End-to-End Adversarial Memory Network for Cross-domain Sentiment Classification
Authors: Zheng Li, Yu Zhang, Ying Wei, Yuxiang Wu, Qiang Yang
IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on the Amazon review dataset demonstrate that AMN can significantly outperform state-of-the-art methods. |
| Researcher Affiliation | Academia | Zheng Li, Yu Zhang , Ying Wei , Yuxiang Wu, Qiang Yang Department of Computer Science and Engineering, Hong Kong University of Science and Technology {zlict,zhangyu,yweiad,ywubw,qyang}@cse.ust.hk |
| Pseudocode | No | The paper describes the model components and training process in detail, but it does not include a formal pseudocode block or algorithm section. |
| Open Source Code | No | The paper does not provide a statement about releasing its own source code or a link to a code repository for the proposed AMN model. It only mentions using source codes for baseline methods (SCL, DANN, DAm SDA). |
| Open Datasets | Yes | Experiments are conducted on the Amazon reviews dataset [Blitzer et al., 2007], which has been widely used for crossdomain sentiment classification. |
| Dataset Splits | Yes | For each transfer pair A B, we randomly choose 800 positive and 800 negative reviews from the source domain A as the training data, the rest from the source domain A as the validation data, and 200 positive and 200 negative reviews from the target domain B for testing. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper mentions using 'word2vec vectors' but does not specify any software dependencies with their version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific library versions) for reproducibility. |
| Experiment Setup | Yes | The memory size m is set to 500 and the number of hops is 3. The regularization weight ρ is set to 0.05, which is obtained via 5-fold cross-validation on the labeled data in the source domain and is used for all transfer pairs. The model is optimized with the stochastic gradient descent over shuffled mini-batches with momentum rate 0.9. Due to different training sizes for the sentiment classifier and domain classifier, we set the batch size bd for the domain classifier with 100, half coming from the source and target domains, and use the same number of batches for both classifiers. Gradients with the ℓ2 norm larger than 40 are normalized to be 40. We define the training progress as p = t T , where t and T are current epoch and the maximum one, respectively, and T is set to 120. By following [Ganin et al., 2016], the learning rate is decayed as η = 0.0075 (1+10p)0.75 and the adaptation rate is increased as λ = 2 1+exp( 10p) 1 during training. |