Sample CodaLab competition for SemEval

This is a sample competition bundle for CodaLab competitions (https://competitions.codalab.org/) designed as a starting point for SemEval task organizers.

Understanding CodaLab competitions

These descriptions and tutorials on CodaLab are a good place to start:

• Tristan Miller's tutorial: https://www.hse.ru/data/2017/05/31/1171931089/CodaLabCompetitions.pdf
• Isabelle Guyon's tutorial: https://www.youtube.com/watch?v=mU1yEEMrMvY
• CodaLab Competition Roadmap: https://github.com/codalab/codalab-competitions/wiki/User_Competition-Roadmap
• CodaLab Running a Competition: https://github.com/codalab/codalab-competitions/wiki/User_Running-a-Competition

Understanding this sample competition

This sample competition looks for submissions with a single file, 'answer.txt', containing the text "Hello World!". If it finds it, it gives it a score of 1, otherwise a score of 0.

To understand how a CodaLab competition bundle is created, take a look at the Makefile. In short, you need to create a ZIP file that includes all the YAML and HTML files in the competition directory, as well as the zipped-up scoring program and reference data. To do this automatically, you can run:

make competition.zip

That ZIP file can then be used to create a new competition at CodaLab: https://competitions.codalab.org/competitions/create. (You will need to register and create an account first.) You may want to upload this competition with minimum (or no) editing first, just to get a feel of how a competition bundle works.

Once the competition has uploaded successfully:

• Note the secure url the system provides below the task title. This is the competition url that can be shared with others while you are still setting up the competition.

• Explore the admin features at the top such as "Edit" and "Submissions". (Note that the edit button takes a few seconds to load the page.)

• Try uploading a submission. You can generate a sample submission at the command line by running make submission.zip, which, as you can see in the Makefile, makes a ZIP file containing just the answer.txt file. Then in your CodaLab competition:

  • Click on "Participate". Click on "Submit / View Results".
  • Enter a suitable description for each submission in the text box provided.
  • Click on the "Submit" button to upload the file.
  • This should execute the evaluation script on the submission and store the results.
  • You can monitor the progress on the "Submit / View Results" page by clicking the "+" beside your submission and then clicking on "Refresh status".
  • Your submission should move from "Submitted" to "Running" to "Finished".
  • When it is "Finished", explore various files generated on running the evaluation script. For example, text sent to STDOUT will be available in "View scoring output log".
  • Click on "Results" to see the leaderboard.

Customizing this sample competition

You will need to customize many things in this sample competition to make it appropriate for your SemEval shared task

competition.yaml

This is the main meta-data for your task.

• Edit the title and description.

• Modify the leaderboard.

  • Change "correct" to the name of your evaluation metric (e.g., fscore, accuracy)
  • Change "numeric_format" for the required digits after the decimal point. For example, using 2 will show two digits after the decimal.
  • If you will have multiple evaluation metrics, add a similar block for each of them.
  • You will need to make sure that your scoring program outputs these same names.

• Add extra phases if needed. If you're not sure whether you need extra phases or not, start a discussion on [email protected].

Other competition files

The other competition files are all referenced by competition.yaml and are included in the bundle you upload. If you change any of their names, be sure to also change their names in competition.yaml.

• Replace the logo file with the logo for your task.
• Add details to each of the webpages.

The scoring program

• Replace the sample scoring program with your task's scoring program. Your scoring program will have to follow the standard CodaLab directory structure for the reference data, the system submission, and the output scores.txt file: https://github.com/codalab/codalab-competitions/wiki/User_Building-a-Scoring-Program-for-a-Competition#directory-structure-for-submissions

• If you change the name of the evaluation script, then change it in the metadata file as well. For example, if your script is in Perl and called score.pl, then use command: perl $program/score.pl $input $output.

• Be sure that your evaluation script gives clear error messages when it fails. For example, if there is a formatting error in someone's submission, your script should explain the problem and exit with an error status (e.g., sys.exit("some error message") in Python).

• Be sure that your evaluation script is writing the output in the right format; it should print lines of the format <metric name>:<score> in the scores.txt file. For example:

    correct:1
    f-score:0.74
  

The reference data

The dev_data and test_data directories contain the reference data on which systems will be evaluated in the Practice and Evaluation phases, respectively. In the sample, the reference data for each takes the form of a single file, truth.txt. You should replace truth.txt with whatever file(s) and format(s) your scoring program expects for the reference data. If you have not yet completed preparation of data for your Practice (dev) or Evaluation (test) phases, put some sort of placeholder data here for now. You can add the real data via the CodaLab graphical interface later.