def gold_analysis(self):
print 'Triggers:',
freq = {}
for trig in self.train_triggers + self.test_triggers + self.val_triggers:
if not freq.has_key(trig.type):
freq[trig.type] = 1
else:
freq[trig.type] += 1
print freq
print 'Total:', sum(freq.values())
d = {'starts_with_aux': [], 'ends_with_aux': []}
all_pos = set()
for s in self.sentences:
for p in s.pos:
all_pos.add(p)
ant_start_pos = set()
for trig in self.train_triggers:
if wc.is_aux_lemma(trig.gold_ant.sub_sentdict.words[0]):
d['starts_with_aux'].append(trig.gold_ant)
ant_start_pos.add(trig.gold_ant.sub_sentdict.pos[0])
ant_end_pos = set()
for trig in self.train_triggers:
if wc.is_aux_lemma(trig.gold_ant.sub_sentdict.words[-1]):
d['ends_with_aux'].append(trig.gold_ant)
ant_start_pos.add(trig.gold_ant.sub_sentdict.pos[-1])
print 'Ants never start with these tags: ', all_pos - ant_start_pos
print 'Percent of ants that start with auxs: ', len(d['starts_with_aux']) / float(len(self.train_triggers))
print 'Ants never END with these tags: ', all_pos - ant_end_pos
print 'Percent of ants that END with auxs: ', len(d['ends_with_aux']) / float(len(self.train_triggers))
def get_head(self, idx=False, idx_in_subsentdict=False):
"""
@type return: str
"""
for i in range(len(self.sub_sentdict)):
if wc.is_verb(self.sub_sentdict.pos[i]) and not wc.is_aux_lemma(self.sub_sentdict.lemmas[i]):
if idx:
if idx_in_subsentdict:
return i
else:
return self.start + i
else:
return self.sub_sentdict.words[i]
if idx and not idx_in_subsentdict:
return self.start
elif idx_in_subsentdict:
return 0
else:
try:
return self.sub_sentdict.words[0]
except IndexError:
return ''
def get_head(self, idx=False, idx_in_subsentdict=False):
"""
@type return: str
"""
for i in range(len(self.sub_sentdict)):
if wc.is_verb(self.sub_sentdict.pos[i]) and not wc.is_aux_lemma(
self.sub_sentdict.lemmas[i]):
if idx:
if idx_in_subsentdict:
return i
else:
return self.start + i
else:
return self.sub_sentdict.words[i]
if idx and not idx_in_subsentdict:
return self.start
elif idx_in_subsentdict:
return 0
else:
try:
return self.sub_sentdict.words[0]
except IndexError:
return ''
def hardt_features(ant, trig, sentences, pos_tags):
"""
This exists to add features that are somewhat based on what Hardt did in 1997.
@type ant: vpe_objects.Antecedent
@type trig: vpe_objects.Auxiliary
@type sentences: vpe_objects.AllSentences
"""
v = []
sent_tree = sentences.get_sentence_tree(ant.sentnum)
ant_sent = sentences.get_sentence(ant.sentnum)
trig_sent = sentences.get_sentence(trig.sentnum)
vp = sentences.nearest_vp(trig)
vp_head = vp.get_head()
vp_head_idx = vp.get_head(idx=True)
ant_head = ant.get_head()
ant_head_idx = ant.get_head(idx=True)
v.append(1.0 if ant == vp else 0.0)
v.append(1.0 if ant_head == vp_head else 0.0)
v.append(1.0 if vp.start <= ant_head_idx <= vp.end else 0.0)
v.append(1.0 if ant.start <= vp_head_idx <= ant.end else 0.0)
v.append(ant.sentnum - vp.sentnum)
v.append(ant.start - vp.start)
v.append(ant.end - vp.end)
# be-do form
try:
v.append(1.0 if wc.is_be(ant_sent.lemmas[ant.start - 1])
or wc.is_be(ant_sent.lemmas[ant.start]) else 0.0)
v.append(1.0 if trig.type == 'do' and v[-1] == 1.0 else 0.0)
except IndexError:
v += [0.0, 0.0]
# quotation features
quote_start_trig, quote_end_trig = None, None
for i, w in enumerate(trig_sent.lemmas):
if w == "\"":
if not quote_start_trig:
quote_start_trig = i
else:
quote_end_trig = i
break
trig_in_quotes = False
if quote_start_trig and quote_end_trig:
trig_in_quotes = quote_start_trig <= trig.wordnum <= quote_end_trig
v.append(1.0 if trig_in_quotes else 0.0)
else:
v.append(0.0)
quote_start_ant, quote_end_ant = None, None
for i, w in enumerate(ant_sent.lemmas):
if w == "\"":
if not quote_start_ant:
quote_start_ant = i
else:
quote_end_ant = i
break
ant_in_quotes = False
if quote_start_ant and quote_end_ant:
ant_in_quotes = quote_start_ant <= ant.start <= quote_end_ant and quote_start_ant <= ant.end <= quote_end_ant
v.append(1.0 if quote_start_ant <= ant.start <= quote_end_ant else 0.0)
v.append(1.0 if quote_start_ant <= ant.end <= quote_end_ant else 0.0)
else:
v += [0.0, 0.0]
v.append(1.0 if trig_in_quotes and ant_in_quotes else 0.0)
# Nielsen features
v.append(1.0 if wc.is_aux_lemma(ant.sub_sentdict.lemmas[0]) else 0.0)
v.append(1.0 if wc.is_aux_lemma(ant.sub_sentdict.lemmas[ant.get_head(
idx=True, idx_in_subsentdict=True)]) else 0.0)
for tag in pos_tags:
v.append(1.0 if tag == ant.sub_sentdict.pos[0] else
0.0) # Sparse encoding of the pos tag of first word in ant
v.append(1.0 if tag == ant.sub_sentdict.pos[-1] else
0.0) # Sparse encoding of the pos tag of last word in ant
v.append(
float(ant.sub_sentdict.pos.count(tag)) /
len(ant.sub_sentdict)) # Frequency of the given pos tag in ant
for fun in [
wc.is_adverb, wc.is_verb, wc.is_adverb, wc.is_noun,
wc.is_preposition, wc.is_punctuation, wc.is_predicative
]:
v.append(1.0 if fun(ant.sub_sentdict.pos[0]) else
0.0) # Sparse encoding of the identity of first word in ant
v.append(1.0 if fun(ant.sub_sentdict.pos[-1]) else
0.0) # Sparse encoding of the identity of last word in ant
v.append(
float(len(map(fun, ant.sub_sentdict.pos))) /
len(ant.sub_sentdict)) # Frequency of the given function in ant
sent_phrases = get_phrases(sent_tree)
ant_phrases = lowest_common_subtree_phrases(sent_tree, ant.get_words())
v.append(float(len(ant_phrases)) / len(sent_phrases))
for phrase in ['NP', 'VP', 'S', 'SINV', 'ADVP', 'ADJP', 'PP']:
v.append(
len(map(lambda s: s.startswith(phrase), ant_phrases)) /
float(len(ant_phrases)))
v.append(
len(map(lambda s: s.startswith(phrase), sent_phrases)) /
float(len(sent_phrases)))
continuation_words = ['than', 'as', 'so']
if ant.sentnum == trig.sentnum:
v.append(1.0)
for word in continuation_words:
v.append(1.0 if word in
ant_sent.words[ant.end:trig.wordnum] else 0.0)
else:
v.append(0.0)
for _ in continuation_words:
v.append(0.0)
try:
v.append(1.0 if ant_sent.words[ant.start - 1] == trig.word else 0.0)
v.append(1.0 if ant_sent.lemmas[ant.start - 1] == trig.lemma else 0.0)
v.append(1.0 if ant_sent.lemmas[ant.start - 1] == trig.type else 0.0)
v.append(1.0 if ant_sent.pos[ant.start - 1] == trig.pos else 0.0)
except IndexError:
v += [0.0, 0.0, 0.0, 0.0]
# Theoretical linguistics features
if ant.sentnum == trig.sentnum:
word_positions = getwordtreepositions(sent_tree)
v.append(1.0)
v.append(1.0 if wc.ccommands(ant.start, trig.wordnum, sent_tree,
word_positions) else 0.0)
v.append(1.0 if wc.ccommands(trig.wordnum, ant.start, sent_tree,
word_positions) else 0.0)
v.append(1.0 if wc.ccommands(ant.end, trig.wordnum, sent_tree,
word_positions) else 0.0)
v.append(1.0 if wc.ccommands(trig.wordnum, ant.
end, sent_tree, word_positions) else 0.0)
# Check if a word in the antecedent c-commands the trig and vice versa.
ant_word_ccommands, trig_ccommands = False, False
for idx in range(ant.start, ant.end + 1):
if wc.ccommands(idx, trig.wordnum, sent_tree, word_positions):
v.append(1.0)
ant_word_ccommands = True
if wc.ccommands(trig.wordnum, idx, sent_tree, word_positions):
v.append(1.0)
trig_ccommands = True
if ant_word_ccommands and trig_ccommands: # speed boost of 0.02ms kek
break
if not ant_word_ccommands:
v.append(0.0)
if not trig_ccommands:
v.append(0.0)
else:
v += [0.0 for _ in range(7)]
return v
