LLM-QA4RE

Data and code for ACL 2023 Findings: Aligning Instruction Tasks Unlocks Large Language Models as Zero-Shot Relation Extractors.

We present LLM-QA4RE, which aligns underrepresented tasks in the instruction-tuning dataset (relation extraction) to a common task (question answering) to unlock instruction-tuned LLMs' abilities on relation extraction.

QA4RE achieves significant and consistent performance gains over 6 LLMs across 4 datasets. In addition, it shows strong transferability to model sizes from 175B (GPT-3.5 series) to even 80M (FLAN-T5 Small).

TODO

• Organize and release Code for GPT-3.5 Series
• Release output results of GPT-3.5 Series
• Organize and release CODE for FLAN T5 Series
• Refactor code to save results as json/jsonl

Installation

Run the following commands to create a conda environment with the required packages.

conda create -n QA4RE python=3.9 pip
conda activate QA4RE
pip install -r requirements.txt
# same env with few-shot-bioIE

Data and Launch

Download data and subsets via Google Drive

Results and prompts are saved in Google Drive

Unzip directly in ./ and then the root folder should organize like this:

.
├───  data
│    ├───  RETACRED
│    ├───  TACRED
│    ├───  TACREV
│    ├───  semeval
├───  outputs
│    ├───  RETACRED
│    ├───  TACRED
│    ├───  TACREV
│    ├───  semeval
├───  projs
│    ├───  QA4RE
│    ├───  vanillaRE
│    ├───  README.md
│    ├───  re_templates.py
│    └───  re_utils.py
├───  utils
│   ...

For running, please refer to the README in ./projs dir.

Results

QA4RE works on GPT-3.5 Series and FLAN-T5 Series, 6 LLMs in total

QA4RE works on smaller instruction-tuned models.

Cite

If you find our paper, code, or data helpful, please cite it.

@inproceedings{Zhang2023LLM-QA4RE,
  title={Aligning Instruction Tasks Unlocks Large Language Models as Zero-Shot Relation Extractors},
  author={Kai Zhang, Bernal Jiménez Gutiérrez, Yu Su},
  booktitle={Findings of ACL 2023},
  year={2023}
}

Question

If you have any questions, please feel free to contact drogozhang[AT]gmail[DOT]com or open an issue so we can help you better and quicker :)