def antecedent_description(trig_sentdict, ant_sentdict, ant, POS_TAGS):
vector = []
ant_words = ant.get_words()
subtree = ant.get_subtree()
# Feature 1.
vector.append(truth(len(subtree.leaves()) == len(ant_words)))
# Feature 2.
vector.append(len(ant_words))
# Feature 3.
vector.append(
truth(
NT.dominates(subtree.root(), subtree,
ant.get_trigger().get_subtree())))
# Features 4.
pos_tags_dict = {}
for tag in POS_TAGS:
pos_tags_dict[tag] = 0
idx = get_antecedent_head_index(ant.get_context(), ant)
for tag in ant.get_context()['pos'][idx:len(ant_words)]:
pos_tags_dict[tag] += 1
vector += [pos_tags_dict[tag] for tag in pos_tags_dict]
# Feature 5: if the antecedent starts with an auxiliary, verb, adj.
vector.append(truth(DV.isauxiliary(ant_sentdict, idx)))
vector.append(truth(DV.isverb(ant_sentdict['pos'][idx])))
vector.append(truth(DV.isadj(ant_sentdict['pos'][idx])))
return vector
def antecedent_description(trig_sentdict, ant_sentdict, ant, POS_TAGS):
vector = []
ant_words = ant.get_words()
subtree = ant.get_subtree()
# Feature 1.
vector.append(truth(len(subtree.leaves()) == len(ant_words)))
# Feature 2.
vector.append(len(ant_words))
# Feature 3.
vector.append(truth(NT.dominates(subtree.root(), subtree, ant.get_trigger().get_subtree())))
# Features 4.
pos_tags_dict = {}
for tag in POS_TAGS:
pos_tags_dict[tag] = 0
idx = get_antecedent_head_index(ant.get_context(), ant)
for tag in ant.get_context()['pos'][idx:len(ant_words)]:
pos_tags_dict[tag] += 1
vector += [pos_tags_dict[tag] for tag in pos_tags_dict]
# Feature 5: if the antecedent starts with an auxiliary, verb, adj.
vector.append(truth(DV.isauxiliary(ant_sentdict, idx)))
vector.append(truth(DV.isverb(ant_sentdict['pos'][idx])))
vector.append(truth(DV.isadj(ant_sentdict['pos'][idx])))
return vector
def alignment_comparison(trig_sentdict, ant_sentdict, ant, trigger,
word2vec_dict):
vector = []
ant_context_sentdict = ant.get_context()
trig_context_sentdict = trigger.get_context()
ant_head_idx = get_antecedent_head_index(ant_sentdict, ant)
# Feature 1.
ant_auxs = []
for i in range(0, len(ant_sentdict['words'])):
if DV.isauxiliary(ant_sentdict, i):
ant_auxs.append(ant_sentdict['lemmas'][i])
found = False
for aux in ant_auxs:
if aux in trig_sentdict['lemmas']:
vector.append(truth(True))
found = True
break
if not found:
vector.append(truth(False))
# Feature 2.
if ant.get_sentnum() == trigger.get_sentnum():
vector.append(truth(ant_head_idx > trigger.get_idx()))
vector.append(truth(ant_head_idx == trigger.get_idx()))
vector.append(truth(ant_head_idx < trigger.get_idx()))
else:
vector += [0, 0, 0]
# Features 3,4,5.
for k in ['words', 'lemmas', 'pos']:
total = len(ant_context_sentdict[k]) + len(trig_context_sentdict[k])
common = len(
set(ant_context_sentdict[k]).intersection(
trig_context_sentdict[k]))
vector.append(common)
vector.append((2.0 * float(common)) / float(total))
# Feature 6 - number of words between trigger and antecedent.
vector.append(ant.get_sentnum() - trigger.get_sentnum())
if ant.get_sentnum() == trigger.get_sentnum():
vector.append(ant_head_idx - trigger.get_idx())
else:
crt_sentnum = trigger.get_sentnum()
distance = ant_head_idx
while crt_sentnum < ant.get_sentnum():
distance += len(trig_sentdict['words'])
crt_sentnum += 1
vector.append(distance)
# Feature 7.
# First we get the vecs from the Ant NP and average them.
blank_np = False
ant_np_word2vec = []
ant_np_location = ant.get_context()['np']
if ant_np_location != (-1, -1):
ant_np_word2vec = get_average_np_vec(word2vec_dict, ant_sentdict,
ant_np_location[0],
ant_np_location[1])
else:
blank_np = True
# Next we do the same for the Trigger NP.
trig_np_word2vec = []
trig_np_location = trigger.get_context()['np']
if trig_np_location != (-1, -1):
trig_np_word2vec = get_average_np_vec(word2vec_dict, trig_sentdict,
trig_np_location[0],
trig_np_location[1])
else:
blank_np = True
# Adding the angle of the vector between the trigger NP and antecedent NP.
if not blank_np:
ant_length = vector_length(ant_np_word2vec)
trig_length = vector_length(trig_np_word2vec)
try:
angle = angle_btwn_vectors(ant_np_word2vec,
trig_np_word2vec,
v1_length=ant_length,
v2_length=trig_length)
except ValueError:
angle = 90.0
vector.append(angle)
vector.append(truth(angle == 0.0))
else:
vector.append(90.0)
vector.append(truth(90.0 == 0.0))
if not ant_np_word2vec:
vector += [0 for _ in range(0, WORD2VEC_LENGTH)]
else:
vector += ant_np_word2vec
if not trig_np_word2vec:
vector += [0 for _ in range(0, WORD2VEC_LENGTH)]
else:
vector += trig_np_word2vec
# Now for what comes after the head.
ant_head_idx = get_antecedent_head_index(ant_sentdict, ant)
ant_post_head_w2vec = get_average_np_vec(word2vec_dict, ant_sentdict,
ant_head_idx,
len(ant_sentdict['words']))
# if not ant_post_head_w2vec: vector += [0 for i in range(0,WORD2VEC_LENGTH)]
# else: vector += ant_post_head_w2vec
stop_idx = len(trig_sentdict['words'])
for i in range(trigger.get_idx(), len(trig_sentdict['words'])):
if DV.ispunctuation(trig_sentdict['lemmas'][i]):
stop_idx = i
break
post_trig_w2vec = get_average_np_vec(word2vec_dict, trig_sentdict,
trigger.get_idx(), stop_idx)
# if not post_trig_w2vec: vector += [0 for i in range(0,WORD2VEC_LENGTH)]
# else: vector += post_trig_w2vec
if ant_post_head_w2vec and post_trig_w2vec:
try:
post_angle = angle_btwn_vectors(ant_post_head_w2vec,
post_trig_w2vec)
except ValueError:
post_angle = 90.0
vector.append(post_angle)
vector.append(truth(post_angle == 0.0))
else:
vector.append(90.0)
vector.append(truth(90.0 == 0.0))
# Sentenial complement check.
tree = NT.maketree(ant_sentdict['tree'][0])
if NT.dominates(tree, ant.get_subtree(), trigger.get_subtree()):
vector.append(
truth(
NT.has_phrases_between_trees(
ant.get_subtree(), trigger.get_subtree(),
NIELSON_SENTENIAL_COMPLEMENT_PHRASES)))
else:
vector.append(truth(False))
# Features to account for the number of each phrase type between the antecedent and trigger.
phrases_between = [0 for _ in ALL_PHRASES]
if ant.get_sentnum() == trigger.get_sentnum():
for i in range(0, len(phrases_between)):
if NT.has_phrases_between_trees(ant.get_subtree(),
trigger.get_subtree(),
[ALL_PHRASES[i]]):
phrases_between[i] += 1
vector += phrases_between
vector.append(sum(phrases_between))
return vector
def alignment_comparison(trig_sentdict, ant_sentdict, ant, trigger, word2vec_dict):
vector = []
ant_context_sentdict = ant.get_context()
trig_context_sentdict = trigger.get_context()
ant_head_idx = get_antecedent_head_index(ant_sentdict, ant)
# Feature 1.
ant_auxs = []
for i in range(0,len(ant_sentdict['words'])):
if DV.isauxiliary(ant_sentdict, i):
ant_auxs.append(ant_sentdict['lemmas'][i])
found = False
for aux in ant_auxs:
if aux in trig_sentdict['lemmas']:
vector.append(truth(True))
found = True
break
if not found:
vector.append(truth(False))
# Feature 2.
if ant.get_sentnum() == trigger.get_sentnum():
vector.append(truth(ant_head_idx > trigger.get_idx()))
vector.append(truth(ant_head_idx == trigger.get_idx()))
vector.append(truth(ant_head_idx < trigger.get_idx()))
else: vector += [0,0,0]
# Features 3,4,5.
for k in ['words','lemmas','pos']:
total = len(ant_context_sentdict[k])+len(trig_context_sentdict[k])
common = len(set(ant_context_sentdict[k]).intersection(trig_context_sentdict[k]))
vector.append(common)
vector.append((2.0*float(common))/float(total))
# Feature 6 - number of words between trigger and antecedent.
vector.append(ant.get_sentnum()-trigger.get_sentnum())
if ant.get_sentnum() == trigger.get_sentnum(): vector.append(ant_head_idx - trigger.get_idx())
else:
crt_sentnum = trigger.get_sentnum()
distance = ant_head_idx
while crt_sentnum < ant.get_sentnum():
distance += len(trig_sentdict['words'])
crt_sentnum += 1
vector.append(distance)
# Feature 7.
# First we get the vecs from the Ant NP and average them.
blank_np = False
ant_np_word2vec = []
ant_np_location = ant.get_context()['np']
if ant_np_location != (-1,-1):
ant_np_word2vec = get_average_np_vec(word2vec_dict, ant_sentdict, ant_np_location[0], ant_np_location[1])
else: blank_np = True
# Next we do the same for the Trigger NP.
trig_np_word2vec = []
trig_np_location = trigger.get_context()['np']
if trig_np_location != (-1,-1):
trig_np_word2vec = get_average_np_vec(word2vec_dict, trig_sentdict, trig_np_location[0], trig_np_location[1])
else: blank_np = True
# Adding the angle of the vector between the trigger NP and antecedent NP.
if not blank_np:
ant_length = vector_length(ant_np_word2vec)
trig_length = vector_length(trig_np_word2vec)
try:
angle = angle_btwn_vectors(ant_np_word2vec, trig_np_word2vec, v1_length=ant_length, v2_length=trig_length)
except ValueError:
angle = 90.0
vector.append(angle)
vector.append(truth(angle == 0.0))
else:
vector.append(90.0)
vector.append(truth(90.0 == 0.0))
if not ant_np_word2vec:
vector += [0 for _ in range(0,WORD2VEC_LENGTH)]
else:
vector += ant_np_word2vec
if not trig_np_word2vec:
vector += [0 for _ in range(0,WORD2VEC_LENGTH)]
else:
vector += trig_np_word2vec
# Now for what comes after the head.
ant_head_idx = get_antecedent_head_index(ant_sentdict, ant)
ant_post_head_w2vec = get_average_np_vec(word2vec_dict, ant_sentdict, ant_head_idx, len(ant_sentdict['words']))
# if not ant_post_head_w2vec: vector += [0 for i in range(0,WORD2VEC_LENGTH)]
# else: vector += ant_post_head_w2vec
stop_idx = len(trig_sentdict['words'])
for i in range(trigger.get_idx(), len(trig_sentdict['words'])):
if DV.ispunctuation(trig_sentdict['lemmas'][i]):
stop_idx = i
break
post_trig_w2vec = get_average_np_vec(word2vec_dict, trig_sentdict, trigger.get_idx(), stop_idx)
# if not post_trig_w2vec: vector += [0 for i in range(0,WORD2VEC_LENGTH)]
# else: vector += post_trig_w2vec
if ant_post_head_w2vec and post_trig_w2vec:
try:
post_angle = angle_btwn_vectors(ant_post_head_w2vec, post_trig_w2vec)
except ValueError: post_angle = 90.0
vector.append(post_angle)
vector.append(truth(post_angle == 0.0))
else:
vector.append(90.0)
vector.append(truth(90.0 == 0.0))
# Sentenial complement check.
tree = NT.maketree(ant_sentdict['tree'][0])
if NT.dominates(tree, ant.get_subtree(), trigger.get_subtree()):
vector.append(truth( NT.has_phrases_between_trees(ant.get_subtree(), trigger.get_subtree(), NIELSON_SENTENIAL_COMPLEMENT_PHRASES)))
else:
vector.append(truth(False))
# Features to account for the number of each phrase type between the antecedent and trigger.
phrases_between = [0 for _ in ALL_PHRASES]
if ant.get_sentnum() == trigger.get_sentnum():
for i in range(0,len(phrases_between)):
if NT.has_phrases_between_trees(ant.get_subtree(), trigger.get_subtree(), [ALL_PHRASES[i]]):
phrases_between[i] += 1
vector += phrases_between
vector.append(sum(phrases_between))
return vector