HATN

Data and source code for our AAAI'18 paper "Hierarchical Attention Transfer Network for Cross-domain Sentiment Classification", which is an extension of our IJCAI'17 paper "End-to-End Adversarial Memory Network for Cross-domain Sentiment Classification".

Demo Video

Click the picture for watching a demo about visualization of attention transfer. => .

Requirements

• Python 2.7.5

• Tensorflow-gpu 1.2.1

• numpy 1.13.3

• nltk 3.2.1

• Google Word2Vec

Environment

• OS: CentOS Linux release 7.5.1804
• GPU: NVIDIA TITAN Xp
• CUDA: 8.0

Tips for Amazon small setting

1. It would be better to use smaller batch size like bs=20.
2. Removing hierarchical position embeddings, which are data-driven.
3. Using some regurizations like dropout to avoid overfitting.

Running

Individual attention learning:

The goal is to automatically capture pos/neg pivots as a bridge across domains based on PNet, which provides the inputs and labels for NPnet. If the pivots are already obtained, you can ignore this step.

python extract_pivots.py --train --test -s dvd [source_domain] -t electronics [target_domain] -v [verbose]

Joint attention learning:

PNet and NPnet are jointly trained for cross-domain sentiment classification. When there exists large domain discrepany, it can demonstrate the efficacy of NPnet.

python train_hatn.py --train --test -s dvd [source_domain] -t electronics [target_domain] -v [verbose]

Training over all transfer pairs:

./all_train.sh

Results

The results are obtained under ramdom seed 0 in this implementation.

Citation

If the data and code are useful for your research, please be kindly to give us stars and cite our paper as follows:

@inproceedings{li2018hatn,
	author = {Zheng Li and Ying Wei and Yu Zhang and Qiang Yang},
	title = {Hierarchical Attention Transfer Network for Cross-Domain Sentiment Classification},
	conference = {AAAI Conference on Artificial Intelligence},
	year = {2018},
}

@inproceedings{li2017end,
  title={End-to-end adversarial memory network for cross-domain sentiment classification},
  author={Li, Zheng and Zhang, Yu and Wei, Ying and Wu, Yuxiang and Yang, Qiang},
  booktitle={Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI 2017)},
  year={2017}
}