processed = self.pipeline.process(sentence, self._error)
if self._error.occurred():
print(f"=== Error occurred: {self._error.message}")
self._error = ProcessingError()
return None
else:
conll_example = [ud_parse for sent_id, ud_parse in load_conllu(processed)][
0
]
ppatt = PredPatt(conll_example, opts=self._opts)
result = []
for predicate in ppatt.instances:
structure = {
"predicate": predicate.tokens,
"arguments": [x.tokens for x in predicate.arguments],
}
result.append(structure)
return result
def main():
# Data Locations
parser = argparse.ArgumentParser(
description='Recast UDS-Time duration to NLI format.')
parser.add_argument('--udstime',
type=str,
default='time_eng_ud_v1.2_2015_10_30.tsv',
help='UDS-Time tsv dataset file location.')
parser.add_argument(
'--split',
type=str,
default='',
help='If specified (train, dev, test), only that split is recasted')
parser.add_argument('--out_train',
type=str,
default='train/',
help='recasted train data folder location ')
parser.add_argument('--out_dev',
type=str,
default='dev/',
help='recasted train data folder location')
parser.add_argument('--out_test',
type=str,
default='test/',
help='recasted train data folder location ')
args = parser.parse_args()
# ### Import UDS Time
uds_time = pd.read_csv(args.udstime, sep="\t")
ewt = doc_utils.Corpus(uds_time=uds_time)
df = ewt.process_data
#######################################################
## Add features to UDS-time dataframe
#######################################################
df['Pred1.UPOS'] = df.apply(
lambda row: get_predicate_pos(row, ewt, event=1), axis=1)
df['Pred2.UPOS'] = df.apply(
lambda row: get_predicate_pos(row, ewt, event=2), axis=1)
## Extract Predicate Full Text
predicate_dict = {}
for ud_data_path in ud_data:
covered_set = set()
fname = ud_data_path.split("/")[-1]
data_name = fname.split(".")[0].split("-")[-1]
#print(f"Start processing: {data_name}")
with open(ud_data_path) as infile:
data = infile.read()
parsed = [(PredPatt(ud_parse, opts=options), sent_id)
for sent_id, ud_parse in load_conllu(data)]
for pred_object, sentid in parsed:
sentnum = sentid.split("_")[-1]
sentenceid = fname + " " + sentnum
for predicate_object in pred_object.instances:
#print(f"sentenceid: {sentenceid}, pred: {predicate_object}")
pred_text, _, pred_root_token, _ = predicate_info(
predicate_object)
predicate_dict[sentenceid + "_" +
str(pred_root_token)] = pred_text
#print(f"error at sentid :{sentenceid}")
print(f"Finished creating predicate dictionary for : {data_name}\n")
df['Pred1.Text.Full'] = df['Event1.ID'].map(lambda x: predicate_dict[x])
df['Pred2.Text.Full'] = df['Event2.ID'].map(lambda x: predicate_dict[x])
#######################################################
## Recast Data
#######################################################
pairid = -1 # count total pair ids
# Count event-pairs skipped due to ambiguous text for highlighting predicate.
skipcount = 0
if args.split:
splits = [args.split]
else:
splits = ['train', 'dev', 'test']
for split in splits:
data = []
metadata = []
curr_df = df[df['Split'] == split]
print(f"Creating NLI instances for Data split: {split}")
event_pair_ids = list(curr_df.groupby(['Event.Pair.ID']).groups.keys())
pbar = tqdm(total=len(event_pair_ids))
for idx, event_pair_id in enumerate(event_pair_ids):
## Predicate 1
recasted_data, recasted_metadata, pairid, skipcount = create_duration_NLI(
event_pair_id,
df,
ewt,
pairid=pairid,
skipcount=skipcount,
event=1,
sliding_window=1)
if recasted_data:
data += recasted_data
metadata += recasted_metadata
## Predicate 2
recasted_data, recasted_metadata, pairid, skipcount = create_duration_NLI(
event_pair_id,
df,
ewt,
pairid=pairid,
skipcount=skipcount,
event=2,
sliding_window=1)
if recasted_data:
data += recasted_data
metadata += recasted_metadata
# if pairid%10000==0:
# print(f"Total pair-ids processed so far: {pairid}, skipped so far: {skipcount}")
pbar.update(1)
out_folder = {
'train': args.out_train,
'dev': args.out_dev,
'test': args.out_test
}
print(
f"Total pair-ids processed so far: {pairid}, skipped so far: {skipcount}"
)
with open(out_folder[split] + "recast_temporal-duration_data.json",
'w') as out_data:
json.dump(data, out_data, indent=4)
with open(out_folder[split] + "recast_temporal-duration_metadata.json",
'w') as out_metadata:
json.dump(metadata, out_metadata, indent=4)
print(f"Total pair-ids: {pairid}")
print(f'Total events skipped: {skipcount}')
]
home = expanduser("~/Downloads/")
parsed = {'train': [], 'devte': []}
out_data = []
options = PredPattOpts(resolve_relcl=True,
borrow_arg_for_relcl=True,
resolve_conj=False,
cut=True) # Resolve relative clause
path = home + '/UD_English-r1.2/en-ud-train.conllu'
with open(path, 'r') as infile:
data = infile.read()
parsed['train'] += [('en-ud-train.conllu' + " " + sent_id,
PredPatt(ud_parse, opts=options))
for sent_id, ud_parse in load_conllu(data)]
for file in files:
path = home + file
with open(path, 'r') as infile:
data = infile.read()
parsed['devte'] += [(file[17:] + " " + sent_id,
PredPatt(ud_parse, opts=options))
for sent_id, ud_parse in load_conllu(data)]
c = {'train': 0, 'devte': 0}
d = {'train': 0, 'dev': 0, 'test': 0}
ign = {'train': 0, 'devte': 0}
prons_incl = [
"you", "they", "yourself", "themselves", "them", "themself", "theirself",
"theirselves"
for line in f.readlines():
feats = line.split('\t')
features[feats[0]] = [feats[1].split(), feats[2].split()]
# Load the predpatt objects for creating features
files = ['/Downloads/UD_English-r1.2/en-ud-train.conllu',
'/Downloads/UD_English-r1.2/en-ud-dev.conllu',
'/Downloads/UD_English-r1.2/en-ud-test.conllu']
options = PredPattOpts(resolve_relcl=True, borrow_arg_for_relcl=True, resolve_conj=False, cut=True) # Resolve relative clause
patt = {}
for file in files:
path = home + file
with open(path, 'r') as infile:
for sent_id, ud_parse in load_conllu(infile.read()):
patt[file[27:] + " " + sent_id] = PredPatt(ud_parse, opts=options)
data['Structure'] = data['Sentence.ID'].map(lambda x: (patt[x], features[x]))
# Split the datasets into train, dev, test
data_test = data[data['Split'] == 'test'].reset_index(drop=True)
data_dev = data[data['Split'] == 'dev'].reset_index(drop=True)
data = data[data['Split'] == 'train'].reset_index(drop=True)
# Ridit scoring annotations and confidence ratings
# for attr in attributes:
# resp = attr_map[attr]
# resp_conf = attr_conf[attr]
# data[resp_conf + ".norm"] = data.groupby('Annotator.ID')[resp_conf].transform(ridit)
# data_dev[resp_conf + ".norm"] = data_dev.groupby('Annotator.ID')[resp_conf].transform(ridit)
print(
"Processing {} of Genre: {}, Progress {}/{} ({} %), Num Skipped: {}"
.format(currbook, currgenre, num_processed, num_books,
num_processed / (num_books * 1.0), num_skipped))
decomp_lines_json = [json.loads(x) for x in decomp_lines]
book_conll_files = [
fi for fi in os.listdir(conlldir)
if parse_conll_filename(fi)[1][0] == currbook
]
for conllfi in book_conll_files: #For each chunk in the book
conllfile = os.path.join(conlldir, conllfi)
genre, book, doc_id = parse_conll_filename(conllfile)
conll_iter = load_conllu(conllfile)
decomp_lines_json_chunk = [
x for x in decomp_lines_json if x['doc-id'] == doc_id
] #get the lines associated with this chunk
line_idx = 0 #Where we are in the decomp json file
valid_instance = True
for sent_id, parse in conll_iter:
sent_id = int(sent_id.split('_')[1])
if line_idx >= len(decomp_lines_json_chunk):
break
if decomp_lines_json_chunk[line_idx][
'sent-id'] == sent_id: #check if there is a matching decomp extraction for this conll line
json_line = decomp_lines_json_chunk[line_idx]
id = 1
files = ['/UD_English-r1.2/en-ud-dev.conllu',
'/UD_English-r1.2/en-ud-test.conllu']
home = expanduser("~/Downloads/")
parsed = {'train': [], 'devte': []}
out_data = []
# Resolve relative clause
options = PredPattOpts(resolve_relcl=True, borrow_arg_for_relcl=True, resolve_conj=False, cut=True)
path = home + '/UD_English-r1.2/en-ud-train.conllu'
with open(path, 'r') as infile:
data = infile.read()
parsed['train'] += [('en-ud-train.conllu' + " " + sent_id, PredPatt(ud_parse, opts=options)) for sent_id, ud_parse in load_conllu(data)]
for file in files:
path = home + file
with open(path, 'r') as infile:
data = infile.read()
parsed['devte'] += [(file[17:] + " " + sent_id, PredPatt(ud_parse, opts=options)) for sent_id, ud_parse in load_conllu(data)]
# random.shuffle(parsed['train'])
c = {'train': 0, 'devte': 0}
d = {'train': 0, 'dev': 0, 'test': 0}
copp = {'train': 0, 'devte': 0}
auxverb = {'train': 0, 'devte': 0}
ign = {'train': 0, 'devte': 0}
adj = {'train': 0, 'devte': 0}
for write_file in ['pred_train_data.csv', 'pred_devte_data.csv']:
def hand_engineering(prot, batch_size, data, data_dev):
'''
Hand engineered feature extraction. Supports the following - UD,
Verbnet classids, Wordnet supersenses, concreteness ratings, LCS
eventivity scores
'''
home = expanduser("~")
framnet_posdict = {
'V': 'VERB',
'N': 'NOUN',
'A': 'ADJ',
'ADV': 'ADV',
'PREP': 'ADP',
'NUM': 'NUM',
'INTJ': 'INTJ',
'ART': 'DET',
'C': 'CCONJ',
'SCON': 'SCONJ',
'PRON': 'PRON',
'IDIO': 'X',
'AVP': 'ADV'
}
# Load the features
features = {}
with open(home + '/Desktop/protocols/data/features-2.tsv', 'r') as f:
for line in f.readlines():
feats = line.split('\t')
features[feats[0]] = (feats[1].split(), feats[2].split())
# Load the predpatt objects for creating features
files = [
'/Downloads/UD_English-r1.2/en-ud-train.conllu',
'/Downloads/UD_English-r1.2/en-ud-dev.conllu',
'/Downloads/UD_English-r1.2/en-ud-test.conllu'
]
home = expanduser("~")
options = PredPattOpts(resolve_relcl=True,
borrow_arg_for_relcl=True,
resolve_conj=False,
cut=True) # Resolve relative clause
patt = {}
for file in files:
path = home + file
with open(path, 'r') as infile:
for sent_id, ud_parse in load_conllu(infile.read()):
patt[file[33:][:-7] + " " + sent_id] = PredPatt(ud_parse,
opts=options)
data['Structure'] = data['Split.Sentence.ID'].map(lambda x:
(patt[x], features[x]))
data_dev['Structure'] = data_dev['Split.Sentence.ID'].map(
lambda x: (patt[x], features[x]))
raw_x = data['Structure'].tolist()
raw_dev_x = data_dev['Structure'].tolist()
all_x = raw_x + raw_dev_x
all_feats = '|'.join(['|'.join(all_x[i][1][0]) for i in range(len(all_x))])
feature_cols = Counter(all_feats.split('|'))
# All UD dataset features
all_ud_feature_cols = list(
feature_cols.keys()) + [(a + "_dep") for a in feature_cols.keys()]
# Concreteness
f = open(home + '/Desktop/protocols/data/concrete.pkl', 'rb')
concreteness = pickle.load(f)
if prot == 'arg':
conc_cols = ['concreteness']
else:
conc_cols = ['concreteness', 'max_conc', 'min_conc']
f.close()
# LCS eventivity
from lcsreader import LexicalConceptualStructureLexicon
lcs = LexicalConceptualStructureLexicon(
home + '/Desktop/protocols/data/verbs-English.lcs')
lcs_feats = ['lcs_eventive', 'lcs_stative']
# Wordnet supersenses(lexicographer names)
supersenses = list(
set(['supersense=' + x.lexname() for x in wordnet.all_synsets()]))
# Framenet
lem2frame = {}
for lm in framenet.lus():
for lemma in lm['lexemes']:
lem2frame[lemma['name'] + '.' +
framnet_posdict[lemma['POS']]] = lm['frame']['name']
frame_names = ['frame=' + x.name for x in framenet.frames()]
# Verbnet classids
verbnet_classids = ['classid=' + vcid for vcid in verbnet.classids()]
# Lexical features
lexical_feats = [
'can', 'could', 'should', 'would', 'will', 'may', 'might', 'must',
'ought', 'dare', 'need'
] + [
'the', 'an', 'a', 'few', 'another', 'some', 'many', 'each', 'every',
'this', 'that', 'any', 'most', 'all', 'both', 'these'
]
dict_feats = {}
for f in verbnet_classids + lexical_feats + supersenses + frame_names + lcs_feats + all_ud_feature_cols + conc_cols:
dict_feats[f] = 0
x_pd = pd.DataFrame([
features_func(sent_feat=sent,
token=token,
lemma=lemma,
dict_feats=dict_feats.copy(),
prot=prot,
concreteness=concreteness,
lcs=lcs,
l2f=lem2frame) for sent, token, lemma in
zip(raw_x, data['Root.Token'].tolist(), data['Lemma'].tolist())
])
dev_x_pd = pd.DataFrame([
features_func(sent_feat=sent,
token=token,
lemma=lemma,
dict_feats=dict_feats.copy(),
prot=prot,
concreteness=concreteness,
lcs=lcs,
l2f=lem2frame)
for sent, token, lemma in zip(raw_dev_x, data_dev['Root.Token'].tolist(
), data_dev['Lemma'].tolist())
])
# Figure out which columns to drop(they're always zero)
todrop1 = dev_x_pd.columns[(dev_x_pd == 0).all()].values.tolist()
todrop = x_pd.columns[(x_pd == 0).all()].values.tolist()
intdrop = [a for a in todrop if a not in todrop1]
cols_to_drop = cols_to_drop = list(set(todrop) - set(intdrop))
x = x_pd.drop(cols_to_drop, axis=1).values.tolist()
dev_x = dev_x_pd.drop(cols_to_drop, axis=1).values.tolist()
x = [[a[:] for a in x[i:i + batch_size]]
for i in range(0, len(data), batch_size)]
dev_x = [[a[:] for a in dev_x[i:i + batch_size]]
for i in range(0, len(data_dev), batch_size)]
return x, dev_x
