jiminsun / pointer-generator

Pytorch implementation of the ACL paper 'Get To The Point: Summarization with Pointer-Generator Networks (See et al., 2017)', adapted to a Korean dataset

Pointer-Generator Network

This repository contains the Pytorch implementation of the Pointer-Generator Network for text summarization, presented in Get To The Point: Summarization with Pointer-Generator Networks (See et al., 2017).

While the original paper trains the model on an English dataset, this project aims at building a Korean summarization model. Thus, we additionally incorporate Korean preprocessing & tokenization techniques to adapt the model to Korean.

Most of the code is implemented from scratch, but we also referred to the following repositories. Any direct references are mentioned explicitly on the corresponding lines of code.

• https://github.com/abisee/pointer-generator - the original author's implementation in tensorflow
• https://github.com/atulkum/pointer_summarizer
• https://github.com/rohithreddy024/Text-Summarizer-Pytorch

Note that the overall pipeline relies on pytorch-lightning.

Requirements

Packages

torch==1.5.1
pytorch-lightning==1.0.3
fasttext==0.9.2

pip install -r requirements.txt

Mecab tokenizer

The model requires an additional installation of the the Mecab tokenizer provided by konlpy package. The guide to install Mecab can be found in this link: https://konlpy.org/en/latest/install/.

How to run

Prepare data

Download the dataset at this link, which is a human-annotated abstractive summarization dataset published by the National Institute of Korean Language. The dataset is arbitrarily split into train, dev, and test.

data
├── nikl_train.pkl
├── nikl_dev.pkl
└── nikl_test.pkl

Run training

First, set up the desired model configurations in config.json.

To begin training your model, run:

python train.py

Details on optional command-line arguments are specified below:

Pointer-generator network

optional arguments:
  -h, --help            show this help message and exit
  -cp CONFIG_PATH, --config-path CONFIG_PATH
                        path to config file
  -m MODEL_PATH, --model-path MODEL_PATH
                        path to load model in case of resuming training from an existing checkpoint
  --load-vocab          whether to load pre-built vocab file
  --stop-with {loss,r1,r2,rl}
                        validation evaluation metric to perform early stopping
  -e EXP_NAME, --exp-name EXP_NAME
                        suffix to specify experiment name
  -d DEVICE, --device DEVICE
                        gpu device number to use. if cpu, set this argument to -1
  -n NOTE, --note NOTE  note to append to result output file name

Running the file will create a subdirectory in logs with the experiment name. All checkpoints, test set predictions, the constructed vocab file, tensorboard logs, and hyperparameter configurations will be saved in this directory.

Run evaluation

python test.py --model-path $PATH-TO-CHECKPOINT

This will report the ROUGE scores on the command-line and save the predicted outputs in .tsv format in the experiment directory where you have loaded the checkpoint.