This is the official repository for COLING 2025 paper
CycleOIE: A Low-Resource Training Framework For Open Information Extraction
conda create -n cycleoie python=3.10
conda activate cycleoie
pip install torch==2.1
pip install transformers
pip install openai
cd ./llm_annotate
Instruct gpt with openie annotation principles
python principles.py
Instruct gpt with few-shot examples, formatted into multi-turn chat
python examples.py
cd ../
python cycle_train_t5.py
(Our designed formulation for extractions.)
We use three additional tokens to split each part of an extraction. For example,
Sentence:
The track was in excellent condition , running through the trees .
Extraction:
The track | was in | excellent condition
The track | was running through | the trees
We formulate it into a sequence:
subject <is> The track <and> relation <is> was in <and> object <is> excellent condition <then> subject <is> The track <and> relation <is> was running through <and> object <is> the trees