code for the paper: A Failure of Aspect Sentiment Classifiers and an Adaptive Re-weighting Solution.

We found Aspect Sentiment Classifier does poorly on ASPECT-LEVEL sentiment classification, because most training examples have the same sentence-level and aspect-level polarity. Sentences with more than 1 opinions of different polarities (we call them contrastive sentences) are rare but truly indicating aspect-level polarity.

We leverage the dataset from SemEval 2014 but augment the testing set of laptop with more contrastive sentences to test aspect-level sentiment. The dataset is self-contained in this repository.

This code base is tested on GTX 1080 Ti, with Ubuntu 16.04, Python3.6, PyTorch 1.0.1 and pytorch transformer 0.4 pip install pytorch-pretrained-bert==0.4.0.

Download weights of BERT-DK:
laptop to pt_model/laptop_pt_review
restaurant to pt_model/rest_pt_review

bash run.sh train

bash run.sh test

@article{xu2019afailure,
  title={A Failure of Aspect Sentiment Classifiers and an Adaptive Re-weighting Solution},
  author={Xu, Hu and Liu, Bing and Shu, Lei and Yu, Philip S},
  journal={arXiv preprint arXiv:1911.01460},
  year={2019}
}