🎉MultiTQ

This is the code for the paper Multi-granularity Temporal Question Answering over Knowledge Graphs (Chen et al., ACL 2023).

MULTITQ is a large-scale dataset featuring ample relevant facts and multiple temporal granularities.

🤗Datasets Link: https://huggingface.co/datasets/chenziyang/MultiTQ

Example questions	Answer
Who condemned Abhisit Vejjajiva in May 2010?	Thailand
Who was the first to visit the Middle East in 2008?	Frank Bainimarama
When did the Aam Aadmi Party first negotiated with Harish Rawat?	2015-12-13
Who expressed intent to engage in diplomatic cooperation with Ethiopia before Jun 25th, 2006?	China

Architecture of MultiQA

Dataset and pretrained models

MultiTQ dataset can be found in ./data folder.

git clone https://github.com/czy1999/MultiTQ.git

cd ./MultiTQ/data
unzip Dataset.zip

cd ../MultiQA/models
unzip Models.zip

Environment setup

Option 1 — one-shot script (recommended)

bash setup_env.sh

This creates a conda environment named multitq (Python 3.10), installs PyTorch 2.1.2 (CUDA 12.1 wheel), and installs the rest of the dependencies from requirements.txt. End-to-end verified on RTX 4090.

Option 2 — manual install

conda create -n multitq python=3.10 -y
conda activate multitq
pip install -r requirements.txt

Running the code

MultiQA runs in two stages: an NER preprocessing pass that adds entity/time annotations to each question (output written to data/MultiTQ/questions/processed_questions/), and the actual QA training.

cd MultiQA
python ner_task.py                          # NER preprocessing (uses GPU if available)
python ./train_qa_model.py --model multiqa  # Train MultiQA

The first run of ner_task.py downloads the flair/ner-english-large checkpoint (~1.4 GB). On an RTX 4090 + 22-core CPU, the full preprocessing of train/dev/test (~500k questions) takes about an hour.

Please explore more argument options in train_qa_model.py.

The implementation is based on TempoQR in TempoQR: Temporal Question Reasoning over Knowledge Graphs and their code from https://github.com/cmavro/TempoQR. You can find more installation details there. We use TComplEx KG Embeddings as implemented in https://github.com/facebookresearch/tkbc.

Cite

If you find our method, code, or experimental setups useful, please cite our paper:

@inproceedings{chen-etal-2023-multi,
    title = "Multi-granularity Temporal Question Answering over Knowledge Graphs",
    author = "Chen, Ziyang  and
      Liao, Jinzhi  and
      Zhao, Xiang",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.637",
    pages = "11378--11392",
}