Exploiting Coarse-to-Fine Task Transfer for Aspect-Level Sentiment Classification
Authors: Zheng Li, Ying Wei, Yu Zhang, Xiang Zhang, Xin Li4253-4260
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, extensive experiments demonstrate the effectiveness of the MGAN. Our full model MGAN consistently and significantly achieves the best results in all target domains, outperforming the strongest baseline RAM by 4.13%, 3.58%, 5.26% for accuracy and 2.99%, 2.94% and 6.23% for Macro-F1 on average. |
| Researcher Affiliation | Collaboration | 1Hong Kong University of Science and Technology, Hong Kong 2Tencent AI Lab, Shenzhen, China 3The Chinese University of Hong Kong, Hong Kong |
| Pseudocode | No | The paper describes the model architecture and components using text and mathematical equations, but it does not include pseudocode or an algorithm block. |
| Open Source Code | No | The dataset is available at https://github.com/hsqmlzno1/MGAN. This link is explicitly stated for the dataset, not the source code for the methodology itself. |
| Open Datasets | Yes | Source: AC-level We build a large-scale, multi-domain dataset named Yelp Aspect as source domains... The dataset is available at https://github.com/hsqmlzno1/MGAN. For target domains, we use three public benchmark datasets: Laptop (L), Restaurant (R2) and Twitter (T). The Laptop and Restaurant are from Sem Eval 14 ABSA challenge (Kiritchenko et al. 2014)... The Twitter dataset is collected by (Dong et al. 2014). |
| Dataset Splits | Yes | For each transfer pair Ds Dt, the training data from domain Ds and randomly sampled 90% training data from domain Dt are used for training, the rest 10% training data from Dt is used for validation, and the testing data from Dt is used for testing. The hyperparameters are tuned on 10% randomly held-out training data of the target domain in R1 L task and are fixed to be used in all transfer pairs. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using GloVE vectors and the Adam optimizer, but it does not specify version numbers for these or other software dependencies. |
| Experiment Setup | Yes | The word embeddings are initialized with 200-dimension Glo VE vectors... de, dh, du are set to be 200, 150 and 100, respectively. The fc layer size is 300. The Adam (Kingma and Ba 2014) is used as the optimizer with the initial learning rate 10 4. Gradients with the ℓ2 norm larger than 40 are normalized to be 40. All weights in networks are randomly initialized from a uniform distribution U( 0.01, 0.01). The batch sizes are 64 and 32 for source and target domains, respectively. The control-off factors λ, ρ are set to be 0.1 and 10 6. To alleviate overfitting, we apply dropout on the word embeddings of the context with dropout rate 0.5. We also perform early stopping on the validation set during the training process. |