Skip to content


Repository files navigation

Text Summarization

This repository contains the code for our work:

Kurniawan, K., & Louvan, S. (2018). IndoSum: A New Benchmark Dataset for Indonesian Text Summarization. In 2018 International Conference on Asian Language Processing (IALP) (pp. 215–220). Bandung, Indonesia: IEEE.


Create a virtual environment from environment.yml file using conda:

$ conda env create -f environment.yml

To run experiments with NeuralSum [CL16], Tensorflow is also required.


Get the dataset from

Preprocessing for NeuralSum

For NeuralSum, the dataset should be further preprocessed using

$ ./ -o neuralsum train.01.jsonl

The command will put the oracle files for NeuralSum under neuralsum directory. Invoke the script with -h/--help to see its other options.

Running experiments

The scripts to run the experiments are named run_<model>.py. For instance, to run an experiment using LEAD, the script to use is All scripts use Sacred so you can invoke each with help command to see its usage. The experiment configurations are fully documented. Run ./run_<model>.py print_config to print all the available configurations and their docs.

Training a model

To train a model, for example the naive Bayes model, run print_config command first to see the available configurations:

$ ./ print_config

This command will give an output something like:

INFO - summarization-bayes-testrun - Running command 'print_config'
INFO - summarization-bayes-testrun - Started
Configuration (modified, added, typechanged, doc):
  cutoff = 0.1                       # proportion of words with highest TF-IDF score to be considered important words
  idf_path = None                    # path to a pickle file containing the IDF dictionary
  model_path = 'model'               # where to load or save the trained model
  seed = 313680915                   # the random seed for this experiment
    dev = None                       # path to dev oracle JSONL file
    encoding = 'utf-8'               # file encoding
    lower = True                     # whether to lowercase words
    remove_puncts = True             # whether to remove punctuations
    replace_digits = True            # whether to replace digits
    stopwords_path = None            # path to stopwords file, one per each line
    test = 'test.jsonl'              # path to test oracle JSONL file
    train = 'train.jsonl'            # path to train oracle JSONL file
    delete_temps = True              # whether to delete temp files after finishes
    on = 'test'                      # which corpus set the evaluation should be run on
    size = 3                         # extract at most this number of sentences as summary
    path = 'test.jsonl'              # path to the JSONL file to summarize
    size = 3                         # extract at most this number of sentences as summary
INFO - summarization-bayes-testrun - Completed after 0:00:00

So, to train the model on a train corpus in /tmp/train.jsonl and save the model to /tmp/models/bayes.model, simply run:

$ ./ train with corpus.train=/tmp/train.jsonl model_path=/tmp/models/bayes.model

Evaluating a model

Evaluating an unsupervised model is simple. For example, to evaluate a LEAD-N summarizer:

$ ./ evaluate with corpus.test=/tmp/test.jsonl

This command will print an output like this:

INFO - run_experiment - Running command 'evaluate'
INFO - run_experiment - Started
INFO - read_jsonl - Reading test JSONL file from /tmp/test.jsonl
INFO - evaluate - References directory: /var/folders/p9/4pp5smf946q9xtdwyx792cn40000gn/T/tmp7jct3ede
INFO - evaluate - Hypotheses directory: /var/folders/p9/4pp5smf946q9xtdwyx792cn40000gn/T/tmpnaqoav4o
INFO - evaluate - ROUGE scores: {'ROUGE-1-R': 0.71752, 'ROUGE-1-F': 0.63514, 'ROUGE-2-R': 0.62384, 'ROUGE-2-F': 0.5502, 'ROUGE-L-R': 0.70998, 'ROUGE-L-F': 0.62853}
INFO - evaluate - Deleting temporary files and directories
INFO - run_experiment - Result: 0.63514
INFO - run_experiment - Completed after 0:00:11

Evaluating a trained model is done similarly with model_path configuration is set to the path to the saved model.

Setting up Mongodb observer

Sacred allows the experiments to be observed and saved to a Mongodb database. The experiment scripts above can readily be used for this, simply set two environment variables SACRED_MONGO_URL and SACRED_DB_NAME to your Mongodb authentication string and database name (to save the experiments into) respectively. Once set, the experiments will be saved to the database. Use -u flag when invoking the experiment script to disable saving.

Reproducing results

All best configurations obtained from tuning on the development set are saved as Sacred's named configurations. This makes it easy to reproduce our results. For instance, to reproduce our LexRank result on fold 1, simply run:

./ evaluate with tuned_on_fold1 corpus.test=test.01.jsonl

Since the best configuration is named as tuned_on_fold1, the command above will use that configuration and evaluate the model on the test set. In general, all run scripts have tuned_on_foldX named configuration, where X is the fold number. For though, there are other named configurations, namely emb300_on_foldX and fasttext_on_foldX, referring to the scenario of using word embedding size of 300 and fastText pretrained embedding respectively. Some run scripts do not have such named configurations; that is because their hyperparameters were not tuned/they do not have any.


Apache License, Version 2.0.


If you're using our code or dataset, please cite:

  place={Bandung, Indonesia},
  title={IndoSum: A New Benchmark Dataset for Indonesian Text Summarization},
  booktitle={2018 International Conference on Asian Language Processing (IALP)},
  author={Kurniawan, Kemal and Louvan, Samuel},

Cheng, J., & Lapata, M. (2016). Neural summarization by extracting sentences and words. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (pp. 484–494). Berlin, Germany: Association for Computational Linguistics. Retrieved from