版权声明:本文为博主原创文章,未经博主允许不得转载。 https://blog.csdn.net/u014665013/article/details/83309191
baseline文件:
run_rc_model_att.py
run_shuffle_everyepoch.py
调参策略:
- baseline: 每个epoch都shuffle
run_shuffle_everyepoch.py
( epoch 3 71.62)run_shuffle_everyepoch2.py
(71.61) - 添加tensorboard观察loss
- 修改dropout的值
run_shuffle_everyepoch_dropout_0_5.py
run_shuffle_everyepoch_dropout_0_5.2.py
+0.3(71.9)
run_shuffle_everyepoch_dropout_0_9.py
(epoch 5 71.34)run_shuffle_everyepoch_dropout_0_5.py
(epoch 5 71.15) - query 和 passage长度调整
- 字向量卷积filter 个数和height
- hidden size 调整 目前是150 base里面是100
run_shuffle_everyepoch_hidden_100.log
(epoch 71.61)run_shuffle_everyepoch_hidden_100.2.log
(epoch 71.46) - 修改最后几层的dense网络,尝试激活函数
run_shuffle_everyepoch_dense_decrease.py
run_shuffle_everyepoch_dense_decrease.2.py
基本无效 - 设置lstm 和cnn是否共享参数
- cos 学习率衰减
run_shuffle_everyepoch_summary.log
(epoch 5 71.91)run_shuffle_everyepoch_summary.2.log
(epoch 71.84) - 停用词试验:
- 简单去除标点符号
- 去除标点和 的 了 吗 这样的词 语气词
run_shuffle_everyepoch_stopwords.log
(epoch 5 71.24)run_shuffle_everyepoch_stopwords.2.log
(epoch 4 71.47) - 使用停用词表
- 分词,使用百度的分词方法,使用jieba分词配合该词表 run_shuffle_everyepoch_userdict.log (epoch 4 71.55)
run_shuffle_everyepoch_userdict.2.log
(epoch 4 71.13) - highway_network cnn卷积窗口1
run_shuffle_everyepoch_height_1.log
(epoch 4 71.89)run_shuffle_everyepoch_height_1.2.log
(epoch 71.69) - word embedding:
- word embedding 初始化非0
run_shuffle_everyepoch_rank_initial_embedding.log
(epoch 5 71.51)run_shuffle_everyepoch_rank_initial_embedding.1.log
(epoch 71.22)run_shuffle_everyepoch_rank_initial_embedding.2.log
(epoch 4 70.68 )run_shuffle_everyepoch_rank_initial_embedding.3.log
(epoch 4 71.30) - 观点态度词分词字典制作
- 使用sg公司分词模型进行分词
- embedding dropout
run_shuffle_everyepoch_dropout_embedding.log
(epoch 6 71.94)run_shuffle_everyepoch_dropout_embedding.2.log
(epoch 4 72.16 epoch 5 72.11) - embedding 后面添加可以trainable的向量
run_shuffle_everyepoch_add_word_embedding_trainable.py
(epoch 4 70.66)run_shuffle_everyepoch_add_word_embedding_trainable.2.log
(epoch 3 70.59) - 对最好的模型再设置为trainable看结果
- 使用100维的词向量和字向量,分开使用
run_shuffle_everyepoch_char100_word100.log
(epoch 6 69.28)run_shuffle_everyepoch_char100_word100.2.log
(epoch 7 69.09)
- word embedding 初始化非0
- normalization:
- qanet layer normalization
- qanet unshared layer normalization
- lcn layer normalization
- lcn unshared layer normalizaion
- 重新train词向量:
- 增加单词 数字
- 添加情感词典到词语后面情感词典
- 数据增强:
- 用cb师兄增强数据方案 1
- 调整句子顺序 2
- 使用baseline里面的embedding的方式
- 在embedding后面添加态度词词向
- 单模型训练区分某一类,比如对不确定像单模型预测,最终进行模型融合
- dropout 和batchnormal
- 修改attention部分,使用baseline 里面的attention
- 修改为常规动态修改学习率 3
融合以上提分点结果: 4