PokeMQA

<img src="fig/PokeMQA.png" width="800">

PokeMQA: Programmable knowledge editing for Multi-hop Question Answering has been accepted by ACL 2024 Main Conference !

We release the PokeMQA-turbo_n_edited.py to run PokeMQA on GPT-3.5-turbo-instruct.

Training Dataset for Scope Detector

In datasets\, we release the dataset cls-filtered.json for training scope detector (mentioned in Section 3.2). The details on the dataset construction are illustrated in Appendix B.

Format

Each instance in cls-filtered.json represents a (1edit,4questions) pair:

{
"edit": "Carl Sagan is employed by British Broadcasting Corporation",
"questions":
[
"Who is the employer of Carl Sagan?",
"What is the name of the organization where Carl Sagan works?",
"Which organization employs Carl Sagan?",
"Where does Carl Sagan work?"
]
}

• edit: an edit $e$ in natural language form extracted from MQuAKE-CF.
• questions: four atomic questions belonging to the scope $S(e)$, three of which are generated by Vicuna-13B, another one by manully-defined template.

Pregenerated Knowledge Prompt

We propose knowledge prompt generator (Section 3.3) to enrich contextual information. Specifically, we leverage ELQ, an off-the-shelf entity linking model. It recognizes the key entity in multi-hop questions, links it to Wikidata, and then retrieves the related knowledge facts to construct knowledge prompt. Here we provide the pregenerated knowledge prompt for both MQA dataset, MQuAKE-CF-3k and MQuAKE-T in kgprompt\.

Detector Checkpoints

We release the detector checkpoints trained by us (pre-detector & conflict disambiguator) in [Google Drive].

Commands

To run PokeMQA:

# Start by finetuning distilbert to get pre-detector & conflict disambiguator
OPENAI_API_KEY=YOUR-API-KEY python -m PokeMQA-turbo_n_edited --edited-num 1 --dataset MQuAKE-CF-3k --retraining_detector --retraining_disambiguator --activate_kgprompt

# Skip the finetune stage and load the existing checkpoint of scope detector
OPENAI_API_KEY=YOUR-API-KEY python -m PokeMQA-turbo_n_edited --edited-num 3000 --dataset MQuAKE-CF-3k --detector_name detector-ckpt --dis_name dis-ckpt --activate_kgprompt

If you have any questions about our paper, feel free to email Hengrui Gu guhr22@mails.jlu.edu.cn.

Citation

If you use our code in your research, please cite our work:

@article{gu2023pokemqa,
  title={PokeMQA: Programmable knowledge editing for Multi-hop Question Answering},
  author={Gu, Hengrui and Zhou, Kaixiong and Han, Xiaotian and Liu, Ninghao and Wang, Ruobing and Wang, Xin},
  journal={arXiv preprint arXiv:2312.15194},
  year={2023}
}