/Pegasus_with_Longformer_summarization

Primary LanguagePython

Pegasus_with_Longformer_summarization

Description

Pegasus is a large Transformer-based encoder-decoder model with a new pre-training objective which is adapted to abstractive summarization. More specifically, the pre-training objective, called "Gap Sentence Generation (GSG)", consists of masking important sentences from a document and generating these gap-sentences.

On the other hand, the Longformer is a Transformer which replaces the full-attention mechanism (quadratic dependency) with a novel attention mechanism which scales linearly with the input sequence length. Consequently, Longformer can process sequences up to 4,096 tokens long (8 times longer than BERT which is limited to 512 tokens).

This project plugs Longformer's attention mechanism to Pegasus in order to perform abstractive summarization on long documents. The conversion is done in loading_scripts/Pegasus_to_4k.py which enables Pegasus to process sequences up to 4,096 tokens long (rather than 512 tokens). Note that the max_pos parameter can be changed to accept even longer sequences (e.g max_pos=16384). The new Pegasus model is then fine-tuned on BigPatent dataset. To assess the model's performance on long documents, all training examples are filtered such that they have a minimum length of 4000 tokens.

This project was built using HuggingFace's Transformers library. The model is trained using model partitioning (with fairscale) and parallel batch processing on a cluster of 8 GPUs.

How to run the project

To run this project, clone the repo and execute the following commands:

  1. cd Pegasus_with_Longformer_summarization
  2. pip install -r requirements.txt
  3. pip install git+https://github.com/allenai/longformer.git
  4. pip install tokenizers==0.10.3
  5. Comment out import 'SAVE_STATE_WARNING' from torch.optim.lr_scheduler in lib/python3.7/site-packages/transformers/trainer_pt_utils.py
  6. Add with torch.no_grad(): above out[:, 0 : dim // 2] = torch.FloatTensor(np.sin(position_enc[:, 0::2])) in lib/python3.7/site-packages/transformers/modeling_bart.py
  7. python loading_scripts/pegasus_to_4k.py
  8. git clone -b v4.5.1-release https://github.com/huggingface/transformers
  9. cd transformers
  10. pip install -e .
  11. cd .. ; python download_long_Big_Patent_data.py
  12. bash tune.sh

Citation

@article{DBLP:journals/corr/abs-1910-03771,
     author = {Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Jamie Brew},
     title = {HuggingFace's Transformers: State-of-the-art Natural Language Processing},
     journal = {CoRR},
     volume = {abs/1910.03771},
     year = {2019},
     url = {http://arxiv.org/abs/1910.03771},
     archivePrefix = {arXiv},
     eprint = {1910.03771},
     timestamp = {Tue, 02 Jun 2020 12:49:01 +0200},
     biburl = {https://dblp.org/rec/journals/corr/abs-1910-03771.bib},
     bibsource = {dblp computer science bibliography, https://dblp.org}
}

@article{DBLP:journals/corr/abs-2004-05150,
     author = {Iz Beltagy and Matthew E. Peters and Arman Cohan},
     title = {Longformer: The Long-Document Transformer},
     journal = {CoRR},
     volume = {abs/2004.05150},
     year = {2020},
     url = {https://arxiv.org/abs/2004.05150},
     archivePrefix = {arXiv},
     eprint = {2004.05150},
     timestamp = {Tue, 14 Apr 2020 16:40:34 +0200},
     biburl = {https://dblp.org/rec/journals/corr/abs-2004-05150.bib},
     bibsource = {dblp computer science bibliography, https://dblp.org}
}

@article{DBLP:journals/corr/abs-1912-08777,
     author = {Jingqing Zhang and Yao Zhao and Mohammad Saleh and Peter J. Liu},
     title = {{PEGASUS:} Pre-training with Extracted Gap-sentences for Abstractive Summarization},
     journal = {CoRR},
     volume = {abs/1912.08777},
     year = {2019},
     url = {http://arxiv.org/abs/1912.08777},
     archivePrefix = {arXiv},
     eprint = {1912.08777},
     timestamp = {Fri, 03 Jan 2020 16:10:45 +0100},
     biburl = {https://dblp.org/rec/journals/corr/abs-1912-08777.bib},
     bibsource = {dblp computer science bibliography, https://dblp.org}
}