In this coursework, I implemented various advanced neural network architectures for a binary text classification task. Based on several successful convolutional neural networks (CNN), I proposed a novel hierarchical neural model, which is compatible with inherent hierarchical properties of documents. The experiment shows that the proposed model outperforms conventional character-level CNNs for text classification on AUC criterion. Even without sufficient training, the stacking tricks integrating existing models help further improve the final performance in the in-class kaggle competition.
In this coursework, I design a hierarchical neural model, which is shown below.
For detailed information and evaluation, see my final report here.
I ran experiments on my PC.
- Operating System: CentOS Linux release 7.3.1611 (Core)
- CPU: Intel(R) Core(TM) i5-4460 CPU @ 3.20GHz
- No GPU
- Total Memory: 16142596 kB. The code is in Python 2.7 and I implemented neural networks with PyTorch.