This is the official implementation of our paper, "Unified Instance and Knowledge Alignment Pretraining for Aspect-based Sentiment Analysis" (in Pytorch).

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

• OS: Ubuntu-16.04.1
• GPU: GeForce RTX 2080
• CUDA: 10.2
• cuDNN: v8.0.2

1. target datasets

   • raw data: "./dataset/"
   • processing data: "./dataset_npy/"
   • word embedding file: "./embeddings/"

2. pretraining datasets

   • Amazon review: Amazon Reviews for Sentiment Analysis | Kaggle
   • Yelp review: Yelp Review Sentiment Dataset | Kaggle
   • For the first time, please run "python ./process_data.py" to process the pretraining datasets (remember modifying the path).

• ds_name: the name of target dataset, ['14semeval_laptop', '14semeval_rest', 'Twitter'], default='14semeval_rest'
• pre_name: the name of pretraining dataset, ['Amazon', 'Yelp'], default='Amazon'
• bs: batch size to use during training, [64, 100, 200], default=64
• learning_rate: learning rate to use, [0.001, 0.0005, 0.00001], default=0.001
• n_epoch: number of epoch to use, [5, 10], default=10
• model: the name of model, ['ABGCN', 'GCAE', 'ATAE'], default='ABGCN'
• is_test: train or test the model, [0, 1], default=1
• is_bert: GloVe-based or BERT-based, [0, 1], default=0
• alpha: value of parameter \alpha in knowledge guidance loss of the paper, [0.5, 0.6, 0.7], default=0.06
• stage: the number of training stage, [1, 2, 3, 4], default=4

1. running for the first stage (pretraining on the document)

   • python ./main.py -pre_name Amaozn -bs 256 -learning_rate 0.0005 -n_epoch 10 -model ABGCN -is_test 0 -is_bert 0 -stage 1

2. running for the second stage

   • python ./main.py -ds_name 14semeval_laptop -bs 64 -learning_rate 0.001 -n_epoch 5 -model ABGCN -is_test 0 -is_bert 0 -alpha 0.6 -stage 2

3. runing for the final stage

   • python ./main.py -ds_name 14semeval_laptop -bs 64 -learning_rate 0.001 -n_epoch 10 -model ABGCN -is_test 0 -is_bert 0 -stage 3

4. training from scratch:

   • python ./main.py -ds_name 14semeval_laptop -bs 64 -learning_rate 0.001 -n_epoch 10 -model ABGCN -is_test 0 -is_bert 0 -stage 4

To have a quick look, we saved the best model weight trained on the target datasets in the "./best_model_weight". You can easily load them and test the performance. Due to the limited file space, we only provide the weight of ABGCN on 14semeval_laptop and 14semeval_rest datasets. You can evaluate the model weight with:

• python ./main.py -ds_name 14semeval_laptop -bs 64 -model ABGCN -is_test 1 -is_bert 0
• python ./main.py -ds_name 14semeval_rest-bs 64 -model ABGCN -is_test 1 -is_bert 0

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

• The pretraining datasets are obtained from www.Kaggle.com.

If you find this work helpful, please consider citing as follows:

@article{liu2021unified,
  title={Unified instance and knowledge alignment pretraining for aspect-based sentiment analysis},
  author={Liu, Juhua and Zhong, Qihuang and Ding, Liang and Jin, Hua and Du, Bo and Tao, Dacheng},
  journal={arXiv preprint arXiv:2110.13398},
  year={2021}
}