In 2019, more than 600 million voters cast their ballots in the world’s largest election. Historically, due to high illiteracy rates, India’s election commission assigned visual symbols to political parties to help voters identify their candidate of choice.
These symbols include everyday items like a flower, an elephant and even a USB stick. To mark this major event, Al Jazeera created its very first machine learning game which challenged our digital audience to draw some of India’s 2,300 political party symbols.
The results were both enjoyable and insightful and demonstrated that innovative technology can engage a global audience.
One of the big discussions in 2019 was how could newsrooms innovate with technologies such as Machine Learning (ML) and Artificial Intelligence (AI).
These technologies can yield tremendous opportunities for journalists but just like any other technology, they should be applied to the right story.
As Al Jazeera, one of our key considerations when producing data-driven stories is to ask ourselves how we can amplify the cultural and human stories from the global south to a global audience.
Following its launch, this interactive game quickly became one of our highest engaged stories on our website as well as on social platforms specifically Facebook. Tens of thousands of users from all over the world competed with one another to draw a variety of symbols ranging from an elephant to a bicycle. Many viewers from India challenged one another to draw their party symbols and to learn the meaning behind these symbols.
Organisationally, and for the region, this opened up the doors to experiment more with interactive storytelling using machine learning. This is vital given that these new techniques are often considered out of reach for most news organisations.
Produced and developed over the course of three weeks by one data journalist using open-source technologies, this story sparked a conversation in the newsroom about the adoption of new interactive techniques to storytelling.
Under the hood was a convolutional neural network trained using Keras, an open-source neural-network library written in Python. The data itself was the open-source Quick Draw dataset which contains around 50 million drawings from 345 classes. https://quickdraw.withgoogle.com/data
For our story, we utilised 20 classes which corresponded to India’s biggest political parties. Once the model was trained we utilised Google’s TensorFlow.js. which allowed us to run the entire deep learning module directly in the user’s web browser.
This made the application very accessible and meant that users did not need to download anything to start drawing on their desktop/mobile.
What was the hardest part of this project?
The most difficult part of the process was making the technology seamless in order for users to just focus on the storytelling and user experience. The success of the project opened up huge possibilities for future machine learning based stories which are expected to become an integral part of the data-driven newsroom.
What can others learn from this project?
The biggest takeaway from this project is that machine learning can have its place in every data journalist’s toolbox. As algorithmic accountability and machine bias is increasingly playing a role in journalism it’s important for data journalists to experiment with new storytelling techniques that captivate a global audience’s attention.