def _preprocess(self, corpus: Corpus, log_stats=True):
xs = []
ys = []
answers = []
question2predicates = []
num_right = {'all': .0, '2hop': .0}
num_2hop = .0
is_train = corpus.name == 'train'
for i, sample in enumerate(corpus):
candidates = sample['candidate_answer']
gold_predicates = sample['gold_tuple']
gold_entities = sample['gold_entitys']
answers.append(sample['answer'])
has_right = False
question2predicates.append({
'gold': gold_predicates,
'candidate': []
})
for answer_predicates, feats in candidates.items():
if cmp(answer_predicates, gold_predicates) == 0:
xs.append(feats[2:])
ys.append([1])
question2predicates[-1]['candidate'].append(
answer_predicates)
else:
prop = random()
if prop < 0.5 or not is_train:
xs.append(feats[2:])
ys.append([0])
question2predicates[-1]['candidate'].append(
answer_predicates)
if cmp(answer_predicates, gold_predicates) == 0:
has_right = True
if has_right:
num_right['all'] += 1
if len(gold_predicates) <= 3 and len(gold_entities) == 1:
num_right['2hop'] += 1
if len(gold_predicates) <= 3 and len(gold_entities) == 1:
num_2hop += 1
xs = np.array(xs, dtype=self._dtype)
ys = np.array(ys, dtype=self._dtype)
if log_stats:
print('* For {}'.format(corpus.name))
print('* Recall ratio of single-subject questions: {:.2f}'.format(
num_right['2hop'] / num_2hop))
print('\tand the one of all questions: {:.2f}'.format(
num_right['all'] / len(corpus)))
return xs, ys, question2predicates, answers
def GetData(corpus):
'''为验证集验证模型使用的数据
X : numpy.array, (num_sample,num_feature)
Y : numpy.array, (num_sample,1)
samples : python-list,(num_sample,)
ans : python-list, (num_question,num_answer)
question2sample : python-dict, key:questionindex , value:sampleindexs
'''
X = []
Y = []
samples = []
ans = []
gold_tuples = []
question2sample = {}
sample_index = 0
true_num = 0
hop2_num = 0
hop2_true_num = 0
for i in range(len(corpus)):
candidate_tuples = corpus[i]['candidate_tuples']
gold_tuple = corpus[i]['gold_tuple']
gold_entitys = corpus[i]['gold_entitys']
answer = corpus[i]['answer']
q_sample_indexs = []
for t in candidate_tuples:
features = candidate_tuples[t]
if len(gold_tuple) == len(set(gold_tuple).intersection(set(t))):
X.append([features[2]])
Y.append([1])
else:
X.append([features[2]])
Y.append([0])
samples.append(t)
q_sample_indexs.append(sample_index)
sample_index += 1
ans.append(answer)
gold_tuples.append(gold_tuple)
question2sample[i] = q_sample_indexs
if_true = 0
#判断gold tuple是否包含在候选tuples中
for thistuple in candidate_tuples:
if cmp(thistuple, gold_tuple) == 0:
if_true = 1
break
#判断单实体问题中，可召回的比例
if if_true == 1:
true_num += 1
if len(gold_tuple) <= 3 and len(gold_entitys) == 1:
hop2_true_num += 1
if len(gold_tuple) <= 3 and len(gold_entitys) == 1:
hop2_num += 1
X = np.array(X, dtype='float32')
Y = np.array(Y, dtype='float32')
print('单实体问题中，候选答案可召回的的比例为:%.3f' % (hop2_true_num / hop2_num))
print('候选答案能覆盖标准查询路径的比例为:%.3f' % (true_num / len(corpus)))
return X, Y, samples, ans, gold_tuples, question2sample
def recall_of(predicts_list, gold_list):
num_correct = .0
num_single = .0
for predicts, gold in zip(predicts_list, gold_list):
if len(gold) <= 3:
num_single += 1
for predict, prob in predicts:
if cmp(predict, gold) == 0:
num_correct += 1
break
return num_correct / num_single
def ComputePrecision(gold_tuples, predict_tuples, predict_props):
'''
计算单实体问题中，筛选后候选答案的召回率，float
'''
true_num = 0
one_subject_num = 0
for i in range(len(gold_tuples)):
gold_tuple = gold_tuples[i]
if len(gold_tuple) <= 3:
one_subject_num += 1
for j in range(len(predict_tuples[i])):
predict_tuple = predict_tuples[i][j]
if cmp(predict_tuple, gold_tuple) == 0:
true_num += 1
break
return true_num / one_subject_num
def GetData_train(corpus):
'''
为训练集的候选答案生成逻辑回归训练数据，由于正负例非常不均衡，对于负例进行0.05的采样
'''
X = []
Y = []
true_num = 0
hop2_num = 0
hop2_true_num = 0
for i in range(len(corpus)):
if i >= 1283 and i < 1683:
continue
candidate_tuples = corpus[i]['candidate_tuples'] #字典
gold_tuple = corpus[i]['gold_tuple']
gold_entitys = corpus[i]['gold_entitys']
for t in candidate_tuples:
features = candidate_tuples[t]
if len(gold_tuple) == len(set(gold_tuple).intersection(set(t))):
X.append([features[9][0][1]])
Y.append([1])
else:
prop = random.random()
if prop < 0.5:
X.append([features[9][0][1]])
Y.append([0])
if_true = 0 #判断答案是否召回
for thistuple in candidate_tuples:
if cmp(thistuple, gold_tuple) == 0:
if_true = 1
break
if if_true == 1:
true_num += 1
if len(gold_tuple) <= 3 and len(gold_entitys) == 1:
hop2_true_num += 1
if len(gold_tuple) <= 3 and len(gold_entitys) == 1:
hop2_num += 1
X = np.array(X, dtype='float32')
Y = np.array(Y, dtype='float32')
print('单实体问题中，候选答案可召回的的比例为:%.3f' % (hop2_true_num / hop2_num))
print('候选答案能覆盖标准查询路径的比例为:%.3f' % (true_num / len(corpus)))
return X, Y