Case study 05

Anti-dirt

Anti-dirt is a key part of the company's risk services. For every product line, it manages, evaluates, detects, extracts and applies sensitive words (political terms, vulgar language and so on) in user posts.

  • Rebuilt the process of ByteDance's sensitive-word platform, starting with its most used feature, Add keyword.
  • Interviewed four regular users and shortened the add-keyword flow from 10.23 to 8.4 minutes (−18%).
  • After the redesign, users rated the platform 4.94 / 5.
Role
Product strategy · User research · User experience
Tools
Figma · Lark
Timeline
Mar 2021 – Oct 2022 (at )
−18%Time to add keywords (10.23 min → 8.4 min)
4.94 / 5User rating
270Unique visitors (UV)
Encapsulated callAdjusts the volume sent for reviewReview page dataSmart assignmentReal-time review (content)Machine review (models)Manual review (review queue)Gets the featuresof the post.✅ User ID✅ Post time…Matches a specific strategy:❓ Check against a wordlist?❓ Does it need to be sent for review?❓ Does it need special handling?…FeaturesyesUser postMatch the review processGet risk featuresDecide whether to sendto manual reviewCreate the data task packageto send for reviewTask assignmentPublish a modelSelect a modelReview modelKeyword selectionDeduplicated sample selectionAdjust the review thresholdSmart hintsHeat map, highlighted words,frame samplingQueue routingMerged queues

What Anti-dirt does

Text-related content is usually handled with NLP models. But training an NLP model takes time, and models make mistakes. So the idea behind Anti-dirt is to find sensitive words with text matching that is fast, accurate and efficient.

Search results page saying that, according to the relevant laws and regulations, the results are not displayed The keyword cannot be searched

A product review where the first word is replaced by two asterisks The keyword is masked

An encyclopedia article about an event A video whose comments ask what happened Posts containing the keyword are taken down

Two kinds of users

Anti-dirt has a stable group of users and a history of more than three years. Its users fall into two main groups: the people who create sensitive wordlists, and the people who use them.

Wordlist creator

Operates and maintains the platform's core wordlists

Basic info

  • Education: bachelor's degree
  • Device: MacBook Pro 13-inch
  • Hours: weekdays 10:00–19:00

Work objectives and scenarios

Add keywords on demand

  • Collect wording needs from various channels (the OWLS platform, online bad cases and so on) and add sensitive words.
  • Evaluate how the content and the matching method affect what is online, to guarantee the quality of the words.

Daily wordlist work

  • Day-to-day consulting, proposals and promotion for onboarding new businesses.
  • Manage feature-level access permissions

Optimize the wordlists

  • Based on online data, optimize and update the details of the words in a wordlist: tags, exemption words, combination words and so on.
  • Special needs for wordlists (June *, the 20th * and so on)
  • Optimize the wordlists according to business needs

Wordlist user

Safety strategy development in the content safety team

Basic info

  • Education: master's degree
  • Device: ThinkPad
  • Hours: weekdays 10:00–19:00

Work objectives and scenarios

Get permission for business wordlists

  • Apply to the wordlist manager for permission

Monitor online data and collect feedback

  • Monitor how the wordlists perform online and collect user feedback about sensitive content
  • Keep the chance of entering the review process low

Online incident investigation

  • Project review, retrospective
  • Based on misses and mistakes, give requirements to the wordlist managers.

Planning the platform redesign

Based on user feedback, I worked out the redesign strategy for the next step.

The previous version’s home page: a dense list of wordlists grouped by product line

The previous version:

  • Feature structure: for some product lines, features were hard to use because of safety restrictions.
  • How to use it: new users found it hard to understand how the platform worked.
  • Features nobody knew: some features had been built long ago, and nobody could explain them any more.

Restructure

Problems of the previous version

Design solutions

Some processes are blocked

  • Special concepts are not explained clearly.
  • Required and optional form fields are unclear, so users often have to redo steps.
  • Some error messages are missing, which adds extra cost to understanding the process.
  • Reduce the cost of understanding the copy and unify the copy across the platform.
  • Strengthen feedback on actions and add error-preventing design to improve efficiency.

Low operating efficiency

  • Features are weakly separated and flattened.
  • User paths are confusing.
  • The information hierarchy is weak, and important information does not stand out.
  • Feedback on actions is lacking.
  • Provide general guides and hints so users have more control over what they do on a page.
  • Shorten the paths and remove unnecessary actions.
  • Optimize how the interface is displayed and bring information together.

No consistency between pages

  • Components of the same kind are used inconsistently.
  • Unify the platform's visual elements and interaction patterns.

Start with the most frequently used module

The previous “edit sensitive word” form: every field stacked on the left, the right side empty

“Add keyword” is the most used feature.

But the previous version had been built in whatever form the back-end engineers came up with, without any research into users’ habits.

User research showed that optimizing this feature would bring the most value. I interviewed four regular users and collected their feedback.

😭 “The whole process is too long for me. Every time I add keywords there are 500 to 1,000 words, and I want to finish in five minutes. But I always have to fill in lots of form fields.”

🤬 “The tagging feature is really hard to use. The tree selector has many layers, and I can’t tell what is under each layer, so I have to expand it again and again.”

🤨 “The right side of the add-keyword page is completely blank, and all the content is squeezed into the left. I don’t like that…”

😃 “Overall it’s fine… I’ve worked here for a year, so I’m used to it.”

What I did to solve the problems

  • First, I let users keep doing what they were already doing, to lower the cost of learning how to use the feature.
  • Next, I made filling in the form less stressful and more comfortable.
  • Then I designed the whole process around their daily work, changed the order of some fields, and made the steps for adding keywords clear.
  • Finally, I improved interactions that are small but affect efficiency.

The new Add keywords flow: three steps (info setting, preview, adding method) with keyword groups, bulk input and remarks

👏🏻👏🏻👏🏻 After launch, I received positive feedback from users.

A chat message from a user: “I quite like the new sensitive-word platform, it’s much more convenient. A good review from me”, with reactions from colleagues

Four more feature improvements

Variant word generator

The word expansion tool generating variant words from a keyword

Variant words are words that look or sound similar to the original, for example “习近平” and “刁近平”.

Giving users a simpler, more direct solution:

  • Help the user create variant words in a few simple steps
  • See the created results directly
  • Select all, or deselect, with a simple click

Approval center

The approval center: task categories, a number badge for pending tasks, and the process of a task

Visual hints and separated information:

  • Visually clarify the different categories of approval tasks
  • Tell users about their to-do tasks with a number badge
  • Let users see all the information of a task by hierarchy, so they can process it efficiently
  • Let users easily see where a task is in its process, and who is in charge of each node

Detection tool

The detection tool: preset wordlist settings on the left and the task list on the right

  • Designed for people with different goals (detection task manager / detection operator)
  • Lets users finish entering the data for detection as efficiently as possible
  • Supports both manual input and file import

The detection result page: hit rate, false hit rate and recall rate at the top, hit details below

The detection results page contains a huge amount of information. I organized the important information and emphasized everything users might care about.

Automatically recommending exempt words

Adding exempt words to a wordlist with automatically recommended words

In many cases, a hit in machine review is a mistake caused by missing exempt words. The result is sent to manual review, which costs manpower.

However, exempt words are currently generated mainly from the wordlist manager’s own judgement, which consumes manpower and lacks accuracy.

The “automatically recommend exempt words” feature saves the wordlist manager’s effort and removes the risk of content being sent to manual review, helping to keep the balance between machine review and manual review.

👏🏻👏🏻👏🏻 Positive feedback after launch:

Chat messages after the announcement of the exempt-word recommendation: “Amazing”, “From zero to one, that’s great”, “A feature we wouldn’t have dared to imagine a year ago”

Results after launch

User rating
4.94/5
UV (unique visitors)
270▲
Less time to add keywords (10.23 min → 8.4 min)
18%▼

Highlight

I rebuilt the process of the whole platform, strengthened the main features and toned down the rarely used ones. Overall, the efficiency of using the platform and user satisfaction went up a lot.

Lowlight

Each product line is fairly satisfied with the redesign, but some issues remain. For example, the process on the page for adding new keywords is too long, and important information gets overlooked at the preview step.

In all

The MVP version is released, but B2B products are complex and come with business background, so I need to keep talking with users to get their feedback.