StructBERT中文情感分类模型介绍
情感分类任务,通常为输入一段句子或一段话,返回该段话正向/负向的情感极性,在用户评价,观点抽取,意图识别中往往起到重要作用。StructBERT中文情感分类模型是基于bdci、dianping、jd binary、waimai-10k四个数据集(11.5w条数据)训练出来的情感分类模型。由于license权限问题,目前只上传了dianping和jd binary两个数据集。
其他数据集:
- bdci:https://www.datafountain.cn/competitions/310/datasets
- waimai-10k:https://github.com/SophonPlus/ChineseNlpCorpus/tree/master/datasets/waimai_10k
模型描述
模型基于Structbert-tiny-chinese,在bdci、dianping、jd binary、waimai-10k四个数据集(11.5w条数据)上fine-tune得到。
期望模型使用方式以及适用范围
你可以使用StructBERT中文情感分类模型模型,对通用领域的中文情感分类任务进行推理。 输入自然语言文本,模型会给出该文本的情感分类标签(0,1),即(negative, positive)以及相应的概率。
如何使用
在安装完成ModelScope-lib之后即可使用
代码范例
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
semantic_cls = pipeline(Tasks.text_classification, 'damo/nlp_structbert_sentiment-classification_chinese-tiny')
semantic_cls(input='启动的时候很大声音,然后就会听到1.2秒的卡察的声音,类似齿轮摩擦的声音')
微调代码范例
import os.path as osp
from modelscope.trainers import build_trainer
from modelscope.msdatasets import MsDataset
from modelscope.utils.hub import read_config
from modelscope.metainfo import Metrics
model_id = 'damo/nlp_structbert_sentiment-classification_chinese-tiny'
dataset_id = 'jd'
WORK_DIR = 'workspace'
max_epochs = 2
def cfg_modify_fn(cfg):
cfg.train.max_epochs = max_epochs
cfg.train.hooks.append({
'type': 'TextLoggerHook',
'interval': 100
})
cfg.evaluation.metrics = [Metrics.seq_cls_metric]
cfg['dataset'] = {
'train': {
'labels': ['负面', '正面'],
'first_sequence': 'sentence',
'label': 'label',
}
}
return cfg
train_dataset = MsDataset.load(dataset_id, namespace='DAMO_NLP', split='train').to_hf_dataset()
eval_dataset = MsDataset.load(dataset_id, namespace='DAMO_NLP', split='validation').to_hf_dataset()
# remove useless case
train_dataset = train_dataset.filter(lambda x: x["label"] != None and x["sentence"] != None)
eval_dataset = eval_dataset.filter(lambda x: x["label"] != None and x["sentence"] != None)
# map float to index
def map_labels(examples):
map_dict = {0: "负面", 1: "正面"}
examples['label'] = map_dict[int(examples['label'])]
return examples
train_dataset = train_dataset.map(map_labels)
eval_dataset = eval_dataset.map(map_labels)
kwargs = dict(
model=model_id,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
work_dir=WORK_DIR,
cfg_modify_fn=cfg_modify_fn)
trainer = build_trainer(name='nlp-base-trainer', default_args=kwargs)
print('===============================================================')
print('pre-trained model loaded, training started:')
print('===============================================================')
trainer.train()
print('===============================================================')
print('train success.')
print('===============================================================')
for i in range(max_epochs):
eval_results = trainer.evaluate(f'{WORK_DIR}/epoch_{i+1}.pth')
print(f'epoch {i} evaluation result:')
print(eval_results)
print('===============================================================')
print('evaluate success')
print('===============================================================')
模型局限性以及可能的偏差
模型训练数据有限,效果可能存在一定偏差。
训练数据介绍
- bdci:CCF汽车行业用户观点主题及情感识别比赛数据集。
- dianping:请参考关联数据集DAMONLP/yfdianping
- jd binary:请参考关联数据集DAMO_NLP/jd
- waimai-10k:某外卖平台收集的用户评价,正向4000条,负向约8000条
数据来源于https://github.com/CLUEbenchmark/CLUEDatasetSearch
数据评估及结果
数据集 | BDCI2018 | Dianping | JD Binary | Waimai-10k |
---|---|---|---|---|
Accuracy | 0.7076 | 0.7406 | 0.8921 | 0.8968 |
@article{wang2019structbert,
title={Structbert: Incorporating language structures into pre-training for deep language understanding},
author={Wang, Wei and Bi, Bin and Yan, Ming and Wu, Chen and Bao, Zuyi and Xia, Jiangnan and Peng, Liwei and Si, Luo},
journal={arXiv preprint arXiv:1908.04577},
year={2019}
}
评论