The dynamic vocavbulary predictor taken from the NLG-RL source code is modified to train long texts in our film_bow and film_raw dataset. The parameters set as following:
- Dimensionality for embeddings and hidden states: 256
- Maximum number of training epochs: 20
- Learning rate for AdaGrad: 8.0e-02
- Learning rate decay for AdaGrad: 0.5
- Batch size: 128
- Dropout rate for residual block: 0.4
- Weight decay rate for internal weight matrices: 1.0e-06
- Clipping value for gradient norm: 1.0
The training is performed with different values of small vocabulary size for evaluating recall (K). The results for BOW is found as follows.
| K | Recal Score |
|---|---|
| 1 | 11.04 % |
| 1000 | 89.92 % |
| 2000 | 94.69 % |
| 3000 | 96.72 % |
| 4000 | 97.77 % |
| 5000 | 98.46 % |
| 6000 | 98.89 % |