A Tensorflow implementation for "Interactive Attention Networks for Aspect-Level Sentiment Classification" (Dehong Ma, IJCAI 2017)
- use
pip install -r requirements.txt
to install required packages - Create three empty folders: 'analysis' for saving analyzing results, 'logs' for saving experiment logs and 'models' for saving experiment models
- Download the 300-dimensional pre-trained word vectors from Glove and save it in the 'data' folder as 'data/glove.840B.300d.txt'
|--- data
| |--- laptop
| |--- restaurant
| |--- data_info.txt - the preprocessing data information file
| |--- test_data.txt - the preprocessing testing data file
| |--- train_data.txt - the preprocessing training data file
|--- main.py
|--- model.py
|--- transfer.py - transfering the origin xml files to text files
|--- utils.py
|--- README.md
Dataset | Accuracy |
---|---|
Laptop | 70.846 |
Restaurant | 79.107 |
Note: In the newest version, the results are worse than the results given above, since the code of the model is revised. I will optimize the model sooner and report the results.
- Implementing by other deep learning frameworks
- Softmax mask
- Optimization to get better performance