• A code for "Unsupervised Abstractive Opinion Summarizationby Generating Sentences with Tree-Structured Topic Guidance", TACL to be published in 2021.

  • Corresponding paper: https://arxiv.org/abs/2106.08007

  • Masaru Isonuma, Juncihiro Mori, Danushka Bollegala, and Ichiro Sakata (The University of Tokyo, University of Liverpool)

• Output examples for all reviews in dev/test splits are avalable at output/yelp and output/amazon.

• Python 3.6

• Run the following script to install required packages.

pip install -r requirements.txt

• Download the raw Yelp data and run the pre-process script to create train, val and test directory following:
  https://github.com/sosuperic/MeanSum

• Download the csv file containing reference summaries from:
  https://s3.us-east-2.amazonaws.com/unsup-sum/summaries_0-200_cleaned.csv
  (distributed in https://github.com/sosuperic/MeanSum)

• Run the following script (you can set any number to n_processes for multiprocessing):

python preprocess_yelp.py \
-n_processes 16 \
-dir_train </path/to/train/dir> \
-dir_val </path/to/val/dir> \
-dir_test </path/to/test/dir> \
-path_ref </path/to/reference.csv>

The preprocessed data will be saved in data/yelp by default.

• Download the raw data from the following URL and unzip it:
  https://abrazinskas.s3-eu-west-1.amazonaws.com/downloads/projects/copycat/data/amazon.zip
  (distributed in https://github.com/abrazinskas/Copycat-abstractive-opinion-summarizer)

• Download dev.csv and test.csv from: https://github.com/abrazinskas/Copycat-abstractive-opinion-summarizer/tree/master/gold_summs

• Run the following script :

python preprocess_amazon.py \
-n_processes 16 \
-dir_train </path/to/train/dir> \
-path_dev </path/to/dev.csv> \
-path_test </path/to/test.csv> 

The preprocessed data will be saved in data/amazon by default.

• Run the following script:

python train.py \
-gpu <index/of/gpu> \
-data <"yelp"/or/"amazon"> \
-n_processes 16

The other arguments and default parameters are defined in configure.py.
The trained parameters will be saved in model by default.

• Run the following script (only single-gpu is available):

python evaluate.py \
-gpu <index/of/gpu> \
-data <"yelp"/or/"amazon"> \
-n_processes 16

You need to set the same arguments as those used for training except for -gpu and -n_processes.
You can also use our checkpoints by downloading them from here.
Put them in model/yelp/recursum-stable or model/amazon/recursum-stable and run the script above with -stable flag.

We would like to acknowledge MeanSum & Copycat authors for providing reference summaries of Yelp & Amazon reviews.

For calculating ROUGE scores, we use the same scripts used in MeanSum (evaluation).
https://github.com/sosuperic/MeanSum/tree/master/evaluation