CCL2020 第二届“小牛杯”幽默计算——情景喜剧笑点识别
- 交叉验证(五折或十折)
- 模型融合 + 伪标签
- Focalloss + 标签平滑
- 对抗训练(FGM)
- 加权Accuracy + F1选择模型
- bert + 胶囊网络
- bert + 图神经网络
- 加载最优模型,学习率*0.1,再跑几个epoch
- 动态衰减
- F1优化
- EMA 指数滑动平均 [参考]
- 预训练的 Word Embedding 与 Bert 结合
- BERT Post-Training [参考]
- GHM-C loss
- Flooding [参考]
- Speaker + Sentence
- Pre-sentence + Cur-sentence
- Pre-sentence + Cur-sentence + Post-sentence
- Speaker_Pre-sentence + Speaker_Post-sentence
- 欠采样
- 过采样
- 数据增强EDA
- 数据增强UDA
- 数据清洗
- dialogue level作为输入(一个dialogue作为一个batch)
- 多卡并行训练
- 混合精度训练
Thanks goes to these wonderful people (emoji key):
airflier 💻 |
This project follows the all-contributors specification. Contributions of any kind welcome!