zk1127 / iaea

This is the code repository for our paper "Preserve Integrity in Realtime Event Summarization", to appear in Transactions on Knowledge Discovery from Data.

As the limitation of LFS, the Glove pre-trained embedding dataset can be downloaded from Google Drive.

HID_Model

Requirements

• Hardwares: a machine with two Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50GHz, 256 GB main memory and a GeForce RTX 2080 Ti graphics card
• OS: Ubuntu 18.04
• Packages:
  • python 3.6
  • tensorflow 1.13.1-gpu
  • keras 2.2.4
  • numpy 1.16.2

• Train

python HID_train.py

• Get HID pre-trained parameters

python getInconsistentWeight.py

After the above steps, data/inconsistent_weight.npy and data/inconsistent_bias.npy are obtained for use in IAEA-Model.

Note: You can directly use these .npy file we provide in data/ folder to train IAEA-Model.

IAEA_Model

Requirements

• Hardwares: a machine with two Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50GHz, 256 GB main memory and a GeForce RTX 2080 Ti graphics card
• OS: Ubuntu 18.04
• Packages:
  • python 3.5
  • tensorflow 1.2.1-gpu

Note: you can use the command to start a tf1.2.1-gpu docker

docker run -itd --gpus all --name tf1.2 -v /:/workspace tensorflow/tensorflow:1.2.1-gpu-py3

docker exec -it tf1.2 /bin/bash

Data preprocess

python data/IAEA/make_datafiles_testdata.py data/IAEA/twitter_final/train

Run this command to get train dataset, to same to valid and test dataset.

Note:you can skip this step and then use the processed dataset of data/IAEA/finished_files_twitter folder

Extractor

• Train

python main.py --model=selector --mode=train --data_path=data/IAEA/finished_files_twitter/chunked/train_* --vocab_path=data/IAEA/finished_files_twitter/vocab --log_root=log_gpu_selector_lr001 --exp_name=exp_sample --max_art_len=110 --max_sent_len=50 --max_train_iter=1500 --batch_size=5 --save_model_every=500 --lr=0.01 --model_max_to_keep=25

Abstractor

• Train

python main.py --model=rewriter --mode=train --data_path=data/IAEA/finished_files_twitter/chunked/train_* --vocab_path=data/IAEA/finished_files_twitter/vocab --log_root=log_rewriter --exp_name=exp_sample --max_enc_steps=400 --max_dec_steps=100 --batch_size=5 --max_train_iter=5000 --save_model_every=1000 --model_max_to_keep=10 --use_temporal_attention=True --intradecoder=True --rl_training=False

• Add reinforcement learning

python main.py --model=rewriter --mode=train --data_path=data/IAEA/finished_files_twitter/chunked/train_* --vocab_path=data/IAEA/finished_files_twitter/vocab --log_root=log_rewriter --exp_name=exp_sample --batch_size=5 --max_train_iter=1000 --intradecoder=True --use_temporal_attention=True --eta=2.5E-05 --rl_training=True --convert_to_reinforce_model=True --max_enc_steps=400 --max_dec_steps=100 --save_model_every=100 --model_max_to_keep=10

python main.py --model=rewriter --mode=train --data_path=data/IAEA/finished_files_twitter/chunked/train_* --vocab_path=data/IAEA/finished_files_twitter/vocab --log_root=log_rewriter --exp_name=exp_sample --batch_size=5 --max_train_iter=1000 --intradecoder=True --use_temporal_attention=True --eta=2.5E-05 --rl_training=True --max_enc_steps=400 --max_dec_steps=100 --save_model_every=100 --model_max_to_keep=10

• Add coverage mechanism

python main.py --model=rewriter --mode=train --data_path=data/IAEA/finished_files_twitter/chunked/train_* --vocab_path=data/IAEA/finished_files_twitter/vocab --log_root=log_rewriter --exp_name=exp_sample --batch_size=5 --max_train_iter=1000 --intradecoder=True --use_temporal_attention=True --eta=2.5E-05 --rl_training=True --max_enc_steps=400 --max_dec_steps=100 --save_model_every=100 --model_max_to_keep=10 --coverage=True --convert_to_coverage_model=True

python main.py --model=rewriter --mode=train --data_path=data/IAEA/finished_files_twitter/chunked/train_* --vocab_path=data/IAEA/finished_files_twitter/vocab --log_root=log_rewriter --exp_name=exp_sample --batch_size=5 --max_train_iter=1000 --intradecoder=True --use_temporal_attention=True --eta=2.5E-05 --rl_training=True --max_enc_steps=400 --max_dec_steps=100 --save_model_every=100 --model_max_to_keep=10 --coverage=True

End2End

• Train

python main.py --model=end2end --mode=train --data_path=data/IAEA/finished_files_twitter/chunked/train_* --vocab_path=data/IAEA/finished_files_twitter/vocab --log_root=log_gpu_endlr0001 --exp_name=exp_sample --max_enc_steps=800 --max_dec_steps=120 --max_train_iter=10000 --batch_size=5 --use_temporal_attention=True --intradecoder=True --eta=2.5E-05 --max_art_len=110 --max_sent_len=50 --selector_loss_wt=5.0 --inconsistent_loss=True --inconsistent_topk=3 --save_model_every=1000 --model_max_to_keep=20 --rl_training=True --coverage=True --pretrained_selector_path=log_gpu_selector_lr001/selector/exp_sample/train/model.ckpt-500 --pretrained_rewriter_path=log_rewriter/rewriter/exp_sample/train/model.ckpt_cov-7000 --lr=0.001

Decode(output final summary)

python main.py --model=end2end --mode=evalall --data_path=data/IAEA/finished_files_twitter/chunked/test_* --vocab_path=data/IAEA/finished_files_twitter/vocab --log_root=log_gpu_endlr0001 --exp_name=exp_sample --max_enc_steps=800 --max_dec_steps=120 --use_temporal_attention=True --intradecoder=True --eta=2.5E-05 --max_art_len=110 --max_sent_len=50 --decode_method=beam --coverage=True --single_pass=1 --save_pkl=True --save_vis=False --inconsistent_loss=True --inconsistent_topk=3 --eval_method=loss --load_best_eval_model=False --coverage=True --rl_training=True --eval_ckpt_path=log_gpu_endlr0001/end2end/exp_sample/train/model.ckpt_cov-10000

Aknowledgement

The code of IAEA-Model is modified on the basis of unified-summarization and RLSeq2Seq.