def _get_features(self, method='treedlib'):
get_feats = compile_relation_feature_generator()
f_index = defaultdict(list)
for j, cand in enumerate(self._candidates):
for feat in get_feats(cand.root, cand.e1_idxs, cand.e2_idxs):
f_index[feat].append(j)
return f_index
def _get_features(self, method='treedlib'):
get_feats = compile_relation_feature_generator()
f_index = defaultdict(list)
for j,cand in enumerate(self._candidates):
for feat in get_feats(cand.root, cand.e1_idxs, cand.e2_idxs):
f_index[feat].append(j)
return f_index
def get_binary_span_feats(sidxs, sentence, stopwords):
"""Get binary (relation) span features from TreeDLib"""
get_tdl_feats = compile_relation_feature_generator()
xmltree = corenlp_to_xmltree(get_as_dict(sentence))
s1_idxs, s2_idxs = sidxs
if len(s1_idxs) > 0 and len(s2_idxs) > 0:
# Apply TDL features
for f in get_tdl_feats(xmltree.root, s1_idxs, s2_idxs,
stopwords=stopwords):
yield 'TDL_' + f, 1
def get_span_feats(candidate):
args = candidate.get_arguments()
if not isinstance(args[0], Span):
raise ValueError("Accepts Span-type arguments, %s-type found.")
# Unary candidates
if len(args) == 1:
get_tdl_feats = compile_entity_feature_generator()
span = args[0]
sent = get_as_dict(span.parent)
xmltree = corenlp_to_xmltree(sent)
sidxs = range(span.get_word_start(), span.get_word_end() + 1)
if len(sidxs) > 0:
# Add DDLIB entity features
for f in get_ddlib_feats(sent, sidxs):
yield 'DDL_' + f, 1
# Add TreeDLib entity features
for f in get_tdl_feats(xmltree.root, sidxs):
yield 'TDL_' + f, 1
# Binary candidates
elif len(args) == 2:
get_tdl_feats = compile_relation_feature_generator()
span1, span2 = args
xmltree = corenlp_to_xmltree(get_as_dict(span1.parent))
s1_idxs = range(span1.get_word_start(), span1.get_word_end() + 1)
s2_idxs = range(span2.get_word_start(), span2.get_word_end() + 1)
if len(s1_idxs) > 0 and len(s2_idxs) > 0:
# Apply TDL features
for f in get_tdl_feats(xmltree.root, s1_idxs, s2_idxs):
yield 'TDL_' + f, 1
else:
raise NotImplementedError(
"Only handles unary and binary candidates currently")
def get_content_feats(candidates):
candidates = candidates if isinstance(candidates, list) else [candidates]
for candidate in candidates:
args = tuple([m.context for m in candidate.get_mentions()])
if not (isinstance(args[0], TemporarySpanMention)):
raise ValueError(
f"Accepts Span-type arguments, {type(candidate)}-type found.")
# Unary candidates
if len(args) == 1:
span = args[0]
if span.sentence.is_lingual():
get_tdl_feats = compile_entity_feature_generator()
sent = get_as_dict(span.sentence)
xmltree = corenlp_to_xmltree(sent)
sidxs = list(
range(span.get_word_start_index(),
span.get_word_end_index() + 1))
if len(sidxs) > 0:
# Add DDLIB entity features
for f in get_ddlib_feats(span, sent, sidxs):
yield candidate.id, f"DDL_{f}", DEF_VALUE
# Add TreeDLib entity features
if span.stable_id not in unary_tdl_feats:
unary_tdl_feats[span.stable_id] = set()
for f in get_tdl_feats(xmltree.root, sidxs):
unary_tdl_feats[span.stable_id].add(f)
for f in unary_tdl_feats[span.stable_id]:
yield candidate.id, f"TDL_{f}", DEF_VALUE
else:
for f in get_word_feats(span):
yield candidate.id, f"BASIC_{f}", DEF_VALUE
# Binary candidates
elif len(args) == 2:
span1, span2 = args
if span1.sentence.is_lingual() and span2.sentence.is_lingual():
get_tdl_feats = compile_relation_feature_generator()
sent1 = get_as_dict(span1.sentence)
sent2 = get_as_dict(span2.sentence)
xmltree = corenlp_to_xmltree(get_as_dict(span1.sentence))
s1_idxs = list(
range(span1.get_word_start_index(),
span1.get_word_end_index() + 1))
s2_idxs = list(
range(span2.get_word_start_index(),
span2.get_word_end_index() + 1))
if len(s1_idxs) > 0 and len(s2_idxs) > 0:
# Add DDLIB entity features for relation
for f in get_ddlib_feats(span1, sent1, s1_idxs):
yield candidate.id, f"DDL_e1_{f}", DEF_VALUE
for f in get_ddlib_feats(span2, sent2, s2_idxs):
yield candidate.id, f"DDL_e2_{f}", DEF_VALUE
# Add TreeDLib relation features
if candidate.id not in binary_tdl_feats:
binary_tdl_feats[candidate.id] = set()
for f in get_tdl_feats(xmltree.root, s1_idxs, s2_idxs):
binary_tdl_feats[candidate.id].add(f)
for f in binary_tdl_feats[candidate.id]:
yield candidate.id, f"TDL_{f}", DEF_VALUE
else:
for f in get_word_feats(span1):
yield candidate.id, f"BASIC_e1_{f}", DEF_VALUE
for f in get_word_feats(span2):
yield candidate.id, f"BASIC_e2_{f}", DEF_VALUE
else:
raise NotImplementedError(
"Only handles unary and binary candidates currently")
def extract_textual_features(
candidates: Union[Candidate, List[Candidate]],
) -> Iterator[Tuple[int, str, int]]:
"""Extract textual features.
:param candidates: A list of candidates to extract features from
"""
candidates = candidates if isinstance(candidates, list) else [candidates]
for candidate in candidates:
args = tuple([m.context for m in candidate.get_mentions()])
if not (isinstance(args[0], (SpanMention, ImplicitSpanMention))):
raise ValueError(
f"Accepts Span/ImplicitSpan-type mentions, {type(args[0])}-type found."
)
# Unary candidates
if len(args) == 1:
span: Union[SpanMention, ImplicitSpanMention] = args[0]
if span.sentence.is_lingual():
get_tdl_feats = _compile_entity_feature_generator()
xmltree = corenlp_to_xmltree(span.sentence)
sidxs = list(
range(span.get_word_start_index(),
span.get_word_end_index() + 1))
if len(sidxs) > 0:
# Add DDLIB entity features
for f in _get_ddlib_feats(span, get_as_dict(span.sentence),
sidxs):
yield candidate.id, f"DDL_{f}", DEF_VALUE
# Add TreeDLib entity features
if span.stable_id not in unary_tdl_feats:
unary_tdl_feats[span.stable_id] = set()
for f in get_tdl_feats(xmltree.root, sidxs):
unary_tdl_feats[span.stable_id].add(f)
for f in unary_tdl_feats[span.stable_id]:
yield candidate.id, f"TDL_{f}", DEF_VALUE
for f in _get_word_feats(span):
yield candidate.id, f"BASIC_{f}", DEF_VALUE
# Multinary candidates
else:
spans = args
if all([span.sentence.is_lingual() for span in spans]):
get_tdl_feats = compile_relation_feature_generator(
is_multary=True)
sents = [get_as_dict(span.sentence) for span in spans]
xmltree = corenlp_to_xmltree(spans[0].sentence)
s_idxs = [
list(
range(span.get_word_start_index(),
span.get_word_end_index() + 1)) for span in spans
]
if all([len(s_idx) > 0 for s_idx in s_idxs]):
# Add DDLIB entity features for relation
for span, sent, s_idx, i in zip(spans, sents, s_idxs,
range(len(spans))):
for f in _get_ddlib_feats(span, sent, s_idx):
yield candidate.id, f"DDL_e{i}_{f}", DEF_VALUE
# Add TreeDLib relation features
if candidate.id not in multinary_tdl_feats:
multinary_tdl_feats[candidate.id] = set()
for f in get_tdl_feats(xmltree.root, s_idxs):
multinary_tdl_feats[candidate.id].add(f)
for f in multinary_tdl_feats[candidate.id]:
yield candidate.id, f"TDL_{f}", DEF_VALUE
for i, span in enumerate(spans):
for f in _get_word_feats(span):
yield candidate.id, f"BASIC_e{i}_{f}", DEF_VALUE
def extract_textual_features(
candidates: Union[Candidate, List[Candidate]],
) -> Iterator[Tuple[int, str, int]]:
"""Extract textual features.
:param candidates: A list of candidates to extract features from
:type candidates: list
"""
candidates = candidates if isinstance(candidates, list) else [candidates]
for candidate in candidates:
args = tuple([m.context for m in candidate.get_mentions()])
if not (isinstance(args[0], (SpanMention, ImplicitSpanMention))):
raise ValueError(
f"Accepts Span/ImplicitSpan-type mentions, {type(args[0])}-type found."
)
# Unary candidates
if len(args) == 1:
span: Union[SpanMention, ImplicitSpanMention] = args[0]
if span.sentence.is_lingual():
get_tdl_feats = _compile_entity_feature_generator()
xmltree = corenlp_to_xmltree(span.sentence)
sidxs = list(
range(span.get_word_start_index(), span.get_word_end_index() + 1)
)
if len(sidxs) > 0:
# Add DDLIB entity features
for f in _get_ddlib_feats(span, get_as_dict(span.sentence), sidxs):
yield candidate.id, f"DDL_{f}", DEF_VALUE
# Add TreeDLib entity features
if span.stable_id not in unary_tdl_feats:
unary_tdl_feats[span.stable_id] = set()
for f in get_tdl_feats(xmltree.root, sidxs):
unary_tdl_feats[span.stable_id].add(f)
for f in unary_tdl_feats[span.stable_id]:
yield candidate.id, f"TDL_{f}", DEF_VALUE
for f in _get_word_feats(span):
yield candidate.id, f"BASIC_{f}", DEF_VALUE
# Binary candidates
elif len(args) == 2:
span1, span2 = args
if span1.sentence.is_lingual() and span2.sentence.is_lingual():
get_tdl_feats = compile_relation_feature_generator()
sent1 = get_as_dict(span1.sentence)
sent2 = get_as_dict(span2.sentence)
xmltree = corenlp_to_xmltree(span1.sentence)
s1_idxs = list(
range(span1.get_word_start_index(), span1.get_word_end_index() + 1)
)
s2_idxs = list(
range(span2.get_word_start_index(), span2.get_word_end_index() + 1)
)
if len(s1_idxs) > 0 and len(s2_idxs) > 0:
# Add DDLIB entity features for relation
for f in _get_ddlib_feats(span1, sent1, s1_idxs):
yield candidate.id, f"DDL_e1_{f}", DEF_VALUE
for f in _get_ddlib_feats(span2, sent2, s2_idxs):
yield candidate.id, f"DDL_e2_{f}", DEF_VALUE
# Add TreeDLib relation features
if candidate.id not in binary_tdl_feats:
binary_tdl_feats[candidate.id] = set()
for f in get_tdl_feats(xmltree.root, s1_idxs, s2_idxs):
binary_tdl_feats[candidate.id].add(f)
for f in binary_tdl_feats[candidate.id]:
yield candidate.id, f"TDL_{f}", DEF_VALUE
for f in _get_word_feats(span1):
yield candidate.id, f"BASIC_e1_{f}", DEF_VALUE
for f in _get_word_feats(span2):
yield candidate.id, f"BASIC_e2_{f}", DEF_VALUE
else:
raise NotImplementedError(
"Only handles unary and binary candidates currently"
)
for i,model in enumerate(roc_models):
plt.plot(model[0],model[1], color=color_vals[i],
lw=2, label='%s (area = %0.2f)' % (model_names[i],model[2]))
plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('ROC Curve')
plt.legend(loc="lower right")
plt.show()
# # Parse Tree Visualization
# In[ ]:
cand = session.query(Candidate).filter(Candidate.id == 19885).one()
print cand
xmltree = corenlp_to_xmltree(get_as_dict(cand.get_parent()))
xmltree.render_tree(highlight=[range(cand[0].get_word_start(), cand[0].get_word_end() + 1), range(cand[1].get_word_start(), cand[1].get_word_end()+1)])
# In[ ]:
get_tdl_feats = compile_relation_feature_generator()
sids = [range(a.get_word_start(), a.get_word_end() + 1) for a in cand.get_contexts()]
tags = list(get_tdl_feats(xmltree.root, sids[0], sids[1]))
print tags