Case Study · Innodata for Twitter

Twitter Spam Violation Workbench

I designed the moderator workbench to bring investigation, evidence, and annotation into one focused experience.

Semantic accuracy rate +22%

Improvement in SAR after 3 months in production, exceeding Twitter’s 10% target.

Role
Senior Product Designer, sole designer
Client
Innodata for Twitter
Collaborators
Twitter stakeholders, moderators, Innodata delivery team
Scope
New annotation workbench, designed from the ground up
Context & ownership

Accurate human annotations were critical to Twitter’s spam detection

Human moderators reviewed flagged tweets and classified violations, producing the annotations Twitter used to improve its spam detection models.

Ownership

End-to-end UX

Research, exploration, interaction design and validation.

Scope

New annotation workbench

Designed from the ground up for high-volume moderation work.

Innodata partnered with Twitter to improve the quality of human annotations used to train its spam detection models. As the sole product designer, I owned the end-to-end UX of a new moderation workbench, collaborating with Twitter stakeholders, moderators, and Innodata’s delivery team to bring investigation and classification into one experience.

Problem & goals

Fragmented investigation workflow

Moderators had to switch between spreadsheets, Twitter profiles, and questionnaires to investigate and classify tweets, repeatedly interrupting their work.

The problem

Too much effort between decisions

Every annotation required navigating separate tools to gather evidence and answer questions, adding unnecessary steps to an already demanding workflow.

The goal

Improve annotation accuracy by 10%

Create a focused workbench that brings investigation and annotation together.

Research

I documented the rules and studied how moderators worked

I reviewed Twitter’s requirements to capture the annotation rules, then studied real moderator sessions to understand how those rules played out in practice.

Twitter annotation project documentation showing rater instructions for reply qualityEnlarge
Annotation rules captured from Twitter’s requirements.

I reviewed recordings of real moderation sessions to understand how annotators moved between tasks, Twitter profiles, translation tools, and classification questions. I focused on repeated actions, unnecessary context switching, and the effort required to gather evidence before making a decision. These observations exposed workflow friction that wasn’t apparent from the annotation requirements alone.

Exploration

I mapped the workflow, then challenged it with stakeholders

I refined the workflow through stakeholder reviews, a collaborative What If workshop, and technical constraints.

General moderator workflow map with scan, completeness, prepare and explore stagesEnlarge
The general moderator workflow, from scanning a task to exploring an account.

I mapped the end-to-end moderation workflow and used it to challenge assumptions with Twitter stakeholders and Innodata’s delivery team. Our discussions helped distinguish steps required for accurate classification from unnecessary navigation and repeated manual work. The resulting workflow became the foundation for a more focused workbench, balancing moderator needs with the project’s technical constraints.

Insights

Research revealed the hidden costs of manual moderation

My screencast analysis revealed repeated interruptions in a high-volume annotation process.

Fragmented evidence

Context was scattered

Tasks, Twitter context and questionnaires lived in different places.

Repeated actions

Too many manual steps

Copying, pasting and translation interrupted the investigation flow.

Moderator wellbeing

Repeated exposure to harmful content

Moderators routinely encountered offensive or disturbing material while investigating flagged accounts.

Design decisions

Testing my own assumptions against how moderators really worked

Exploration 1

I initially optimized for convenience

Wireframe of a four-pane layout with Tasks, Homepage, History and Q&A panesEnlarge
The opportunity
Moderators sometimes needed older tweets to investigate an account.
My design
I added a History pane with tweet search and content filters.
The assumption
Fewer clicks to access historic tweets would make investigations easier.
Iteration 1

I removed complexity moderators didn’t need

Wireframe of the simplified three-pane layout with Tasks, Homepage and Q&AEnlarge
The feedback
Four panes felt too busy. Recent Tweets and Replies were usually enough.
My decision
I removed the History pane and reused the existing Tweets and Replies functionality.
The outcome
Moderators preferred the simpler layout. I avoided building search and filtering tools that added little value to their workflow.
Exploration 2

I started with a compact performance snapshot

The requirement
SAR score was Twitter’s primary KPI for annotation quality.
My design
I put the SAR score, elapsed time, completed tasks, and last saved in the header.
The assumption
A single status summary would help moderators understand their performance and progress.
Iteration 2

I made progress visible where work happens

The opportunity
Header totals didn’t show progress through long task and question sets.
My decision
I added visual progress indicators to the Tasks and Q&A panes.
The outcome
Stakeholders approved the change, giving moderators clearer visibility into task and question progress.
Design decision 3

I brought translation into the investigation workflow

The problem
Moderators occasionally needed translation to understand ambiguous or unfamiliar content.
My decision
I introduced a contextual Translate and Copy helper, available when users selected text.
Why it mattered
Moderators could understand difficult content without leaving their investigation or choosing a language.
A tweet in English offering a launch-day giveaway
The same tweet with text selected and a Translate and Copy helper shown
The tweet text translated inline into Filipino
Select, translate, keep investigating, all in context.
Final design

Investigate, understand, and classify in one workbench

Load a job before starting the investigation

Job queue
See available annotation jobs and resume saved work.
Preferences
Use the saved translation language and select how many tweets to load.
Load job
Enter the workbench with the job and preferences ready.
Record Classification dialog showing the job queue, translation language preference and number of tweets to loadEnlarge
Final Spam Violation Workbench with Tasks, the account’s Tweets and Replies, and the Q&A questionnaire with progressEnlarge
The final three-pane workbench.
  1. 1

    Investigate

    Review the account and recent Tweets and Replies.

  2. 2

    Understand

    Translate selected text inline when needed.

  3. 3

    Classify

    Answer violation questions, track progress, and submit.

The workbench supported moderators and managers

Moderators used the workbench to investigate and classify tweets. Managers used reports to track accuracy and productivity.

Moderator experience

Investigate and classify

Load jobs, review Twitter context, translate when needed, and complete violation questions.

Manager reporting

Quality and throughput

Reports covered SAR, throughput, completed jobs, and time spent, helping managers see performance patterns.

Outcomes & impact

Twitter targeted a 10% improvement in accuracy. The reported result was 22%.

Semantic accuracy rate+22%

Reported SAR improvement after three months in production, against a 10% target.

Annotation qualityMore consistent decisions

Twitter reported more consistent annotations used to support spam detection models.

Manager reportingBetter performance insight

Managers could track SAR and operational metrics to identify top and bottom performers.

Reflection

The workbench improved. Moderator wellbeing remained a challenge.

What worked

Watch people work

Screen recordings exposed friction that interviews alone might have missed.

What mattered

Challenge my own assumptions

My initial four-pane design seemed more convenient, but feedback showed that a simpler three-pane layout better supported how moderators actually worked.

What I’d revisit

Protect moderators

I’d explore ways to limit exposure to disturbing content without compromising annotation accuracy.

Working on something ambitious?

I’d love to hear what you’re building. Tell me about the problem and where you’re headed.