Comparative Opinion Summarization via Collaborative Decoding

This repository contains the dataset, source code and trained model Comparative Opinion Summarization via Collaborative Decoding.

Citations

@inproceedings{iso22aclfindings,
    title = {{C}omparative {O}pinion {S}ummarization via {C}ollaborative {D}ecoding},
    author = {Hayate Iso and
              Xiaolan Wang and
              Stefanos Angelidis and
              Yoshihiko Suhara},
    booktitle = {Findings of the Association for Computational Linguistics (ACL)},
    month = {May},
    year = {2022}
}

Setup

Please use the command below to setup and install requirements.

conda create -n cocosum python=3.8
conda activate cocosum
pip install -r requirements.txt 

CoCoTrip dataset

The script prep.py automatically downloads the public TripAdvisor dataset by Wang+(2010), and builds the CoCoTrip dataset, which includes self-supervised training and evaluation sets.

python prep.py ./data/
head -n 50000 ./data/train_cont_all.jsonl > ./data/train_cont_50k.jsonl 
head -n 1000 ./data/train_comm_pair.jsonl > ./data/train_comm_pair_1k.jsonl

Available models

All models are hosted on huggingface 🤗 model hub (https://huggingface.co/megagonlabs/).

Model name	Task	Training setting
megagonlabs/cocosum-cont-self	Contrastive	Self-supervision
megagonlabs/cocosum-cont-few	Contrastive	Few-shot
megagonlabs/cocosum-comm-self	Common	Self-supervision
megagonlabs/cocosum-comm-few	Common	Few-shot

Usage

You can generate contrastive and common opinion summaries by combining two base models, as referred to Collaborative-decoding. After making the CoCoTrip dataset, you can just run the following to download our pre-trained model from huggingface model hub and generate the summaries.

# Co-Decoding for Contrastive Opinion Summarization
python decode.py \
  ./data/ \
  cont \
  gen/cont/codec/ \  # directory to store generated summaries
  megagonlabs/cocosum-cont-few \  # Target model
  --counter_model_checkpoint megagonlabs/cocosum-cont-few \  # Counterpart model
  --alpha 0.2 \  # hyper-parameter
  --top_p 0.9

# Co-Decoding for Common Opinion Summarization  
python decode.py \
  ./data/ \
  comm \
  gen/comm/codec \
  megagonlabs/cocosum-comm-few \  # Target model
  --counter_model_checkpoint megagonlabs/cocosum-cont-few \  # Counterpart model. The contrastive summarization model is used in this case.
  --alpha 0.4 \  # hyper-parameter
  --top_p 0.9 \
  --do_ens_tgt \
  --do_ens_cnt \
  --ens_method add  # combining the output by summing up

Reproduction

Training

After building CoCoTrip dataset, you can train the base contrastive and common opinion summarization models by running the following commands using self-supervised dataset!!

# Contrastive Summarization Model -- Self-supervision
python train.py \
  ./data/train_cont_50k.jsonl \
  --default_root_dir ./log/cont/self \
  --accumulate_grad_batches 8 \
  --gradient_clip_val 1.0 \
  --max_steps 50000 \
  --warmup 1000 \
  --val_check_interval 5000 \
  --task cont \
  --gpus 1
  
# Common Opinion Summarization -- Self-supervision
python train.py \
  ./data/train_comm_pair_1k.jsonl \
  --default_root_dir ./log/comm/self \
  --accumulate_grad_batches 8 \
  --gradient_clip_val 1.0 \
  --max_steps 5000 \
  --warmup 100 \
  --val_check_interval 500 \
  --task comm \
  --use_pair \
  --gpus 1

If you want to further train the model on top of self-supervised opinion summarization model, you can run the followings:

# Contrastive Summarization Model -- Few-Shot
python train.py \
  ./data/few_cont.jsonl \
  --default_root_dir ./log/cont/few \
  --accumulate_grad_batches 8 \
  --gradient_clip_val 1.0 \
  --max_steps 1000 \
  --warmup 100 \
  --val_check_interval 100 \
  --task cont \
  --ckpt log/cont/self/lightning_logs/version_0/checkpoints/ \
  --gpus 1

# Common Opinion Summarization -- Few-Shot
python train.py \
 ./data/few_comm_pair.jsonl \
 --default_root_dir ./log/comm/few \
 --accumulate_grad_batches 8 \
 --gradient_clip_val 1.0 \
 --max_steps 1000 \
 --warmup 100 \
 --val_check_interval 100 \
 --task comm \
 --use_pair \
 --ckpt log/comm/self/lightning_logs/version_0/checkpoints/ \
 --gpus 1 

Decoding

Once you train the models for both contrastive and common opinion summarizations, it's time to generate summaries using Co-decoding!

# Co-Decoding for Contrastive Opinion Summarization
python decode.py \
  ./data/ \
  cont \
  gen/cont/codec/ \  # directory to store generated summaries
  log/cont/few/lightning_logs/version_0/checkpoints/ \  # Target model
  --counter_model_checkpoint log/cont/few/lightning_logs/version_0/checkpoints/ \  # Counterpart model
  --alpha 0.2 \  # hyper-parameter
  --top_p 0.9

# Co-Decoding for Common Opinion Summarization  
python decode.py \
  ./data/ \
  comm \
  gen/comm/codec \
  ./log/comm/few/lightning_logs/version_0/checkpoints/ \  # Target model
  --counter_model_checkpoint log/cont/few/lightning_logs/version_0/checkpoints/ \  # Counterpart model. The contrastive summarization model is used in this case.
  --alpha 0.4 \  # hyper-parameter
  --top_p 0.9 \
  --do_ens_tgt \
  --do_ens_cnt \
  --ens_method add  # combining the output by summing up

Evaluation

Finally, you can evaluate your generated summaries with all the evaluation metrics by running the following command!

python evaluate.py \ 
  ./data/ \
  ./gen/cont/codec/outputs.json \  # path of the generated contrastive summaries
  ./gen/comm/codec/outputs.json  # path of the generated common summaries

Acknowledgement

The repository is build based on the naacl2021-longdoc-tutorial .

Disclosure

Embedded in, or bundled with, this product are open source software (OSS) components, datasets and other third party components identified below. The license terms respectively governing the datasets and third-party components continue to govern those portions, and you agree to those license terms, which, when applicable, specifically limit any distribution. You may receive a copy of, distribute and/or modify any open source code for the OSS component under the terms of their respective licenses. In the event of conflicts between Megagon Labs, Inc. Recruit Co., Ltd., license conditions and the Open Source Software license conditions, the Open Source Software conditions shall prevail with respect to the Open Source Software portions of the software. You agree not to, and are not permitted to, distribute actual datasets used with the OSS components listed below. You agree and are limited to distribute only links to datasets from known sources by listing them in the datasets overview table below. You are permitted to distribute derived datasets of data sets from known sources by including links to original dataset source in the datasets overview table below. You agree that any right to modify datasets originating from parties other than Megagon Labs, Inc. are governed by the respective third party’s license conditions. All OSS components and datasets are distributed WITHOUT ANY WARRANTY, without even implied warranty such as for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, and without any liability to or claim against any Megagon Labs, Inc. entity other than as explicitly documented in this README document. You agree to cease using any part of the provided materials if you do not agree with the terms or the lack of any warranty herein. While Megagon Labs, Inc., makes commercially reasonable efforts to ensure that citations in this document are complete and accurate, errors may occur. If you see any error or omission, please help us improve this document by sending information to contact_oss@megagon.ai.

All dataset and code used within the product are listed below (including their copyright holders and the license conditions). For Datasets having different portions released under different licenses, please refer to the included source link specified for each of the respective datasets for identifications of dataset files released under the identified licenses.

ID	OSS Component Name	Modified	Copyright Holder	Upstream Link	License
1	Longformer encoder-decoder model	Yes	Hugging Face	link	Apache License 2.0

ID	Dataset	Modified	Copyright Holder	Source Link	License
1	TripAdvisor dataset	Yes	University of Virginia	source