EventExtraction

Extract Service-related Event From Social Media News.

Usage

Preparation

Clone this repository

git clone --recursive http://192.168.1.104:12345/serviceecosystem/eventextraction.git

download albert pre-trained model

wget https://storage.googleapis.com/albert_zh/albert_tiny_489k.zip

Note: This is albert tiny. For more pre-trained albert model, please visit https://github.com/brightmart/albert_zh

Format your file

If your data is news data like this:

<title> \t timestamp
鲜丰水果完成红杉领投B轮融资，已拥有1100+家全国门店	2018-01-22

Simply run the preprocess script.

python src/preprocess.py

After this you need to split your data into train and test(dev) set. You will get a json file like below:

{
"text": "<sentence1>,<sentence2>,<timestamp>",
"labels": [[0, 1, "LOC"], [2, 3, "ORG"], [15, 25, "Cause"]]
}
{
"text": "<sentence1>,<timestamp>",
"labels": [[0, 1, "LOC"], [2, 3, "ORG"], [14, 24, "Cause"]]
}

Note:

You must attach the timestamp (if you do not need it, just treat it as a placeholder)
I have label the sentence pair relationship on the timestamp. (Don't worry, input_fn will process this problem)

Train joint learning model

cd src/albert_zh
python joint_learning.py --do_train true \
    --data_dir your/path/to/data/directory/ \
    --vocab_file your/path/to/vocab.txt  \
    --bert_config_file your/path/to/albert_config_tiny.json \
    --max_seq_length 128 \
    --train_batch_size 32 \
    --learning_rate 1e-5 \
    --num_train_epochs 3 \
    --init_checkpoint your/path/to/albert_model.ckpt \
    --output_dir your/path/to/output_dir/

Normally you do not need modify the max_predictions_per_seq, train_batch_size, learning_rate.

Using the model to do prediction or evaluation

Do Predict

cd src/albert_zh
python joint_learning.py --do_predict true \
    --data_dir your/path/to/data_dir/ \
    --vocab_file your/path/vocab.txt \
    --bert_config_file your/path/to/albert_config_tiny.json \
    --max_seq_length 128 \ 
    --output_dir your/path/to/output_dir/ \
    --predict_batch_size 32

If you need to do evaluation just change --do_predict to --do_eval.

How to Use Other Pre-training models (BERT, RoBERT)?:

I use the albert as pre-trained model in my joint learning model. Because ALBERT is much smaller than BERT. So that it much faster to train/predict/eval.

This code can be easily modified to suit BERT or RoBERT. Only need to change some import in src/albert_zh/joint_learning.py.

What should I do if I only want to do NER/Text-Pair-Classification?

You can modify the souce code. In src/albert_zh/joint_learning.py, you can find

total_loss = task1_loss + task2_loss

where task1_loss is the NER job loss and task2_loss is the Text-Pair-Classification loss. You can modify the loss weight to control which job is more important.

Specifically, when you set task1_loss weight to 0, then this only do text pair classification task.

Can I use this model to do single sentence classification?

Of course you can.

What you need to do is change get_labelrs. and make sure your each of your input sentence doesn't contain ,.

How can I use custom NER-tag and sentence-pair-label?

You can modify get_labeles and get_labelrs respectively.

icecity96/eventextraction