Interfacing with a new crowd

On this page:

The crowd server is designed to be easily extensible to send tasks to other crowd systems. Each crowd is implemented as a Django app that can re-use models, views, and templates from our generic ‘basecrowd’ implementation. To add support for a new crowd to the server, you must:

  1. Assign your crowd a CROWD_NAME that can be referenced from within URLs.

  2. Create your django app with python startapp CROWD_NAME.

  3. In CROWD_NAME/, define your models. A crowd must have at a minimum four models: one for Tasks, one for Task Groups, one for Workers, and one for worker Responses. Luckily, we provide abstract implementations of all of these in basecrowd/, so simply subclass models.AbstractCrowdTask, models.AbstractCrowdTaskGroup, models.AbstractCrowdWorker, and models.AbstractCrowdWorkerResponse and you’re good to go. You can add custom fields to your subclasses if you’d like, but it’s probably not necessary.

  4. Create a new file CROWD_NAME/ This will be the bulk of your work. In the file, create a subclass of basecrowd.interface.CrowdInterface, and implement any methods necessary to support your crowd. Feel free to look at basecrowd/ for the full list of available methods, but you’re likely to need only a few, including:

    • create_task: create a new task on your crowd platform.

    • delete_tasks: delete one or more tasks on your crowd platform.

    • get_assignment_context: when your crowd platform requests a task interface, provide enough context to render your interface template (described more below).

    • get_response_context: when a crowd worker submits data, extract it from the request and put it in a format that can be saved to your models.

    Finally, create an instance of your interface as a module-level variable, e.g.:

  5. In CROWD_NAME/, register your new interface with the server:

    from interface import MYCROWD_INTERFACE
    from models import *
    from basecrowd.interface import CrowdRegistry
  6. Now all we need are templates to render the task interfaces to the crowd. Again, we provide a base template that should do almost all of the heavy lifting. Create your app’s template directories and a new template in CROWD_NAME/templates/CROWD_NAME/base.html. Inherit from our base template with {% extends "basecrowd/base.html" %}, and implement as many blocks in that template as you’d like. Our base template (located in basecrowd/templates/basecrowd/base.html) contains many inheritable blocks, but you probably need to implement only one of them: {% block get_submit_context_func %}. This block should contain a javascript function definition, get_submit_context(urlParamStrings) that takes a list of URL parameters (unsplit in the form 'key=value'), and should return a JSON object containing any context that needs to be submitted to the server with the data (e.g. the task, worker, and assignment ids relevant to this interface). See our implementation in amt/templates/amt/base.html for details.

  7. The interfaces for individual task types (sentiment analysis, deduplication, etc.) should just fit into your base.html template with no additional work. If you completely customized base.html, however, you might need custom templates for the task types as well. To create them, simply follow the steps for creating a new task type, but place your template in your crowd’s template directory (e.g., CROWD_NAME/templates/CROWD_NAME/sa.html).

  8. And that’s it! Now you should be able to use the APIs described above to create and complete tasks on your new crowd platform. If you run into trouble, take a look at our implementation of the Amazon Mechanical Turk crowd (in amt/) for inspiration.