MXNet implementation of ICLR 2018 paper: A new method of region embedding for text classification.
Official implementation in TensorFlow.
- I implemented both Word-Context and Context-Word Region Embedding in the paper.
- Please see the original papar about the datasets and pre-pocessing.
- All the hyper-parameters I used are copied from the official implementation.
- Python2 or Python3
- Mxnet 1.2.1
| Datasets | Accuracy(%) WordContext |
Best Epoch WordContext |
Accuracy(%) ContextWord |
Best Epoch ContextWord |
Running Time Per Epoch(mins) |
|---|---|---|---|---|---|
| Yahoo Answer | 73.07(73.7) | 2 | 73.42(73.4) | 3 | 110 |
| Amazon Polarity | 95.27(95.1) | 2 | 95.36(95.3) | 3 | 247 |
| Amazon Full | 61.58(60.9) | 2 | 61.59(60.8) | 2 | 183 |
| Ag news | 92.96(92.8) | 6 | 92.89(92.8) | 8 | 2 |
| DBPedia | 98.91(98.9) | 4 | 98.88(98.9) | 3 | 23 |
| Yelp Full | 64.98(64.9) | 3 | 64.94(64.5) | 2 | 25 |
Note:
- The accuracy in brackets are results reported in the original paper.
- The running speed is much faster than the origin implementation in Tensorflow.
The running time was tested on the model of context-word region embedding, which run roughly the same as the word-context region embedding. - The code run on a Titan Xp GPU.