def make_contract_value_attention_vectors(self, subdoc):
sumphrase_attention_vector = max_exclusive_pattern_by_prefix(
subdoc.distances_per_pattern_dict, '_phrase')
sumphrase_attention_vector = momentum(sumphrase_attention_vector, 0.99)
value_attention_vector, _c1 = rectifyed_sum_by_pattern_prefix(
subdoc.distances_per_pattern_dict, '_sum.work', relu_th=0.4)
value_attention_vector = cut_above(value_attention_vector, 1)
value_attention_vector = relu(value_attention_vector, 0.6)
value_attention_vector = momentum(value_attention_vector, 0.8)
novalue_attention_vector = max_exclusive_pattern_by_prefix(
subdoc.distances_per_pattern_dict, 'sum_neg')
novalue_attention_vector_local_contrast = relu(
novalue_attention_vector, 0.6)
novalue_attention_vector_local_contrast = momentum(
novalue_attention_vector_local_contrast, 0.9)
value_attention_vector_tuned = (value_attention_vector -
novalue_attention_vector * 0.7)
value_attention_vector_tuned = (value_attention_vector_tuned +
sumphrase_attention_vector) / 2
value_attention_vector_tuned = relu(value_attention_vector_tuned, 0.2)
return {
'sumphrase_attention_vector': sumphrase_attention_vector,
'value_attention_vector': value_attention_vector,
'novalue_attention_vector': novalue_attention_vector,
'novalue_attention_vector_local_contrast':
novalue_attention_vector_local_contrast,
'value_attention_vector_tuned': value_attention_vector_tuned,
}
def make_constraints_attention_vectors(subdoc):
# TODO: move to notebook, too much tuning
value_attention_vector, _c1 = rectifyed_sum_by_pattern_prefix(
subdoc.distances_per_pattern_dict, 'sum_max', relu_th=0.4)
value_attention_vector = cut_above(value_attention_vector, 1)
value_attention_vector = relu(value_attention_vector, 0.6)
value_attention_vector = momentum(value_attention_vector, 0.7)
deal_attention_vector, _c2 = rectifyed_sum_by_pattern_prefix(
subdoc.distances_per_pattern_dict, 'd_order', relu_th=0.5)
deal_attention_vector = cut_above(deal_attention_vector, 1)
deal_attention_vector = momentum(deal_attention_vector, 0.993)
margin_attention_vector, _c3 = rectifyed_sum_by_pattern_prefix(
subdoc.distances_per_pattern_dict, 'sum__', relu_th=0.5)
margin_attention_vector = cut_above(margin_attention_vector, 1)
margin_attention_vector = momentum(margin_attention_vector, 0.95)
margin_attention_vector = relu(margin_attention_vector, 0.65)
margin_value_attention_vector = relu(
(margin_attention_vector + value_attention_vector) / 2, 0.6)
deal_value_attention_vector = (deal_attention_vector +
margin_value_attention_vector) / 2
deal_value_attention_vector = relu(deal_value_attention_vector, 0.75)
return {
'value_attention_vector': value_attention_vector,
'deal_attention_vector': deal_attention_vector,
'margin_attention_vector': margin_attention_vector,
'margin_value_attention_vector': margin_value_attention_vector,
'deal_value_attention_vector': deal_value_attention_vector
}
def _find_charter_section_start(self, doc, headline_pattern_prefix, headlines_attention_vector, additional_attention):
assert headlines_attention_vector is not None
vectors = filter_values_by_key_prefix(doc.distances_per_pattern_dict, headline_pattern_prefix)
# v = rectifyed_sum(vectors, 0.3)
v = max_exclusive_pattern(vectors)
v = relu(v, 0.6)
if additional_attention is not None:
additional_attention_s = smooth_safe(additional_attention, 6)
v += additional_attention_s
# v, _ = improve_attention_vector(doc.embeddings, v, relu_th=0.1)
v *= (headlines_attention_vector + 0.1)
if max(v) > 0.75:
v, _ = improve_attention_vector(doc.embeddings, v, relu_th=0.0)
doc.distances_per_pattern_dict["ha$." + headline_pattern_prefix] = v
# span = 100
best_id = np.argmax(v)
# dia = slice(max(0, best_id - span), min(best_id + span, len(v)))
bounds = get_sentence_bounds_at_index(best_id, doc.tokens)
confidence = v[best_id]
return bounds, confidence, v
def improve_attention_vector(embeddings, vv, relu_th=0.5, mix=1): assert vv is not None meta_pattern, meta_pattern_confidence, best_id = make_smart_meta_click_pattern(vv, embeddings) meta_pattern_attention_v = make_pattern_attention_vector(meta_pattern, embeddings) meta_pattern_attention_v = relu(meta_pattern_attention_v, relu_th) meta_pattern_attention_v = meta_pattern_attention_v * mix + vv * (1.0 - mix) return meta_pattern_attention_v, best_id
def normalize_headline_attention_vector(self, headline_attention_vector_pure): # XXX: test it # _max_head_threshold = max(headline_attention_vector_pure) * 0.75 _max_head_threshold = 1 # max(headline_attention_vector_pure) * 0.75 # XXX: test it # print(_max_head) headline_attention_vector = cut_above(headline_attention_vector_pure, _max_head_threshold) # headline_attention_vector /= 2 # 5 is the maximum points a headline may gain during headlne detection : TODO: return relu(headline_attention_vector)
def make_subj_attention_vectors(self, subdoc, subj_types_prefixes):
r = {}
for subj_types_prefix in subj_types_prefixes:
attention_vector = max_exclusive_pattern_by_prefix(
subdoc.distances_per_pattern_dict, subj_types_prefix)
attention_vector_l = relu(attention_vector, 0.6)
r[subj_types_prefix + 'attention_vector'] = attention_vector
r[subj_types_prefix + 'attention_vector_l'] = attention_vector_l
return r
def make_subject_attention_vector_3(self,
section,
subject_kind: ContractSubject,
addon=None) -> List[float]:
from ml_tools import max_exclusive_pattern
pattern_prefix, attention_vector_name, attention_vector_name_soft = self.__sub_attention_names(
subject_kind)
vectors = filter_values_by_key_prefix(
section.distances_per_pattern_dict, pattern_prefix)
x = max_exclusive_pattern(vectors)
section.distances_per_pattern_dict[attention_vector_name_soft] = x
section.distances_per_pattern_dict[attention_vector_name] = x
# x = x-np.mean(x)
x = relu(x, 0.6)
return x