This software is a state-of-the-art semantic role labeler for the Portuguese language (PropBank-Br corpus). It relies on a deep bidirectional long short-term memory neural network architecture.
| Semantic role | Precision | Recall | F-Score |
|---|---|---|---|
| A0 | 81.82 | 86.90 | 84.28 |
| A1 | 71.15 | 72.29 | 71.71 |
| A2 | 52.73 | 42.03 | 46.77 |
| A3 | 28.57 | 40.00 | 33.33 |
| A4 | 100.00 | 50.00 | 66.67 |
| AM-ADV | 42.86 | 50.00 | 46.15 |
| AM-CAU | 50.00 | 33.33 | 40.00 |
| AM-DIS | 44.44 | 28.57 | 34.78 |
| AM-EXT | 0.00 | 0.00 | 0.00 |
| AM-LOC | 54.17 | 72.22 | 61.90 |
| AM-MED | 0.00 | 0.00 | 0.00 |
| AM-MNR | 34.78 | 47.06 | 40.00 |
| AM-NEG | 90.00 | 94.74 | 92.31 |
| AM-PNC | 42.86 | 66.67 | 52.17 |
| AM-PRD | 100.00 | 33.33 | 50.00 |
| AM-TMP | 66.67 | 73.47 | 69.90 |
| Overall | 67.62 | 68.75 | 68.18 |
Make sure that you have a python 2.7 installed. After cloning this repository, you should download the resources using data_utils.py. Specifically, it will download pre-tained word embeddings and network weigths to be used as an off-the-self semantic role labeler.
You should also install the required dependencies listed below:
The next step is to pre-process the original dataset. Run the prepare_features.py file. It will generate 20 folds from the original training data. This setting is to replicate the paper results.
The you can execute train.py to start training the model.
To come
The paper is currently under review and soon we will have the final bib item