Skip to content
Indonesian part-of-speech (POS) tagging
Jupyter Notebook Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
data Update dataset to include README Sep 11, 2018
models Make a separate script for each model May 20, 2018
notebooks Update neural fine analysis notebook Jul 17, 2018
.gitignore Implement memorization baseline model Apr 30, 2018
environment.yml Update new best config on each fold Jul 15, 2018 Reorganize models package May 19, 2018 No need to record per label metric at each CRF iteration May 26, 2018

Indonesian Part-of-Speech (POS) Tagging

This repository contains the implementation of our paper:

Kurniawan, K., & Aji, A. F. (2018). Toward a Standardized and More Accurate Indonesian Part-of-Speech Tagging. 2018 International Conference on Asian Language Processing (IALP), 303–307.


Make sure you have conda package manager. Then, create a conda virtual environment with

$ conda env create -f environment.yml

The command will create a virtual environment named id-pos-tagging and also install all the required packages. Once it is done, activate the virtual environment to get started.

$ source activate id-pos-tagging


The dataset is available in data/dataset.tar.gz. Decompress this file and you will have train.X.tsv, dev.X.tsv, and test.X.tsv files for all 5 folds with X replaced with the fold number.

Running experiments

Scripts to run our models are prefixed with run_. So, for example, to run the CRF model, use script. All scripts use Sacred to manage the experiment configuration and results. We will explain in a more detail and use this script as the example.


A minimal command to train a model is

$ ./ train with corpus.train=train.01.tsv

This will train a CRF model on the given training corpus and save the model in model file in the current directory. There are many configuration that can be set, which can all be listed with

$ ./ print_config

The command above will show all the configuration for the script, including those that might be needed for commands other than train. The print_config command is available for other run_*.py scripts as well.

To make reproduction easier, we already named our best configuration reported in the paper as tuned_on_foldX where X is the fold number. For instance, to get our result on fold 1, run

$ ./ train with tuned_on_fold1 corpus.train=train.01.tsv

These named configurations are also available for and

Evaluation and prediction

To evaluate/predict, use evaluate and predict command respectively. The available configuration is still the same as that of training.

Observing experiments

Sacred allows us to save experiment runs to a MongoDB database. To enable this for our scripts, simply set SACRED_MONGO_URL and SACRED_DB_NAME to your MongoDB instance. Once this is done, an experiment run will be saved everytime you run any run_*.py scripts.




If you use our work, please cite:

  place={Bandung, Indonesia},
  title={Toward a Standardized and More Accurate Indonesian Part-of-Speech Tagging},
  note={arXiv: 1809.03391},
  booktitle={2018 International Conference on Asian Language Processing (IALP)},
  author={Kurniawan, Kemal and Aji, Alham Fikri},
You can’t perform that action at this time.