def predict_one(self, sentence):
"""
对输入的句子进行ner识别，取batch中的第一行结果
:param sentence:
:return:
"""
if self.configs.use_bert:
X, y, att_mask, Sentence = self.dataManager.prepare_single_sentence(sentence)
if self.configs.finetune:
model_inputs = (X, att_mask)
else:
model_inputs = self.bert_model(X, attention_mask=att_mask)[0]
else:
X, y, Sentence = self.dataManager.prepare_single_sentence(sentence)
model_inputs = X
inputs_length = tf.math.count_nonzero(X, 1)
logits, log_likelihood, transition_params = self.ner_model(
inputs=model_inputs, inputs_length=inputs_length, targets=y)
label_predicts, _ = crf_decode(logits, transition_params, inputs_length)
label_predicts = label_predicts.numpy()
sentence = Sentence[0, 0:inputs_length[0]]
y_pred = [str(self.dataManager.id2label[val]) for val in label_predicts[0][0:inputs_length[0]]]
if self.configs.use_bert:
# 去掉[CLS]和[SEP]对应的位置
y_pred = y_pred[1:-1]
entities, suffixes, indices = extract_entity(sentence, y_pred, self.dataManager)
return entities, suffixes, indices
def metrics(X, y_true, y_pred, configs, data_manager, tokenizer):
precision = -1.0
recall = -1.0
f1 = -1.0
hit_num = 0
pred_num = 0
true_num = 0
correct_label_num = 0
total_label_num = 0
label_num = {}
label_metrics = {}
measuring_metrics = configs.measuring_metrics
# tensor向量不能直接索引，需要转成numpy
y_pred = y_pred.numpy()
for i in range(len(y_true)):
if configs.use_bert:
x = tokenizer.convert_ids_to_tokens(X[i].tolist(),
skip_special_tokens=True)
else:
x = [
str(data_manager.id2token[val]) for val in X[i]
if val != data_manager.token2id[data_manager.PADDING]
]
y = [
str(data_manager.id2label[val]) for val in y_true[i]
if val != data_manager.label2id[data_manager.PADDING]
]
y_hat = [
str(data_manager.id2label[val]) for val in y_pred[i]
if val != data_manager.label2id[data_manager.PADDING]
] # if val != 5
correct_label_num += len([1 for a, b in zip(y, y_hat) if a == b])
total_label_num += len(y)
true_labels, labeled_labels_true, _ = extract_entity(
x, y, data_manager)
pred_labels, labeled_labels_pred, _ = extract_entity(
x, y_hat, data_manager)
hit_num += len(set(true_labels) & set(pred_labels))
pred_num += len(set(pred_labels))
true_num += len(set(true_labels))
for label in data_manager.suffix:
label_num.setdefault(label, {})
label_num[label].setdefault('hit_num', 0)
label_num[label].setdefault('pred_num', 0)
label_num[label].setdefault('true_num', 0)
true_lab = [
x for (x, y) in zip(true_labels, labeled_labels_true)
if y == label
]
pred_lab = [
x for (x, y) in zip(pred_labels, labeled_labels_pred)
if y == label
]
label_num[label]['hit_num'] += len(set(true_lab) & set(pred_lab))
label_num[label]['pred_num'] += len(set(pred_lab))
label_num[label]['true_num'] += len(set(true_lab))
if total_label_num != 0:
accuracy = 1.0 * correct_label_num / total_label_num
if pred_num != 0:
precision = 1.0 * hit_num / pred_num
if true_num != 0:
recall = 1.0 * hit_num / true_num
if precision > 0 and recall > 0:
f1 = 2.0 * (precision * recall) / (precision + recall)
# 按照字段切分
for label in label_num.keys():
tmp_precision = 0
tmp_recall = 0
tmp_f1 = 0
# 只包括BI
if label_num[label]['pred_num'] != 0:
tmp_precision = 1.0 * label_num[label]['hit_num'] / label_num[
label]['pred_num']
if label_num[label]['true_num'] != 0:
tmp_recall = 1.0 * label_num[label]['hit_num'] / label_num[label][
'true_num']
if tmp_precision > 0 and tmp_recall > 0:
tmp_f1 = 2.0 * (tmp_precision * tmp_recall) / (tmp_precision +
tmp_recall)
label_metrics.setdefault(label, {})
label_metrics[label]['precision'] = tmp_precision
label_metrics[label]['recall'] = tmp_recall
label_metrics[label]['f1'] = tmp_f1
results = {}
for measure in measuring_metrics:
results[measure] = vars()[measure]
return results, label_metrics