Semantic Role Labeling is a Natural Language Processing problem that consists in the assignment of semantic roles to words in a sentence. Specifically, given the main predicate of a sentence, the task requires the identification (and correct labeling) of the predicate's semantic arguments. A simple example is the sentence "the cat eats a fish", with cat and fish rispectively the agent and the patient of the main predicate eats.
The project consists in the implementation of a Semantic Role Label classifier inspired by the article "Encoding Sentences with Graph Convolutional Networks for Semantic Role Labeling" by Marcheggiani and Titov. An in detail report about the project and the assignment's specification can be found in the docs folder.
Tensorflow (either for cpu or gpu, version >= 1.9 and < 2.0) is required in order to run the system.
The other software dependencies can be found in requirements.txt and installed by running the command:
pip install -r requirements.txt
The system can be used to train a model, evaluate it, or predict the semantic labels for some unseen data.
To do so, the module run.py should be invoked, using the necessary input arguments;
A brief explenation of the software's options can be obtained by running
python run.py -h
IMPORTANT: In order to work properly, the system requires the download of this data. After downloading the content, place it into the data directory.
In order to train the system on the Semantic Role Labeling task, run the command:
python run.py --train <epochs> --params <param_folder>
The argument <epochs> is the number of epochs that will be used during training.
<param_folder> is the folder that will contain the trained parameters (weights) used by the classifier.
Pre-trained models are available in this link. After download, place these models in the models directory. You can then use these through the commands
python run.py --params ../models/original <...>
or
python run.py --gated --params ../models/gated <...> ,
depending on the model you are using.
It is possible to assess the performance of a trained classifier by invoking
python run.py --eval --params <param_folder>
The argument <param_folder> should contain the trained parameters (weights) used by the SRL classifier. A good classifier should have Precision, Recall and F1 around
| Precision | Recall | F1 |
|---|---|---|
| 0.83 | 0.81 | 0.82 |
By invoking the command
python run.py --predict <data-file> --params <param_folder>,
it is possible to predict the classifier output with respect to the data stored in
<data-file> (file that must follow the CoNLL 2009 data format). The predicted labels will be stored in the file <data-file>.out.