4/2/2024

• Our code is updated, i.e., fixing the scripts for BERT-based models.

6/3/2023

• Our paper has been accepted by IEEE Transactions on Knowledge and Data Engineering. (Paper)

26/4/2022

• Knowledge graph embeddings is added. Additionally, we update the implementation of KGNN-BERT and ASGCN-BERT. The corresponding training code is also updated.

• python 3.7
• pytorch-gpu 1.7
• numpy 1.19.4
• pytorch_pretrained_bert 0.6.2
• nltk 3.3
• spacy 2.2.0
• en_core_web_sm 2.1.0
• GloVe.840B.300d
• bert-base-uncased

• OS: Windows 10
• GPU: NVIDIA GeForce GTX 1660 SUPER
• CUDA: 11.0
• cuDNN: v8.0.4

• raw data: "./dataset/"
• processing data: "./dataset_npy/"
• word embedding file: "./embeddings/"
• Knowledge graph embedding file: download from Google Drive

• ds_name: the name of target dataset, ['14semeval_laptop','14semeval_rest','15semeval_rest','16semeval_rest','Twitter'], default='14semeval_rest'
• bs: batch size to use during training, [32, 64], default=64
• learning_rate: learning rate to use, [0.001, 0.0005], default=0.001
• dropout_rate: dropout rate for sentimental features, [0.1, 0.3, 0.5], default=0.05
• n_epoch: number of epoch to use, default=20
• model: the name of model, default='KGNN'
• dim_w: the dimension of word embeddings, default=300
• dim_k: the dimension of graph embeddings, [200,400], default=200
• is_test: train or test the model, [0, 1], default=1
• is_bert: GloVe-based or BERT-based, [0, 1], default=0

More training options can be found in "./main_total.py".

• python -m main_total -ds_name 14semeval_laptop -bs 32 -learning_rate 0.001 -dropout_rate 0.5 -n_epoch 20 -model KGNN -dim_w 300 -dim_k 400 -kge analogy -gcn 0 -is_test 0 -is_bert 0
• python -m main_total -ds_name 14semeval_rest -bs 64 -learning_rate 0.001 -dropout_rate 0.5 -n_epoch 20 -model KGNN -dim_w 300 -dim_k 200 -kge distmult -gcn 0 -is_test 0 -is_bert 0

• python -m main_total -ds_name 14semeval_laptop -bs 32 -learning_rate 0.00003 -n_epoch 20 -model KGNN -dim_w 768 -dim_k 400 -kge analogy -gcn 0 -is_test 0 -is_bert 1
• python -m main_total -ds_name 14semeval_rest -bs 64 -learning_rate 0.00003 -n_epoch 20 -model KGNN -dim_w 768 -dim_k 200 -kge distmult -gcn 0 -is_test 0 -is_bert 1

The detailed training scripts can be found in "./scripts".

To have a quick look, we saved the best model weight trained on the evaluated datasets in the "./model_weight/best_model_weight". You can easily load them and test the performance. You can evaluate the model weight with:

• python -m main_total -ds_name 14semeval_laptop -bs 32 -model KGNN -dim_w 300 -dim_k 400 -is_test 1
• python -m main_total -ds_name 14semeval_rest -bs 64 -model KGNN -dim_w 300 -dim_k 200 -is_test 1

• The datasets and more than 50% of the code are borrowed from TNet-ATT (Code) (Tang et.al, ACL2019).

@article{zhong2022knowledge,
  title={Knowledge Graph Augmented Network Towards Multiview Representation Learning for Aspect-based Sentiment Analysis}, 
  author={Zhong, Qihuang and Ding, Liang and Liu, Juhua and Du, Bo and Jin, Hua and Tao, Dacheng},
  journal={IEEE Transactions on Knowledge and Data Engineering}, 
  year={2023},
  pages={1-14},
  doi={10.1109/TKDE.2023.3250499}
  }