Comma-SRL: Comma Role Labeler

This software extracts relations commas participate in, expanding on previous work in this area. Commas and the surrounding sentence structure often express relations that are essential to understanding the meaning of the sentence.

Structure

There are 2 models:

1. Local Classifier : A classifier that classifies individual commas based on the sentence-parse, shallow parse, POS tags, and NER tags
2. Sequential Model : Uses structured prediction to label a sequence of commas that are siblings in the parse tree of the sentence

There are 2 sources for annotation data:

1. Comma Resolution Data (data/corpus): A set of around a 1000 sentences from section 00 of the PTB in which all the commas have been labeled with their roles. The comma that were labeled as Other have been refined and the annotations are in data/otherFile.txt
2. Comma-Syntax-Pattern annotations (data/Bayraktar-SyntaxToLabel):
   A mapping from a list of the most frequent syntax patterns, extracted from the context of a comma in the parse of the sentence, to comma labels

Execute ./scripts/annotate.sh in the project directory to annotate the commas in the data/infile.txt and receive output in the data/outfile.txt. You can edit the infile to add more sentences. Each sentence must be on a different line. NB: This script requires Maven to be installed.

Run ClassifierComparison to get the performance of different models as evaluated over 5-fold cval.

Use CommaLabeler to obtain a comma View for a sentence represented as a TextAnnotation (must have the views required to extract features for the classifier).

If you use this software please cite our work:

@inproceedings{arivazhagan2016labeling,
  title={Labeling the Semantic Roles of Commas.},
  author={Arivazhagan, Naveen and Christodoulopoulos, Christos and Roth, Dan},
  booktitle={AAAI},
  pages={2885--2891},
  year={2016}
}