/labeling-challenge

Primary LanguagePython

labeling-challenge

About

The goal of this challenge is to get experience doing quality control for labeled data.

The original source of this data and the project is a UPenn class: http://crowdsourcing-class.org/. Highly suggest checking it out to learn more about crowdsourcing / data labeling!

Getting started

0. Set up a pyenv virtual environment (recommended)

Note: it is not necessary to use pyenv, but you do need python 3.6.

Follow this tutorial to install pyenv and learn about it:

https://amaral.northwestern.edu/resources/guides/pyenv-tutorial

Then create a virtual env for this project:

pyenv virtualenv 3.6.5 labeling-challenge
pyenv activate 3.6.5/envs/labeling-challenge

1. Install requirements

Run the following:

pip install -r requirements.txt

2. Get familiar with the data

This data is from a real mTurk project for Adjectives and Attribute Matching.

First, take a close look at the instructions that were provided to the labelers:

Instructions pt1

Instructions pt2

Now inspect the raw data file, raw_data.csv.

A few things to note:

  • You can get a unique id for the worker in column WorkerId
  • Lifetime approval rate: percentage of times requester has approved work of this mTurker across all tasks
  • Input.attr_id is the unique id for the attribute, Input.adj_* are the adjectives and Answer.adj_* the labeler's answers
  • If the labeler answered 'No' or 'not an adj', these are both listed as no in the dataset

3. Run the starter code

Open summarize_labels.py and take a look.

Then try to run it:

python summarize_labels.py

And inspect the answers it produces by opening summarized_data.csv.

How good are the labels? Run

python evaluate_results.py

4. Can you do better?

Write your own label summarization algorithm in summarize_labels.py.

Want a hint?
Think about how you can assess whether certain labelers are reliable or not.

You could also think about whether 50% is the right threshold to use.

Want another hint?
Columns Input.adj_11 through Input.adj_16 have known ground truth. 11-15 are True and 16 is False. How can you use this to evalute the labelers?
One last hint!
Come up with a "reliability score" for the labelers by assessing their performance on columns 11-16. Predict the label based on the weighted average of scores, not the simple average. You can also consider dropping unreliable labelers.