README.md

BERT-Question_Answering

Tensorflow solution of Question Answering task Using BERT model with Google BERT Embeddings

The Elements Contract Extraction training data(bert/contract)

Try to implement Question Answering work based on google's BERT code!

How to train

1. Download BERT for multilingual model (include Japanese):

https://storage.googleapis.com/bert_models/2018_11_03/multilingual_L-12_H-768_A-12.zip

2. Extract the above zip file and put it into the following path

bert

3. Create dataset for training and evaluating model

You can change the paths contain train and test set

  cd data_tools
  python create_dataset.py '../bert/contract/train.json' '../bert/contract/dev.json'

4. Train model

You can change the "train_file" path and "predict_file" path by your own paths (Default: train_file=bert/contract/train.json, predict_file=bert/contract/dev.json)

  python run_contract_qa.py   \   
                  --vocab_file=bert/bert_based/vocab.txt   \
                  --bert_config_file=bert/bert_based/bert_config.json   \
                  --init_checkpoint=bert/bert_based/bert_model.ckpt   \
                  --do_train=True   \
                  --train_file=[path_to_train_file]   \
                  --do_predict=True   \
                  --predict_file=[path_to_test_file]   \
                  --train_batch_size=8  \
                  --predict_batch_size=8   \
                  --learning_rate=3e-5   \
                  --num_train_epochs=100.0   \
                  --max_seq_length=384   \
                  --doc_stride=128   \
                  --output_dir=bert/contract_output    \
                  --max_query_length=15     \
                  --max_answer_length=70    \
                  --version_2_with_negative=False   \
                  --n_best_size=10

How to evaluate

  python post_processing.py [path_to_test_file] [path_to_n_best_output_json]

How to run API service

If you train your own model, ignore step1, step2

1. Download trained model from the following link:

2. Extract the above zip file and put it into the root path

3. Run API

All model's paramters are in config path (default: config_file_path = config.ini), you can re-configure by yourself.

  python run_contract_service.py [config_file_path] [port]

You have to check the configured port which does exist or not by the following command:

  netstat -anp|grep [port]

reference:

• https://github.com/google-research/bert

Note: Currently, this model is run on 1 GPU GTX and can process one document per minute. In the future work, we can speed up the process by experimenting model with less layers and run on multi-gpu.