IDK-MRC

Code and dataset for EMNLP 2022 paper titled "IDK-MRC: Unanswerable Questions for Indonesian Machine Reading Comprehension".

Dataset Description

I(n)dontKnow-MRC (IDK-MRC) is an Indonesian Machine Reading Comprehension dataset that covers answerable and unanswerable questions. Based on the combination of the existing answerable questions in TyDiQA, the new unanswerable question in IDK-MRC is generated using a question generation model and human-written question. Each paragraph in the dataset has a set of answerable and unanswerable questions with the corresponding answer.

Besides IDK-MRC dataset, several baseline datasets also provided:

Trans SQuAD : machine translated SQuAD 2.0 (Muis and Purwarianti, 2020)
TyDiQA: Indonesian answerable questions set from the TyDiQA-GoldP (Clark et al., 2020)
Model Gen: TyDiQA + the unanswerable questions output from the question generation model
Human Filt: Model Gen dataset that has been filtered by human annotator

You can find all datasets on the dataset directory. You can also load it using NusaCrowd or HuggingFace.

Running Models

Environment Setup

We encourage you to create a virtual environment using conda by running this script:

conda create --n "<your_env_name>" python=3.8.13
conda activate <your_env_name>

Install all dependencies:

pip install -r requirements.txt

Then, run this python script to download stanza resources:

import stanza
stanza.download('id')

Available Models

3 MRC models finetuned on IDK-MRC dataset can be downloaded from HuggingFace:

IndoBERT (base, uncased): rifkiaputri/indobert-base-id-finetune-idk-mrc
m-BERT (base, cased): rifkiaputri/mbert-base-id-finetune-idk-mrc
XLM-R (base, cased): rifkiaputri/xlmr-base-id-finetune-idk-mrc

Also, 1 Unanswerable Question Generation (Unanswerable QG) model trained on translated SQuAD dataset can be downloaded from HuggingFace:

mT5 (base, cased): rifkiaputri/mt5-base-id-finetune-unans-qg

Train Your Own Model

Run this following script to train MRC model:

python qa/run_qa.py \
    --train_path dataset/idk_mrc/train.json \
    --val_path dataset/idk_mrc/valid.json \
    --test_path dataset/idk_mrc/test.json \
    --epoch 10 \
    --model_name bert \
    --pretrain_name indobenchmark/indobert-base-p2 \
    --n_gpu 1 \
    --wb_name idk-mrc --wb_name_test idk-mrc-test \
    --out_path outputs \
    --out_name indobert_squad \
    --lang id \
    --uncased \
    --seed 42

Run this following script to train Unanswerable QG model:

python run_mt5.py \
    --train_path dataset/qg/train_id_aligned.json \
    --val_path dataset/qg/dev_id_aligned.json \
    --val_path_for_writing dataset/qg/dev_id_aligned_writing.json \
    --pretrain_name google/mt5-base \
    --qa_model outputs/qa/qgen-squad-id-trans/ \
    --n_gpu 1 \
    --wb_train_name mt5_squad_train \
    --wb_eval_name mt5_squad_ubleu_eval \
    --wb_eval_run_name mt5_base_id \
    --out_dir outputs/ \
    --cache_dir cache_dir/ \
    --seed 42 \
    --num_train_epochs 5 \
    --train_batch_size 8 \
    --max_seq_length 512 \
    --top_k 50 \
    --top_p 0.95 \
    --num_return_seq 10 \
    --do_sample

Citation

Please cite this paper if you use any code or dataset in this repository:

@inproceedings{putri-oh-2022-idk,
    title = "{IDK}-{MRC}: Unanswerable Questions for {I}ndonesian Machine Reading Comprehension",
    author = "Putri, Rifki Afina  and
      Oh, Alice",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.465",
    pages = "6918--6933",
}

rifkiaputri / IDK-MRC