The training data is from https://xueqiu.com and is in Chinese language. 正样本(pos):6873条,负样本(neg):3591条;
模型结构为:原始预训练语言模型后接两层dense;
- 使用albert_chinese_small,参数全调,epoch=10, lr=0.001, 准确率97%;
- 使用albert_chinese_small,只调节原始模型的后面的一层dense层与后接的两层dense,epoch=30, lr=0.004, 准确率92%;