illinois-srl: Semantic Role Labeler

Running

You can use the illinois-srl system in either interactive or annotator mode.

Interactive mode

In interactive mode the user can input a single piece of text and get back the feedback from both the Nominal or Verbal SRL systems in plain text.

To run the system in interactive mode see the class edu.illinois.cs.cogcomp.srl.SemanticRoleLabeler or simply execute the run-interactive script:

For linux:

scripts/run-interactive.sh

For windows:

cd scripts
run-interactive-win.bat

As an Annotator component

illinois-srl can also be used programmatically through the SemanticRoleLabeler class which implemnts CogComp's Annotator interface.

The main method is getView(TextAnnotation) inside SemanticRoleLabeler. This will add a new PredicateArgumentView for either Nominal or Verbal SRL.

Training

To train the SRL system you will require access to the Propbank or Nombank corpora. You need to set pointers to these in the config/srl-config.properties file. (To train the system with a non-Prop/Nombank corpus, you need to extend AbstractSRLAnnotationReader)

To perform the whole training/testing suite, run the Main class with parameters <config-file> expt Verb|Nom true. This will:

1. Read and cache the datasets (train/test)
2. Annotate each TextAnnotation with the required views (here you can set the useCurator flag to false to use the CogComp's standalone NLP pipeline)
3. Pre-extract and cache the features for the classifiers
4. Train the classifiers
5. Evaluate on the (cached) test corpus

IMPORTANT After training, make sure you comment-out the pre-trained SRL model dependencies inside pom.xml (lines 27-38).