This code is for paper Hierarchical Transformers for Multi-Document Summarization in ACL2019
Python version: This code is in Python3.6
Package Requirements: pytorch tensorboardX pyrouge
Some codes are borrowed from ONMT(https://github.com/OpenNMT/OpenNMT-py)
Trained Model: Please use this url for a trained model https://drive.google.com/open?id=1suMaf7yBZPSBtaJKLnPyQ4uZNcm0VgDz
The raw version of the WikiSum dataset can be built following steps in https://github.com/tensorflow/tensor2tensor/tree/5acf4a44cc2cbe91cd788734075376af0f8dd3f4/tensor2tensor/data_generators/wikisum
Since the raw version of the WikiSum dataset is huge (~300GB), we only provide the ranked version of the dataset: https://drive.google.com/open?id=1iiPkkZHE9mLhbF5b39y3ua5MJhgGM-5u
- The dataset is ranked by the ranker as described in the paper.
- The top-40 paragraphs for each instance is extracted.
- Extract the
.ptfiles into one directory.
The sentencepiece vocab file is at: https://drive.google.com/open?id=1zbUILFWrgIeLEzSP93Mz5JkdVwVPY9Bf
To train with the default setting as in the paper:
python train_abstractive.py -data_path DATA_PATH/WIKI -mode train -batch_size 10500 -seed 666 -train_steps 1000000 -save_checkpoint_steps 5000 -report_every 100 -trunc_tgt_ntoken 400 -trunc_src_nblock 24 -visible_gpus 0,1,2,3 -gpu_ranks 0,1,2,3 -world_size 4 -accum_count 4 -dec_dropout 0.1 -enc_dropout 0.1 -label_smoothing 0.1 -vocab_path VOCAB_PATH -model_path MODEL_PATH -accum_count 4 -log_file LOG_PATH -inter_layers 6,7 -inter_heads 8 -hier
- DATA_PATH is where you put the
.ptfiles - VOCAB_PATH is where you put the sentencepiece model file
- MODEL_PATH is the directory you want to store the checkpoints
- LOG_PATH is the file you want to write training logs in
To evaluated the trained model:
python train_abstractive.py -data_path DATA_PATH/WIKI -mode validate -batch_size 30000 -valid_batch_size 7500 -seed 666 -trunc_tgt_ntoken 400 -trunc_src_nblock 40 -visible_gpus 0 -gpu_ranks 0 -vocab_path VOCAB_PATH -model_path MODEL_PATH -log_file LOG_PATH -inter_layers 6,7 -inter_heads 8 -hier -report_rouge -max_wiki 100000 -dataset test -alpha 0.4 -max_length 400
This will load all saved checkpoints in the model_path and calculate validation losses (this could take a long time). Then it will select the top checkpoints and generate real summaries. Rouge scores will be calculated.