Skip to content

README.md

Latest commit

 

History

History
115 lines (102 loc) · 3.96 KB

README.md

File metadata and controls

115 lines (102 loc) · 3.96 KB

Large Language Models for Continual Relation Extraction

Published at Applied Research Paper Track of IEEE Access Journal!

Cite

@ARTICLE{efeoglu_2026,
  author={Efeoglu, Sefika and Paschke, Adrian and Schimmler, Sonja},
  journal={IEEE Access}, 
  title={Large Language Models for Continual Relation Extraction}, 
  year={2026},
  volume={},
  number={},
  pages={1-1},
  keywords={Semantic Web;Computer networks;Continual Relation Extraction;Schema-Level Errors;Large Language Models;Knowledge Graph Construction},
  doi={10.1109/ACCESS.2026.3682652}}

CRE Note Trained models are public on HuggingFace, as stated in the journal article.

Folder Structure

.
├── LICENSE
├── README.md
├── config.ini
├── data                            -> settings and data split setting here for tacred and fewrel like relation types per task
├── doc                             -> figures
├── results                         -> results for TACRED with Flan-T5 and All Results for FewRel
├── logs                            -> time cost logs for each experiment and FewRel's in side of FewRel results
├── main.py
├── requirements.txt                -> dependecies like libraries
└── src
    ├── CRE                         -> continual training of Flan T5 Base, Llama2 and Mistral
    ├── analysis_viz                -> Visualization like logs and  section 4 figures.
    ├── clean                       -> cleaning of results of llama and mistral from explainations and instructions.
    ├── data_preparetation          -> prompt dataset generation
    ├── metrics                     -> bwt, whole and average accuracy calculation
    ├── utils.py                    -> read and write
    └── zero_shot_prompting         -> ablation study, but not in the paper.

How it works

Setup configuration in config.ini according to your needs before starting running experiments.

$ python main.py

or follow the steps below.

1.) Prepare datasets:

TACRED:

  • This command with convert data row to (sentence, subject, object, object_type and subject_type)
$ python src/data_preparetation/data_prepare_tacred.py
  • Split datasets according to setting Cui et al. 2021
$ python src/data_preparetation/instruction_ft_data_same_setting_tacred.py

FewRel

  • Same steps with TACRED
$ python src/data_preparetation/data_preparation_fewrel.py
  • split
$ python src/data_preparetation/instruction_ft_data_same_setting_fewrel.py

2.)Trainer

  • Decoder only models(Llama2-7B-chat-hf and Mistral-Instruct-7B-v2.0)
$ python python src/CRE/trainer_decoder.py
  • Encoder-Decoder model(Flan T5-Base)
$ python src/CRE/trainer_t5.py

3.) Clean decoder-only models results from explainations

$ python src/clean/clean_decoder_results.py

4.) Metrics

Average and Whole Accuracy Metrics

$ python src/metrics/cl_metrics.py

Backward Knowledge Transfer Computation

$ python src/metrics/bwt.py

References

@inproceedings{cui-etal-2021-refining,
  title     = {{R}efining {S}ample {E}mbeddings with {R}elation {P}rototypes to {E}nhance {C}ontinual {R}elation {E}xtraction},
  author    = {Cui, Li and Yang, Deqing and Yu, Jiaxin and Hu, Chengwei and Cheng, Jiayang and Yi, Jingjie and Xiao, Yanghua},
  editor    = {Zong, Chengqing and Xia, Fei and Li, Wenjie and Navigli, Roberto},
  booktitle = {Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)},
  month     = {8},
  year      = {2021},
  address   = {Online},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2021.acl-long.20},
  doi       = {10.18653/v1/2021.acl-long.20},
  pages     = {232--243}
}