def conllup_to_starsem(fname, sherlock_train="data/sherlock/cdt.conllup", semcue=False):
sentences = cd.read_col_data(fname)
train = cd.read_col_data(sherlock_train)
_vocabs = vcb.make_vocabs(train, 0)
vocabs = vcb.Vocabs(*_vocabs)
w2i = vocabs.scoperels.w2i
for sentence in sentences:
#sid, story = sentence.id.split(maxsplit=1)
story = "_"
sid = sentence.id.split()[0]
matrix = sentence.make_matrix("scope", label=True, w2i=w2i)
if semcue:
cmatrix = sentence.make_matrix("sem", label=True, w2i=w2i)
cues = [i for i in range(len(cmatrix)) if cmatrix[0,i] == w2i["cue"] and not w2i["mwc"] in cmatrix[:,i]]
else:
#cues = [i for i in range(len(matrix)) if matrix[0,i] == w2i["cue"] and not w2i["mwc"] in matrix[:,i]]
cues = [i for i in range(len(matrix)) if matrix[0,i] == w2i["cue"]]
#for h in range(len(matrix)):
# if sum(matrix[h,:]) > 0:
# cues.append(h)
if len(cues) > 0:
# cue scope event
for word in sentence:
negs = ["_", "_", "_"] * len(cues)
for i,c in enumerate(cues):
p = len(cues) - 1 - i
p = i
myev = word.form
if c == word.id:
is_incue, mycue, myev = check_cue(word.form)
if not is_incue:
mycue = word.form
negs[3*p] = mycue
if matrix[c,word.id] == w2i["event"]:
negs[3*p+1] = myev
negs[3*p+2] = myev
elif matrix[c,word.id] == w2i["scope"]:
negs[3*p+1] = myev#word.form
elif matrix[c,word.id] == w2i["mwc"]:
negs[3*p+0] = myev#word.form
elif semcue:
if cmatrix[c,word.id] == w2i["mwc"]:
negs[3*p+0] = myev#word.form
print("\t".join([story, sid, str(word.id-1), word.form, word.lemma, word.xpos, "_", *negs]))
else:
for word in sentence:
print("\t".join([story, sid, str(word.id-1), word.form, word.lemma, word.xpos, "_", "***"]))
print()
def conllup_to_epe(fname):
import col_data as cd
s = 0
for sentence in cd.read_col_data(fname):
sid, story = sentence.id.split(maxsplit=1)
epe = {"id": sid, "nodes": []}
cues = {}
nodes = epe["nodes"]
c = 0
for token in sentence:
node = {"id": token.id,
"form": token.form,
"start": s,
"end": s + len(token.form),
"properties": {"xpos": token.xpos,
"upos": token.upos,
"lemma": token.lemma},
"edges": [],
"negation": []}
s += len(token.form) + 1
nodes.append(node)
if "cue" in [l for h,l in token.scope] and not "mwc" in [l for h,l in token.scope]:
if token.id not in cues:
cues[token.id] = c
c += 1
elif "cue" in [l for h,l in token.scope] and "mwc" in [l for h,l in token.scope]:
try:
cues[token.id] = cues[[h for h,l in token.scope if l == "mwc"][0]]
except KeyError:
cues[[h for h,l in token.scope if l == "mwc"][0]] = c
c += 1
cues[token.id] = cues[[h for h,l in token.scope if l == "mwc"][0]]
for token in sentence:
if token.head == 0:
nodes[token.id-1]["top"] = True
elif token.head > 0:
nodes[token.head-1]["edges"].append({"label": token.deprel, "target": token.id})
for h,l in token.deps:
nodes[h-1]["edges"].append({"label": l, "target": token.id})
for h,l in token.scope:
if h == token.id:
is_incue, mycue, myev = check_cue(token.form)
if not is_incue:
mycue = token.form
myev = token.form
if l == "scope":
nodes[token.id-1]["negation"].append({"id": cues[token.id], "cue": mycue, "scope": myev})
elif l == "event":
nodes[token.id-1]["negation"].append({"id": cues[token.id], "cue": mycue, "scope": myev, "event": myev})
elif l == "cue":
#print((cues[token.id], [x["id"] for x in nodes[token.id-1]["negation"]]))
if not (cues[token.id] in [x["id"] for x in nodes[token.id-1]["negation"]]):
nodes[token.id-1]["negation"].append({"id": cues[token.id], "cue": token.form})
elif l == "scope":
nodes[token.id-1]["negation"].append({"id": cues[h], "scope": token.form})
elif l == "event":
nodes[token.id-1]["negation"].append({"id": cues[h], "scope": token.form, "event": token.form})
print(json.dumps(epe))
def predict(model, settings, to_predict, elmo, vocabs):
pred_path = settings.dir + to_predict.split("/")[-1] + ".pred"
entries, predicted, other_predicted = model.predict(to_predict, elmo)
f1, _ = sc.score(*zip(*((entry[1][settings.pt].numpy(), predicted[entry[0]].numpy()) for entry in entries)))
print("F1 is {:.2%}".format(f1))
if len(other_predicted) > 0:
other_f1, _ = sc.score(*zip(*((entry[1][settings.ot].numpy(), other_predicted[entry[0]].numpy()) for entry in entries)))
print("Other F1 is {:.2%}".format(other_f1))
with open(pred_path, "w") as fh:
for sentence in cd.read_col_data(to_predict):
pred = predicted[sentence.id].numpy()
if settings.target_style == "scope-":
cue_matrix = sentence.make_matrix("cues", True, vocabs[settings.td["cue"]].w2i)
pred = np.maximum(pred, cue_matrix)
#pred = other_predicted[sentence.id].numpy()
sentence.update_parse(pred, settings.target_style, vocabs[settings.pt].i2w)
if len(other_predicted) > 0:
pred = other_predicted[sentence.id].numpy()
# NOTE sem == sem hopefully
if settings.target_style == settings.other_target_style:
sentence.update_parse(pred, "syn", vocabs[settings.pt].i2w)
else:
sentence.update_parse(pred, settings.other_target_style, vocabs[settings.pt].i2w)
print(sentence, file=fh)
return True
def _load_data(self, data_path, pos_style, target_style,
other_target_style, elmo):
print("Loading data from {}".format(data_path))
data = cd.read_col_data(data_path)
#with h5py.File(elmo, 'r') as f:
# for sen in f:
# #print(sen)
# for word, vec in zip(sen.split("\t"), f[sen]):
# print(word, vec)
if self.use_elmo:
felmo = h5py.File(elmo, "r")
self.index_entries = []
for sentence in data:
#print(sentence.id)
#print(len(sentence), len(felmo[sentence.id]))
if self.use_elmo:
self.index_entries.append(
IndexEntry(sentence, self.vocabs, self.external,
self.settings, felmo[sentence.id],
self.vec_dim))
else:
self.index_entries.append(
IndexEntry(sentence, self.vocabs, self.external,
self.settings, None))
if self.use_elmo:
felmo.close()
print("Done")
def coldata_to_starsem(fn_in: str, fn_out: str) -> None:
with open(fn_out, "w") as fh_out:
for sentence in read_col_data(fn_in):
# print(sentence, file=fh_out)
negs = negations_from_matrix(negation_matrix(sentence), sentence)
out = []
for token in sentence:
out.append(
f"_\t{sentence.id}\t{token.id-1}\t{token.form}\t_\t_\t_\t")
if not negs:
out.append("***\n")
else:
out_neg = []
for neg in negs.values():
if token.id in neg["Cue"]:
if token.form not in no_affix:
form, span = affixer(token.form)
else:
form = token.form
out_neg.append(form)
else:
out_neg.append("_")
if token.id in neg["Scope"] and \
token.id not in neg["Cue"]: # cue in other negation's scope
form = token.form
out_neg.append(form)
elif token.id in neg["Cue"]:
form, (start, end) = affixer(token.form)
if token.form in no_affix:
form = "_"
elif start == 0 and end != len(
token.form): # prefix
# print(sentence.id, end="\t")
# print("prefix", token.form, form, span, neg)
# print([(t.id, t.form) for t in sentence])
form = token.form[end:]
elif start != 0 and end == len(
token.form): # suffix
# print(sentence.id, end="\t")
# print("suffix", token.form, form, span, neg)
# print([(t.id, t.form) for t in sentence])
form = token.form[:start]
else:
form = "_"
out_neg.append(form)
else:
out_neg.append("_")
out_neg.append("_") # event remnant
out.append("\t".join(out_neg))
out.append("\n")
print("".join(out), file=fh_out)
for i in range(n):
for j in range(n):
C[int(gl[i, j]), int(pl[i, j])] += 1
print(C)
for i in range(len(C)):
print(i2w[i])
fscore(i, C)
#for j in range(len(C)):
# print("\t", i2w[j], C[i,j])
if __name__ == "__main__":
import col_data as cd
import vocab as vcb
import sys
try:
with open("vocabs.pk", "rb") as fh:
vocabs = pickle.load(fh)
except FileNotFoundError:
train = cd.read_col_data(sys.argv[1])
_vocabs = vcb.make_vocabs(train, 0)
vocabs = vcb.Vocabs(*_vocabs)
gold = cd.read_col_data(sys.argv[2])
pred = cd.read_col_data(sys.argv[3])
gms = [g.make_matrix("scope", True, vocabs.scoperels.w2i) for g in gold]
pms = [p.make_matrix("scope", True, vocabs.scoperels.w2i) for p in pred]
confuse(gms, pms, vocabs.scoperels.i2w)
def run_parser(args):
# For now, assume there always is train, val, and glove data
if args.seed == -1:
args.seed = np.random.randint(1234567890)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
device = torch.device("cuda:0" if torch.cuda.is_available()
and not args.force_cpu else "cpu")
print(device)
args.device = device
if torch.cuda.is_available():
print(torch.cuda.get_device_capability(device))
args.td = {None: 0, "syn": 1, "sem": 2, "cue": 3, "scope": 4, "scope-": 5}
args.ot = args.td[args.other_target_style]
args.pt = args.td[args.target_style]
args.helpers = None
if args.help_style:
args.helpers = [args.td[x] for x in args.help_style.split(",")]
if not args.dir.endswith("/"):
args.dir += "/"
if args.load:
with open(args.dir + "vocabs.pk", "rb") as fh:
vocabs = pickle.load(fh)
#args.vocabs = vocabs
model = ModelInteractor.factory(args, vocabs)
model.load(args.load)
else:
sentences = cd.read_col_data(args.train)
if args.vocab is not None:
with open(args.vocab, "rb") as fh:
vocabs = pickle.load(fh)
else:
_vocabs = make_vocabs(sentences, 0)
vocabs = Vocabs(*_vocabs)
with open(args.dir + "vocabs.pk", "wb") as fh:
pickle.dump(vocabs, fh)
#args.vocabs = vocabs
model = ModelInteractor.factory(args, vocabs)
if args.recycle is not None:
with open(args.recycle + "vocabs.pk", "rb") as fh:
other_vocabs = pickle.load(fh)
with open(args.recycle + "settings.json") as fh:
other_settings = json.load(fh)
other_settings = Namespace(**other_settings)
other_settings.device = args.device
other = ModelInteractor.factory(other_settings, other_vocabs)
other.load(args.recycle + "best_model.save")
model.upd_from_other(other, *args.recycle_layers.split(","))
if args.freeze is not None:
model.freeze_params(*args.freeze.split(","))
if (args.load and args.cont) or args.load is None:
model.train()
# load the best_model.save instead of using the current one
model = ModelInteractor.factory(args, vocabs)
model.load(args.dir + "best_model.save")
if (args.load and args.cont) or args.load is None:
predict(model, args, args.val, args.elmo_dev, vocabs)
predict(model, args, args.predict_file, args.elmo_test, vocabs)
return len(self.rels)
def __getitem__(self, index):
return self.rels[index]
def __setitem__(self, index, value):
self.rels[index] = value
if __name__ == "__main__":
import sys
import col_data as cd
import pickle
sentences = []
for fn in sys.argv[2:]:
sentences.extend(cd.read_col_data(fn))
#sentences = cd.read_col_data(sys.argv[1])
forms, norms, lemmas, uposs, xposs, synrels, semrels, chars, scoperels = make_vocabs(
sentences)
print([
len(v.w2i) for v in [
forms, norms, lemmas, uposs, xposs, synrels, semrels, chars,
scoperels
]
])
vocabs = Vocabs(forms, norms, lemmas, uposs, xposs, synrels, semrels,
chars, scoperels)
#print(synrels.w2i, semrels.w2i, scoperels.w2i)
with open(sys.argv[1], "wb") as fh:
pickle.dump(vocabs, fh)
for scope in scopes.values():
scope = sorted(scope)
#print(scope)
for i, j in enumerate(scope):
if i < len(scope) - 1:
next_n = scope[i + 1]
dist = next_n - j
#print(dist)
if dist > 1:
return True
return False
if __name__ == "__main__":
gold = dict([(l.id, l) for l in cd.read_col_data(
"../data/neg_graphs/point_to_root/test.conllu")])
pred = dict([(l.id, l) for l in cd.read_col_data(
"../experiments/point_to_root/2/test.conllu.pred")])
gold_neg = get_only_negated(gold)
pred_neg = get_only_negated(pred)
# which cues in gold but not predicted?
print("Which cues in gold but not predicted?")
gmissed, gsids = which_gold_cues_are_missed(gold_neg, pred)
for cue, count in gmissed.most_common():
print("-- {}:{}".format(cue, count))
# --subquestion: does the model EVER predict these cues?
never_predicted = [cue for cue in gmissed]
for sent_id, sent in pred_neg.items():
# Find which experiments have been run
#experiment_names = set(name_map.keys())
#experiments_run = set(os.listdir(args.preddir))
#to_check = experiment_names.intersection(experiments_run)
for setup in args.experiments:
metric = []
metric.append("")
metric.append(name_map[setup])
#goldfile = os.path.join(args.golddir, setup, "test.conllu")
#predfile = os.path.join(args.preddir, setup, "test.conllu.pred")
goldfile = os.path.join(args.golddir, setup, "dev.conllu")
predfile = os.path.join(args.preddir, setup, "dev.conllu.pred")
gold = list(cd.read_col_data(goldfile))
pred = list(cd.read_col_data(predfile))
for label in ["holder", "targ", "exp"]:
prec, rec, f1 = span_f1(gold, pred, mapping, test_label=label)
metric.append(f1 * 100)
#print("{0}: {1:.1f}".format(label, f1 * 100))
lgold = read_labeled(goldfile)
lpred = read_labeled(predfile)
ugold = read_unlabeled(goldfile)
upred = read_unlabeled(predfile)
#print("Targeted F1")
f1 = targeted_f1(lgold, lpred)
metric.append(f1 * 100)