Tracking Hate Speech & Disinformation
My work from designing and developing a moderation software for the 2020 National Myanmar Election.
Matthew Lim, Harshil Shah, Max Karpawitch, John Britti, Arpit Mathur, Kevin Cooper, Dr. Michael Best, Saira Poonnen, Daniel Nkemelu, Reynold Kyaw
Adobe XD, Figma, Matomo, IBM Equal Access Accessibility Checker, Cognitive Walk-through, A/B Study, Semi-structured Interviews, Bootstrap, NodeJS, Angular, Express, EJS, MongoDB
UX Designer, Full-Stack Developer, UX Researcher, Product Manager, Hiring Manager
Non-profit employees of the New Myanmar Foundation and the Carter Center.
Myanmar has experienced a recent explosion in mobile internet access. Enabled by this has been a wave of disinformation, hate speech, attacks on free speech and human rights groups, and physical violence. Facebook, in particular, has been accused of allowing this content to amplify on their platform.
In conjunction with the New Myanmar Foundation, Facebook's CrowdTangle, and The Carter Center, I worked to develop interventionist technologies and teams to monitor Burmese social media during the past Myanmar 2020 national election. Read my medium post for a better story!
The Solution: Aggie
Aggie is a social media content aggregation and triage platform originally created for election monitoring. Aggie is currently being used in Liberia, Ethiopia, Myanmar, and in the United States. All features displayed are recorded off the live Myanmar instance.
Distribute Work using Batch Mode
On large teams, it is important to avoid having several sets of eyes on the same social media content. Our batch mode allows a user to pull unread reports out of a queue and prevent others from accessing those reports until they have been analyzed or returned back to the queue.
Tag, Flag, and Act on Dangerous Content
Categorize posts with customizable and searchable tags. Flag posts to elevate them to a manager. Act on posts by reporting them on their platform or responding quickly.
Hate Speech Classifier
Several members of the research team trained an NLP algorithm to assign a score to posts. The algorithm was ~80% accurate in rating hate speech with high scores. In order to support
Simplified Content View
Improving the old Aggie meant updating the core UI design of the platform, the posts themselves. By reducing space necessary for flagging, making selection larger and shifting source info to one column, users can focus more on the content of posts.
Visualize Data During and After Elections
After the Myanmar election and during we wanted users to be able to visualize trends and statistics about the social media content being tracked. We created a live dashboard in the platform so users can quickly see live data.
View Post Replies without leaving the App
Research suggests hate speech and disinformation is often spread through comments. We incorporated comments for potentially dangerous posts into our workflow and designed comment views to review comments for dangerous speech.
Before & After
Aggie was originally an election monitoring platform. See some of the changes I made to the UI to increase the usability of the platform.
The Aggie Platform has found over 4,800 incidents of hate speech, disinformation, and harassment of public officials on heavily trafficked Facebook groups.
These incidents are compiled by the New Myanmar Foundation, verified by local sources, and escalated (when necessary) to Facebook and/or to the Myanmar government. Many of these posts are not caught by Facebook's moderation team and can be incredibly popular and widespread.
We conducted research through competitive analysis, user analytics, semi-structured interviews, and cognitive walkthroughs.
I created mockups of Aggie in Figma and Adobe XD to aid feature development and receive feedback from users and stakeholders.
Software Development & Technical Support
I worked primarily on the front-end of Aggie but developed tags, multimedia thumbnails, and upgraded the system to Bootstrap 4.5.0. I also helped maintain several instances of Aggie servers.
Product & Project Management
I took a managerial role in the lab, running weekly standups and setting priorities.
Due to Covid-19 epidemic lockdowns, language barriers, international ethical concerns, nine thousand miles of physical distance, and eleven and a half hour time difference, access to users was scarce at best.
It is often said that in Myanmar “Facebook is the internet and the internet is Facebook”. Myanmar’s dependence on Facebook resulted from their Free Basics program. With this knowledge in mind, many of our UI/UX elements were based on elements of Facebook’s UI/UX.
In addition, much of the UX of the platform is based on a typical email UX. Content can be flagged, grouped, tagged. Content is shown by the most recent post date and searchable with fairly complex queries.
We used Matomo software to analyze user behavior in relation to the platform. By analyzing heat maps and user flows, we found that users spent a significant amount of time outside the platform, often used post grouping, and used customized notes.
We are also able to observe which buttons and elements do not get used at all. This allows us to better design visual hierarchies (like bluing the "Mark All As Read and Grab Batch" button).
Remote Semi-structured Interviews and
Cognitive Walkthroughs in Burmese
These are only preliminary results of the 20 hours of user studies we conducted, 2 hours for ten participants with a Burmese translator named Soe. We will conduct two more interviews on the management of the user staff and then encode the interviews with more academic rigor. These interviews covered everything from user demographics, backgrounds, and usability testing.
Hate speech & disinformation is often repeated and shared
Very often users creating groupings of posts that were the same content shared repeatedly. This creates a need for users to share posts containing hate speech/disinformation so that they can avoid taking duplicate actions.
Hate speech & disinformation is heavily contextual and requires investigation
We found that users always left the platform to investigate the posts because they would check a multitude of sources of information on Facebook. This suggests that we should support a seamless transition between the Aggie platform and social media.
Hate speech & disinformation is multi-modal and often hidden
Often hate speech and disinformation appear as text on a photo or visual/auditory content in videos. In addition, a user mentioned that hate speech and disinformation was hard to immediate discern on our platform because we cut off long text posts and hate speech and disinformation was usually in the middle or end of long posts.
Fact Checking is decentralized and a grassroots effort
Users often mentioned that the terminology of reports indicated that something was finished and said it would be better to call it "news". They also mentioned that they relied on a wide network of individuals stationed throughout the country to check on claims. Centralized news outlets were not considered reliable and unbiased and required fact-checking in themselves.
Mid-fidelity Mockups allow us to show key stakeholders, like the Carter Center, potential updates to Aggie functionality and guide software developers. We also used them to experiment with UI elements. We used Adobe XD originally, but moved to Figma due to its support of collaborative remote work.
Designing a Better Workflow
To moderate a massive flow of posts, we needed an efficient workflow. I ran a remote 2.5 hour collaborative brainstorming workshop with stakeholders. I moderated the workshop to ensure that a viable workflow was produced.
As a full-stack developer, I added four features currently in production (tags, media thumbnails, platform-specific details, and front-end comment integration) and all major visual updates (Bootstrap 3 to Bootstrap 4.5) to the front-end. My work debugging production problems in deployments lead me to form the "Vector Team", a core group of developers who were responsible for fixing any issues that came up during the weeks before and during the election.
Aggie, which is open-source software, recently surpassed 3000 commits and has reached version 3.0.0 with the addition of comment functionality.
On the day of the election (which is our night, our team stayed up with food and drink to make sure that the platform worked throughout the day of the Myanmar national election.
Product & Project Management
I was responsible for finding initial issues with the platform as well as for deciding which new major features would be added. Additionally, I helped filter and hire two undergraduate students and two graduate students for the lab. We conducted resume screens and formal interviews.
During the development process, I managed all front-end related features for myself, and the new hires. We decided to switch the lab from a conventional research workflow to a scrum workflow (Github, Trello) to better support software development.
As of October 27th, the team has archived 200 cards on the Trello board and has archived most of them. I currently lead standups for the Aggie research team.