Home

Awesome

PreSumm

This code is for EMNLP 2019 paper Text Summarization with Pretrained Encoders

Updates Jan 22 2020: Now you can Summarize Raw Text Input!. Swith to the dev branch, and use -mode test_text and use -text_src $RAW_SRC.TXT to input your text file. Please still use master branch for normal training and evaluation, dev branch should be only used for test_text mode.

Results on CNN/DailyMail (20/8/2019):

<table class="tg"> <tr> <th class="tg-0pky">Models</th> <th class="tg-0pky">ROUGE-1</th> <th class="tg-0pky">ROUGE-2</th> <th class="tg-0pky">ROUGE-L</th> </tr> <tr> <td class="tg-c3ow" colspan="4">Extractive</td> </tr> <tr> <td class="tg-0pky">TransformerExt</td> <td class="tg-0pky">40.90</td> <td class="tg-0pky">18.02</td> <td class="tg-0pky">37.17</td> </tr> <tr> <td class="tg-0pky">BertSumExt</td> <td class="tg-0pky">43.23</td> <td class="tg-0pky">20.24</td> <td class="tg-0pky">39.63</td> </tr> <tr> <td class="tg-0pky">BertSumExt (large)</td> <td class="tg-0pky">43.85</td> <td class="tg-0pky">20.34</td> <td class="tg-0pky">39.90</td> </tr> <tr> <td class="tg-baqh" colspan="4">Abstractive</td> </tr> <tr> <td class="tg-0lax">TransformerAbs</td> <td class="tg-0lax">40.21</td> <td class="tg-0lax">17.76</td> <td class="tg-0lax">37.09</td> </tr> <tr> <td class="tg-0lax">BertSumAbs</td> <td class="tg-0lax">41.72</td> <td class="tg-0lax">19.39</td> <td class="tg-0lax">38.76</td> </tr> <tr> <td class="tg-0lax">BertSumExtAbs</td> <td class="tg-0lax">42.13</td> <td class="tg-0lax">19.60</td> <td class="tg-0lax">39.18</td> </tr> </table>

Python version: This code is in Python3.6

Package Requirements: torch==1.1.0 pytorch_transformers tensorboardX multiprocess pyrouge

Updates: For encoding a text longer than 512 tokens, for example 800. Set max_pos to 800 during both preprocessing and training.

Some codes are borrowed from ONMT(https://github.com/OpenNMT/OpenNMT-py)

Trained Models

CNN/DM BertExt

CNN/DM BertExtAbs

CNN/DM TransformerAbs

XSum BertExtAbs

System Outputs

CNN/DM and XSum

Data Preparation For XSum

Pre-processed data

Data Preparation For CNN/Dailymail

Option 1: download the processed data

Pre-processed data

unzip the zipfile and put all .pt files into bert_data

Option 2: process the data yourself

Step 1 Download Stories

Download and unzip the stories directories from here for both CNN and Daily Mail. Put all .story files in one directory (e.g. ../raw_stories)

Step 2. Download Stanford CoreNLP

We will need Stanford CoreNLP to tokenize the data. Download it here and unzip it. Then add the following command to your bash_profile:

export CLASSPATH=/path/to/stanford-corenlp-full-2017-06-09/stanford-corenlp-3.8.0.jar

replacing /path/to/ with the path to where you saved the stanford-corenlp-full-2017-06-09 directory.

Step 3. Sentence Splitting and Tokenization

python preprocess.py -mode tokenize -raw_path RAW_PATH -save_path TOKENIZED_PATH

Step 4. Format to Simpler Json Files

python preprocess.py -mode format_to_lines -raw_path RAW_PATH -save_path JSON_PATH -n_cpus 1 -use_bert_basic_tokenizer false -map_path MAP_PATH

Step 5. Format to PyTorch Files

python preprocess.py -mode format_to_bert -raw_path JSON_PATH -save_path BERT_DATA_PATH  -lower -n_cpus 1 -log_file ../logs/preprocess.log

Model Training

First run: For the first time, you should use single-GPU, so the code can download the BERT model. Use -visible_gpus -1, after downloading, you could kill the process and rerun the code with multi-GPUs.

Extractive Setting

python train.py -task ext -mode train -bert_data_path BERT_DATA_PATH -ext_dropout 0.1 -model_path MODEL_PATH -lr 2e-3 -visible_gpus 0,1,2 -report_every 50 -save_checkpoint_steps 1000 -batch_size 3000 -train_steps 50000 -accum_count 2 -log_file ../logs/ext_bert_cnndm -use_interval true -warmup_steps 10000 -max_pos 512

Abstractive Setting

TransformerAbs (baseline)

python train.py -mode train -accum_count 5 -batch_size 300 -bert_data_path BERT_DATA_PATH -dec_dropout 0.1 -log_file ../../logs/cnndm_baseline -lr 0.05 -model_path MODEL_PATH -save_checkpoint_steps 2000 -seed 777 -sep_optim false -train_steps 200000 -use_bert_emb true -use_interval true -warmup_steps 8000  -visible_gpus 0,1,2,3 -max_pos 512 -report_every 50 -enc_hidden_size 512  -enc_layers 6 -enc_ff_size 2048 -enc_dropout 0.1 -dec_layers 6 -dec_hidden_size 512 -dec_ff_size 2048 -encoder baseline -task abs

BertAbs

python train.py  -task abs -mode train -bert_data_path BERT_DATA_PATH -dec_dropout 0.2  -model_path MODEL_PATH -sep_optim true -lr_bert 0.002 -lr_dec 0.2 -save_checkpoint_steps 2000 -batch_size 140 -train_steps 200000 -report_every 50 -accum_count 5 -use_bert_emb true -use_interval true -warmup_steps_bert 20000 -warmup_steps_dec 10000 -max_pos 512 -visible_gpus 0,1,2,3  -log_file ../logs/abs_bert_cnndm

BertExtAbs

python train.py  -task abs -mode train -bert_data_path BERT_DATA_PATH -dec_dropout 0.2  -model_path MODEL_PATH -sep_optim true -lr_bert 0.002 -lr_dec 0.2 -save_checkpoint_steps 2000 -batch_size 140 -train_steps 200000 -report_every 50 -accum_count 5 -use_bert_emb true -use_interval true -warmup_steps_bert 20000 -warmup_steps_dec 10000 -max_pos 512 -visible_gpus 0,1,2,3 -log_file ../logs/abs_bert_cnndm  -load_from_extractive EXT_CKPT   

Model Evaluation

CNN/DM

 python train.py -task abs -mode validate -batch_size 3000 -test_batch_size 500 -bert_data_path BERT_DATA_PATH -log_file ../logs/val_abs_bert_cnndm -model_path MODEL_PATH -sep_optim true -use_interval true -visible_gpus 1 -max_pos 512 -max_length 200 -alpha 0.95 -min_length 50 -result_path ../logs/abs_bert_cnndm 

XSum

 python train.py -task abs -mode validate -batch_size 3000 -test_batch_size 500 -bert_data_path BERT_DATA_PATH -log_file ../logs/val_abs_bert_cnndm -model_path MODEL_PATH -sep_optim true -use_interval true -visible_gpus 1 -max_pos 512 -min_length 20 -max_length 100 -alpha 0.9 -result_path ../logs/abs_bert_cnndm