The Structured Weighted Violation MIRA

Code for "The Structured Weighted Violation MIRA" by Dor Ringel, Rotem Dror, and Roi Reichart

Dependencies

In order to run the program the following dependencies need to be installed:

• Java 1.8:

  • Install Java 1.8 SDK and choose it as the project's SDK.

• Gurobi optimization package:

  • Download Gurobi from their website
  • Install it
  • Create and activate a license (free for academic purposes)
  • Add installed gurobi.jar to the project's SDKs

• Weka machine learning toolkit:

  • Download Weka from their website
  • Install it
  • Add the installed weka.jar to the project's SDKs

• Mallet machine learning toolkit:

  • Install via Maven, using address: cc.mallet:mallet:2.0.8

• Gson Java to Json converter

  • Install via Maven, using address: com.google.code.gson:gson:2.8.2

Running the code

Endpoint

Running the code is done by the main method in Main class, located in: src/main/java/examples/Main.java

The parameters can be passed into the various experiments using the cvParams object (defined in src/main/java/examples/Main.java)

General parameters

There are three mandatory general parameters:

• dataset: dataset to use. can be any of the following:

  • genia
  • bc2gm
  • conll2002
  • conll2000-chunking

• numTrainIterations: number of iteration for the training process.

  • Has to be a positive integer between 1 and 30.

• maxFolds: number of folds to run.

  • Has to be a positive integer between 1 and 9.

Algorithm specific parameters

Additional algorithm-specific parameters include:

• UpdatorClass - which updator algorithm to use.

• kMira: K for MIRA (when choosing updatorClass KBestMiraUpdator).

  • Has to be a positive integer between 1 and 9.

• topK: top K inference results (when choosing updatorClass SWVPUpdator or SWVMUpdaor).

  • Has to be a positive integer between 1 and 9.

• maType: type of modification templates to use (when choosing updatorClass SWVPUpdator or SWVMUpdaor).

  • can be aggresive, passive, or all.

• jjTypw: size of modification templates to use (when choosing updatorClass SWVPUpdator or SWVMUpdaor).

  • can be single or double.

• gammaObjectiveType: type of objective function (when choosing updatorClass SWVPUpdator or SWVMUpdaor, and gammaCalculationMethod four).

  • can be MAXIMIZE or MINIMIZE.

• gammaCalculationMethod: method for selecting gammas (when choosing updatorClass SWVPUpdator or SWVMUpdaor).

  • can be uniform, wm, wmr, softmin or four.

• gammaCalculationBeta: beta value for gamma calculation (when choosing updatorClass SWVPUpdator or SWVMUpdaor).

  • can be a positive decimal number.

Logging and output files

Each run creates a output-directory in src/main/java/examples/output_files/cv/{dataset}, where {dataset} refers to the chosen dataset. This output-directory contains average, fold-level, and sentence-level statistics and also the raw log file.

Logging is also sent to stdout.

Examples

Example #1 - CSP:

In order to run CSP algorithm on the BC2GM dataset for 10 iteration and 2 on folds the following parameters should be specified:

cvParams.put("dataset", "bc2gm");
cvParams.put("numTrainIterations", "10");
cvParams.put("maxFolds", "2");
cvParams.put("updatorClass", "PerceptronUpdator");

Example #2 - SWVM:

In order to run SWVM algorithm on the CONLL2002 dataset for 15 iteration and 5 on folds, using size one modification templates, updating in respect to top three inference results, aggressive approach, and uniform gamma selection, the following parameters should be specified:

gammaObjectiveType and gammaCalculationBeta parameters will be ignored (due to the selection of uniform approach) but a legal value has to be specific for both of them regardless.

cvParams.put("dataset", "conll2002");
cvParams.put("numTrainIterations", "15");
cvParams.put("maxFolds", "5");
cvParams.put("updatorClass", "SWVMUpdator");
cvParams.put("topK", "3");
cvParams.put("maType", "aggresive");
cvParams.put("jjType", "single");
cvParams.put("gammaCalculationMethod", "uniform");

cvParams.put("gammaObjectiveType", "MAXIMIZE");
cvParams.put("gammaCalculationBeta", "1");

Citation

The Structured Weighted Violation MIRA

[1] D. Ringel, R. Dror, R. Reichart The Structured Weighted Violation MIRA

@misc{ringel2020structured,
    title={The Structured Weighted Violations MIRA},
    author={Dor Ringel and Rotem Dror and Roi Reichart},
    year={2020},
    eprint={2005.04418},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

Contact: dorringel@cs.technion.ac.il