Skip to content

Tools to prepare linguistic surveys for Amazon Mechanical Turk

Notifications You must be signed in to change notification settings

mitcho/turktools

Repository files navigation

turktools

Tools for preparing linguistic surveys for Amazon Mechanical Turk (AMT), as described in Erlewine and Kotek (2016). Website: turktools.net

Features

  • Simplifies the full workflow for Turk-style experiment construction: randomized counterbalanced lists, local experiment simulation, and results in a format optimized for analysis
  • Template skeletons which maximizes template reusability
  • Cross-platform (Windows, Mac, and Linux), requiring only Python (2.6 or 2.7)
  • Easy-to-use tools designed for non-programmers
  • Experiments can be used without Mechanical Turk, with the use of turkserver

Usage

To download everything here, click on the "Download ZIP" button on the right sidebar. The following tools are included:

  • Sample HTML skeletons
  • templater.py
  • simluator.py
  • lister.py
  • decoder.py
  • analysis.r: sample R analysis code

We have additionally provided sample item files for three of the skeletons in the Examples folder.

If you would like to host Mechanical Turk-style surveys on your own server instead of using Amazon Mechanical Turk, this can be done using turktools' sister project, turkserver.

The Python tools described here require Python 2.6.x or 2.7.x, available here. Here is how to execute a Python script (here templater.py) in different platforms:

  • UNIX shell: Move (cd) to the directory that contains the script and execute python templater.py.
  • Mac OS X: Right-click on the file in the Finder and choose Open With… > Python Launcher.
  • Windows: Double-click on the file.

We recommend saving all of the files you create for a given experiment in the same folder, dedicated to that experiment.

In the remainder of this section, we will describe the usage of these tools. For a more detailed description of the tools and a proposed workflow, please read:

Erlewine, Michael Yoshitaka and Hadas Kotek (2016). A streamlined approach to online linguistic surveys. Natural Language & Linguistic Theory 34:481–495. (Preprint)

Skeletons

A skeleton is an HTML file that contains substitution tags that will be filled in by the Templater to create an HTML template that can be uploaded onto AMT. The following template skeletons are currently included in turktools, each corresponding to a different experimental design:

  • binary.skeleton.html: grammaticality judgment, yes-no task
  • slider.skeleton.html: grammaticality judgment, gradient judgment task
  • likert.skeleton.html: grammaticality judgment, Likert scale task
  • magnitude-estimation.skeleton.html: grammaticality judgment, magnitude estimation task
  • constant-sum.skeleton.html: grammaticality judgment, constant-sum scale task
  • image-choice.skeleton.html: picture selection task
  • binary-image.skeleton.html: picture judgment, yes-no task
  • binary-audio.skeleton.html: audio stimuli, yes-no task
  • completion.skeleton.html: word-completion task
  • sentence-choice.skeleton.html: sentence-completion task
  • completion-menu.skeleton.html: completion task with drop-down menu
  • audio-transcription.skeleton.html: audio transcription task

The skeleton contains substitution tags, wrapped in double curly braces, i.e. {{…}}, will be filled in by the Templater. Read more about supported substitution tags.

The provided skeletons all share the same basic structure: they include an experiment code, an instruction block with practice items, a consent statement, an items block, and demographic questions. At the bottom is a JavaScript counter to help participants ensure that they have answered all the questions in the survey. The items block, beginning with {{#items}} and ending with {{/items}}, contains one sample item of the shape that all items in the survey will take. The item block will contain {{field_n}} tags, corresponding to different fields in your items. (See the items file section below for more on fields.) When a template is created out of the skeleton, the items block will be duplicated as many times as there are items in your survey.

Choose an appropriate skeleton for your experiment and edit it in a text editor such as Notepad++ (for Windows) or TextWrangler (for Mac), or create your own to accommodate a different type of experiment. Change the instructions, including any practice items, and add your consent statement and contact information for the experimenters. Everything about the experiment's layout, design, and presentation can be changed by modifying the skeleton.

Templater

The Templater will take the skeleton you have edited and turn it into an HTML template that can be uploaded onto AMT.

The Templater will ask for (a) the file name of your skeleton, (b) the total number of items in your survey (including all experimental and filler items but not practice items, which are coded in the skeleton), and (c) a survey code: any letter-number combination you choose. The Templater will then generate a template from your skeleton.

The items file

Lister creates randomized items lists out of an items file that must be formatted a certain way.

Each item in the raw items file has two parts, each beginning on a new line.

  • The item header. The item header consists of four components, separated by space:
    • the symbol #;
    • the name of the section (e.g., target, filler);
    • the item set number within the section (we recommend 1, 2, 3, …);
    • the condition name
  • The item body. The item body consists of fields, each on its own line. Line n corresponds to the text that will be substituted for the text {{field_n}} in your HTML template. A field may specify a sentence to be judged, choices for completions, a picture, an audio file, a context, a comprehension question, etc.

Condition names are not constrained in any way, but we recommend a naming convention where the values for each of your experimental factors is listed and separated from other values by a hyphen:

<factor 1 value>-<factor 2 value>-<factor 3 value>-...

The Decoder tool will create separate factors from the Condition names when it is run. Note that all items in a given experimental section must have the same number of conditions, but it is not necessary to keep the same names. As long as each item set has the same number of conditions, you can use different names across different item sets. The Lister will produce a warning when it is run, but you may choose to proceed.

Each item must specify information for at least as many fields as there are in the HTML template that you'll be using. It is possible to have hidden fields that do not correspond to fields in the template. Those fields will not be shown to participants but they will be carried over to the randomized file and to the results file. You can use hidden fields to specify expected correct answers to fillers or comprehension questions, and once you have a results file it will be easy to calculate accuracy for your participants.

You can find sample items file that match the binary.skeleton.html, the binary-image.skeleton.html and the completion.skeleton.html in the Examples folder.

Lister

lister.py takes an items file and turns it into a xxxxxx.turk.csv file with randomized lists of items, that can be uploaded onto AMT. The Lister creates Latin Square counterbalanced randomized lists from the items that it is given.

Once it is run, lister.py will ask for (a) the name of your raw items file, (b) which sections should be treated as fillers, (c) how many filler items you would like placed between each target item, (d) how many filler items you would like placed in the beginning and end of the experiment, (e) how many lists you would like to create, and (f) whether or not you would like the reverse of each list to also be created, to help reduce any ordering effects that may occur.

Note that as a default, no section is singled out to be used as fillers. You may to designate one or more sections as fillers and all the other sections will be treated as targets. Fillers and targets will be randomized separately and then combined according to the conditions you specified.

Simulator

Once you have created an HTML template and a randomized xxxxxx.turk.csv items list file for your experiment, you are in principle ready to upload your experiment onto AMT. However, before doing so, we recommend simulating at least one of the lists in your experiment using simulator.py. The Simulator will ask you for (a) the file name of your template, (b) the file name of your xxxxxx.turk.csv items list file, and (c) which list you would like to simulate. It will create a simulation HTML file that can be opened using your web browser and will contain the experiment as it will be seen by your participants.

We recommend simulating the survey to verify that it does not contain any errors: that the buttons or menus in the items work properly, that the counter is working, that all fields are shown properly and that hidden fields are not showing. We also recommend completing your own study at least once. This will give you a good idea of how long it takes to complete your study (which will also help determine payment on AMT) and whether you can detect other strategies for completing your study that are compatible with satisfycing behavior - that is, a strategy for solving the questions you are asking that does not involve performing the linguistic task that you are interested in.

Decoder

After uploading and running your survey, AMT will produce a raw results file which you can download from the Manage tab. The file will have a name like Batch_999999_result.csv. Save that file in your experiment's folder and then run decoder.py to convert this file into a format that can be easily read and analyzed in statistical software such as R.

The Decoder will ask for the results file name and will produce a decoded xxxxxx.decoded.csv results file that contains one row for each item in each submission. The metadata about the submission (including AssignmentId, SubmissionStatus, WorkerId, WorkTimeInSeconds and answers to demographic questions) will be duplicated across all rows within an individual assignment, so each of these values will be in as many rows as there were items in your experiment. For each item, information about the Section, Condition, Item, List, PresentationOrder and and ratings or answers to questions are logged. The Decoder also creates columns for factors in the analysis based on your condition names, following the convention described above.

analysis.r

As past of turktools we have provided a sample analysis script for R that takes care of basic data filtering and preparation for statistical testing. The script implements several exclusion criteria and assists in basic data visualization. It also produces aggregate count data for your results. The script does not attempt to implement statistical tests, as those depend on the experimental design that you are using and on your research questions.

At the end of the day...

By the time you're done preparing, running, and analyzing an experiment, your folder for that experiment will include:

  • an HTML skeleton (an edited version of the HTML skeletons provided here, or created on your own),
  • an HTML template (created by templater.py from your skeleton),
  • a text file with your raw items
  • a xxxxxx.turk.csv randomized item lists file (created by lister.py),
  • a simulation of one or more of your lists (created by simulator.py),
  • a raw results file (downloaded from AMT or from turkserver),
  • a decoded xxxxxx.decoded.csv results file (created by decoder.py), and
  • an analysis script (possibly based on the sample analysis.r code provided here).

Known issues

  • Issue #2: If the results CSV file from Turk is modified and saved in Excel, decoder.py will not be able to read it. This has to do with the line-endings which are used by Excel's CSV output.

Technical information

Design goals

turktools is designed with the non-technical user in mind. Therefore, the following principles are adopted in its design and development:

  • Be portable: each tool is a stand-alone script, and can be moved or copied to a different filesystem location and continue to function.
  • Be graceful: catch failures and present useful warnings and errors to the user.
  • No dependencies: just Python, out of the box.

The current development target is Python 2.6 and 2.7. Python 3 support would be great in the future.

Note that an unfortunate consequence of the portability goal is to explicitly eschew a shared code library, forcing code duplication across different scripts, in violation of DRY. The current approach is to at least ensure the integrity of duplicated code across tools, by subjecting them to the exact same tests. (Perhaps a build tool will be used in the future to cut down on such redundancies.) Note also that the no dependencies goal and Python 2.6 target means that nice libraries like argparse cannot be used.

Contributing

Contributions are welcome! Bug reports, feedback, documentation improvements, and code changes are all welcome forms of contributions. Thank you (in advance) for making turktools better for everyone.

Bug reports and feature requests:

New bug reports and feature requests can be added on the turktools issue tracker. Please check whether your issue is already reported by someone else before opening a new issue. You must be logged into GitHub to create an issue.

Contributing code:

turktools is developed on GitHub. The best way to hack on turktools is to open a GitHub account, fork this repository, and modify your own "fork" of the turktools. To submit changes, you can then initiate a pull request. Within reason, pull requests should include new test cases, in order to avoid later regressions.

Contributors should be familiar with the technical design goals above.

Documentation:

The turktools website is where we post additional documentation. The code for this site is on the gh-pages branch of our GitHub repository and therefore can be edited via pull request, described above. Feel free to contribute any materials there that you think may be helpful to a broader audience.

Changes to this README file can also be made as code changes.

Testing Test Status

turktools includes unit tests using the Python-standard unittest library. Tests can be run by running python tests.py. With the coverage module installed, run coverage run tests.py and then use coverage report -m to see a code coverage report.

Citation

If you use turktools or turkserver we ask that you cite the following paper:

Erlewine, Michael Yoshitaka and Hadas Kotek (2016). A streamlined approach to online linguistic surveys. Natural Language & Linguistic Theory 34:481–495.

The MIT License (MIT)

Copyright (c) 2013--2017 Michael Yoshitaka Erlewine [email protected] and contributors

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Code contributors

  • Hadas Kotek
  • Erin Olson
  • Chung-hye Han and Anoop Sarkar

About

Tools to prepare linguistic surveys for Amazon Mechanical Turk

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published