KaGRMN-DSG_ABSA

This repository contains the PyTorch source Code for our paper: Understand me, if you refer to Aspect Knowledge: Knowledge-aware Gated Recurrent Memory Network. Bowen Xing and Ivor Tsang.

Under Review in TETCI.

Architectures

Our framework:

Details of KaGRMN cell:

Requirements

Our code relies on Python 3.6 and following libraries:

• transformers==1.1.0
• torch==1.2.0
• numpy==1.16.0
• tqdm==4.50.2
• scikit_learn==0.23.2

Run

    # Restaurant14
    python run.py --cuda_id 0 --dataset_name rest --seed 2021  --embedding_type bert --self_num_heads 3 --rel_num_heads 4 --dropout 0.3 --bert_lr 5e-5 --learning_rate 1e-5 --weight_decay 0.05 --stack_num 4 --n_gcn 2
  
    # Laptop14
    python run.py --cuda_id 0 --dataset_name laptop --seed 2021  --embedding_type bert --self_num_heads 3 --rel_num_heads 2 --dropout 0.3 --bert_lr 1e-5 --learning_rate 5e-5 --weight_decay 0.001 --stack_num 4 --n_gcn 2

    # Restaurant15
    python run.py --cuda_id 0 --dataset_name res15 --seed 2021 --embedding_type bert --self_num_heads 6 --rel_num_heads 6 --dropout 0.3 --bert_lr 3e-5 --learning_rate 2e-5 --weight_decay 0.05 --stack_num 2 --n_gcn 2 --logging_steps 25

Citation

If you use our source code in this repo in your work, please cite the following paper. The bibtex are listed below:

@misc{bowen2021KaGRMN,
      title={Understand me, if you refer to Aspect Knowledge: Knowledge-aware Gated Recurrent Memory Network}, 
      author={Bowen Xing and Ivor Tsang},
      year={2021},
      eprint={2108.02352},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}