The implementation of Hierarchical Attention Networks for Document Classification
requirements: pytorch pyltp jieba
The documents and sentences can be any length, I do not cut them off.
train:
The format of train and val data is csv, each row is a sample.
The format is as follows:
label '\t' token1 token2 ... tokenN
The label and text are string and split by '\t'.
The text must be tokenized, and split by space.
(训练和验证数据都是csv格式，每一行为一个样本，格式如下：样本的label和text用\t分隔。针对text，事先分句，用空格分隔。 具体可见train_data中的csv文件。)
In config.py, set the parameter mode to 'train',
and run:
python main.py
(在config.py中，将参数mode设置为'train',
执行：
python main.py即可)
dev：
In config.py, set the parameter mode to 'eval',
and run:
python main.py
The dev file is a txt file in test_data, it can be a normal document.
(在config.py中，将参数mode设置为'eval', 参考dev_data_file为制定的csv文件
执行： python main.py即可)
Experiments:
The train data has 40k and 4 classes. It achieves 91.83% accuracy. I have not adjusted parameters systematically.
(实验： 训练数据共40K，分4类，每类10k左右，在验证集上的准确率在91.83%，由于时间有限，还未系统地调参。)