Created by: JenningsL

我在 sequence_tagging_for_ner 示例代码的基础上增加了几列特征，但是没有改变模型结构。在训练时，训练集上的 Cost, error, 以及其他指标都有变化，但是测试集上的F1-score 等指标从第一个batch开始就是一个非零值，并且没有任何变化。 [INFO 2017-11-01 19:25:05,183 train.py:103] Pass 9, Batch 0, Cost 11.496690, {'ner_chunk.precision': 0.7249544858932495, 'ner_chunk.F1-score': 0.8158245086669922, 'ner_chunk.recall': 0.9327396154403687, 'error': 0.22525805234909058} [INFO 2017-11-01 19:25:08,022 train.py:107] Test with Pass 9, Batch 0, {'ner_chunk.precision': 0.7743763327598572, 'ner_chunk.F1-score': 0.8488954901695251, 'ner_chunk.recall': 0.9392839074134827, 'error': 0.177949458360672} [INFO 2017-11-01 19:25:36,543 train.py:103] Pass 9, Batch 10, Cost 10.829126, {'ner_chunk.precision': 0.7487720847129822, 'ner_chunk.F1-score': 0.8320754766464233, 'ner_chunk.recall': 0.936234712600708, 'error': 0.20192831754684448} [INFO 2017-11-01 19:25:39,371 train.py:107] Test with Pass 9, Batch 10, {'ner_chunk.precision': 0.7743763327598572, 'ner_chunk.F1-score': 0.8488954901695251, 'ner_chunk.recall': 0.9392839074134827, 'error': 0.177949458360672} [INFO 2017-11-01 19:26:08,431 train.py:103] Pass 9, Batch 20, Cost 9.446608, {'ner_chunk.precision': 0.7925390601158142, 'ner_chunk.F1-score': 0.8627714514732361, 'ner_chunk.recall': 0.9466617703437805, 'error': 0.16465938091278076} [INFO 2017-11-01 19:26:11,258 train.py:107] Test with Pass 9, Batch 20, {'ner_chunk.precision': 0.7743763327598572, 'ner_chunk.F1-score': 0.8488954901695251, 'ner_chunk.recall': 0.9392839074134827, 'error': 0.177949458360672} [INFO 2017-11-01 19:26:33,026 train.py:117] Test with Pass 9, {'ner_chunk.precision': 0.7743763327598572, 'ner_chunk.F1-score': 0.8488954901695251, 'ner_chunk.recall': 0.9392839074134827, 'error': 0.177949458360672} 在调用 infer.py 进行预测时，发现所有标签都预测为 O。