This repository contains the code for the paper Large-Scale and Multi-Perspective Opinion Summarization with Diverse Review Subsets (https://arxiv.org/abs/2310.13340) accepted to EMNLP 2023 Findings. SubSumm is a supervised summarization framework for large-scale multi-perspetcive opinion summarization with limited computational resource.

Please use the command below to setup the environment needed.

conda create -n subsumm python=3.7.0
conda activate subsumm
pip install -r requirements.txt

Then download this version of FairSeq and store it to ./fairseq_lib. In the unzipped directory are the installation instructions.

export PYTHONPATH=./fairseq_lib/.:$PYTHONPATH
export MKL_THREADING_LAYER=GNU

The preprocessed datasets, models fine-tuned for sentiment analysis and information valuation, and checkpoints of SubSumm are available here.

To create the FairSeq formatted datasets, firstly set global variables as shown below.

PROJECT_WORK_DIR=
TASK=

ENCODER_PATH=$PROJECT_WORK_DIR/artifacts/bart/encoder.json
VOCAB_BPE=$PROJECT_WORK_DIR/artifacts/bart/vocab.bpe
DICT_PATH=$PROJECT_WORK_DIR/artifacts/bart/dict.txt

Secondly, run the BPE encoding of sequences:

for SPLIT in train valid test
do
  for LANG in source target
    do
       python -m examples.roberta.multiprocessing_bpe_encoder \
       --encoder-json $ENCODER_PATH \
       --vocab-bpe $VOCAB_BPE \
       --inputs "$TASK/$SPLIT.$LANG" \
       --outputs "$TASK/$SPLIT.bpe.$LANG" \
       --workers 60 \
       --keep-empty;
  done
done

Finally, binarize sequences:

fairseq-preprocess \
  --source-lang "source" \
  --target-lang "target" \
  --trainpref "${TASK}/train.bpe" \
  --validpref "${TASK}/valid.bpe" \
  --destdir "${TASK}-bin/" \
  --workers 60 \
  --srcdict $DICT_PATH \
  --tgtdict $DICT_PATH;

With the fine-tuned models, you can modify the scripts below, then conduct sentiment analysis and information valuation on the review sets for .senti, .dist, and .rank files:

python subsumm/scripts/inference/sentiment_analysis.py

python subsumm/scripts/inference/information_valuation.py --data-path= \
--checkpoint-path= \
--bart-dir=artifacts/bart \
--output-folder-path= \
--split= \
--ndocs=10 

The code for fine-tuning a BART-base encoder for information valuation is also provided in this repo.

The three review sampling strategies in our paper, i.e., Random Sampling, Sentiment-Random Sampling, and Sentiment-Information Ranking can be used by this command:

python subsumm/scripts/inference/sampling_strategies.py --data-path= \
--dataset= \
--target= \
--output-folder-path= \
--split= \
--ndocs=10 \
--batch-size= \
--select-strategy=

The summarizer of SubSumm is initialized with BART-base. Please download it to artifacts/bart.

bash subsumm/scripts/training/train_sentisel.sh

You may adjust the hyper-parameters and the paths in the scripts if needed. The command below is to generate summaries with a checkpoint from stage I:

python subsumm/scripts/inference/gen_summs.py --data-path= \
--bart-dir=artifacts/bart \
--checkpoint-path= \
--output-folder-path= \
--split= \
--batch-size= \
--min-length= \
--length-penalty=

The outputs will be available at artifacts/summs.

Before fine-tuning the checkpoints from stage I, use the modified optimal strategy, i.e., Sentiment-Information Ranking (modified) to prepare the input. The code implementation is the function sentiment_information_sampling.

Since we followed the practice of BRIO to design the training stage II, the modified scripts are available for substitution in this repo.

For comparability, we use the evaluation method of SelSum, including the same wrapper over ROUGE and the coreNLP tokenizer.

@misc{jiang2023largescale,
      title={Large-Scale and Multi-Perspective Opinion Summarization with Diverse Review Subsets}, 
      author={Han Jiang and Rui Wang and Zhihua Wei and Yu Li and Xinpeng Wang},
      year={2023},
      eprint={2310.13340},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

We use the resources from SelSum, BRIO and FairSeq. We sincerely thank the authors for opensourcing their fantastic works.

For details not mentioned in this document, you may move to the above repositories.