Podcast
Questions and Answers
What are the two ways users can express their interests in TikTok's new feature?
What are the two ways users can express their interests in TikTok's new feature?
What type of prompts can users utilize to request relevant content recalls?
What type of prompts can users utilize to request relevant content recalls?
What is the purpose of the fixed first option in the survey?
What is the purpose of the fixed first option in the survey?
What happens when users enter their preferences in the freeform editing flow?
What happens when users enter their preferences in the freeform editing flow?
Signup and view all the answers
During which scenario is the positive or negative survey displayed to users?
During which scenario is the positive or negative survey displayed to users?
Signup and view all the answers
What does the fourth option in the survey allow users to do?
What does the fourth option in the survey allow users to do?
Signup and view all the answers
What primarily focuses on the request for 'More posts like this' in the i2i prompt?
What primarily focuses on the request for 'More posts like this' in the i2i prompt?
Signup and view all the answers
What do users confirm after AI processes their input in the freeform flow?
What do users confirm after AI processes their input in the freeform flow?
Signup and view all the answers
What should be done with postal addresses that belong to businesses?
What should be done with postal addresses that belong to businesses?
Signup and view all the answers
What happens when an ID number is only partially identified?
What happens when an ID number is only partially identified?
Signup and view all the answers
Which types of account numbers should be retained?
Which types of account numbers should be retained?
Signup and view all the answers
What characterizes an i2i prompt?
What characterizes an i2i prompt?
Signup and view all the answers
How should user prompts that request 'more posts like this' be categorized?
How should user prompts that request 'more posts like this' be categorized?
Signup and view all the answers
What is a guideline for handling usernames?
What is a guideline for handling usernames?
Signup and view all the answers
What type of prompt is classified as an entity prompt?
What type of prompt is classified as an entity prompt?
Signup and view all the answers
Which of the following should be discarded if there is uncertainty?
Which of the following should be discarded if there is uncertainty?
Signup and view all the answers
What is defined as the text or input that users provide to express their preferences?
What is defined as the text or input that users provide to express their preferences?
Signup and view all the answers
Which of the following describes the format of a TikTok post?
Which of the following describes the format of a TikTok post?
Signup and view all the answers
What should happen if a user prompt contains personal data?
What should happen if a user prompt contains personal data?
Signup and view all the answers
Under which condition should a user prompt be discarded for being unintelligible?
Under which condition should a user prompt be discarded for being unintelligible?
Signup and view all the answers
Which of the following is NOT a personal data identifier that needs to be discarded?
Which of the following is NOT a personal data identifier that needs to be discarded?
Signup and view all the answers
When is it acceptable to discuss public figures in a user prompt without discarding it?
When is it acceptable to discuss public figures in a user prompt without discarding it?
Signup and view all the answers
What is the purpose of the Recalled Content?
What is the purpose of the Recalled Content?
Signup and view all the answers
What happens if most of the words in a user prompt are not in the target language?
What happens if most of the words in a user prompt are not in the target language?
Signup and view all the answers
What does the User Prompt refer to in the context of content generation?
What does the User Prompt refer to in the context of content generation?
Signup and view all the answers
Which of the following is NOT an aspect that users can specify when inputting their preferences?
Which of the following is NOT an aspect that users can specify when inputting their preferences?
Signup and view all the answers
What happens when certain aspects are left as 'unknown'?
What happens when certain aspects are left as 'unknown'?
Signup and view all the answers
What defines the 'Action' aspect in content preferences?
What defines the 'Action' aspect in content preferences?
Signup and view all the answers
Which of the following is considered a 'Format' in content recommendations?
Which of the following is considered a 'Format' in content recommendations?
Signup and view all the answers
What is the role of the 'Creator' aspect in content requests?
What is the role of the 'Creator' aspect in content requests?
Signup and view all the answers
How does the model generate outputs from user input?
How does the model generate outputs from user input?
Signup and view all the answers
What is the purpose of combining different entities in content instructions?
What is the purpose of combining different entities in content instructions?
Signup and view all the answers
What is indicated by the term 'closed set' in aspect definitions?
What is indicated by the term 'closed set' in aspect definitions?
Signup and view all the answers
What does the aspect labeled 'Content Language' specify?
What does the aspect labeled 'Content Language' specify?
Signup and view all the answers
What is the maximum number of recommended posts that are generally expected to be labeled?
What is the maximum number of recommended posts that are generally expected to be labeled?
Signup and view all the answers
If the original video does not clearly convey its meaning, how should it be labeled?
If the original video does not clearly convey its meaning, how should it be labeled?
Signup and view all the answers
How is the relevancy of a recommended post determined?
How is the relevancy of a recommended post determined?
Signup and view all the answers
What label should be used if a recommended video link is not available to open and play?
What label should be used if a recommended video link is not available to open and play?
Signup and view all the answers
Which label indicates that you cannot determine if a recommended video is relevant?
Which label indicates that you cannot determine if a recommended video is relevant?
Signup and view all the answers
What does the overall accuracy represent?
What does the overall accuracy represent?
Signup and view all the answers
What should be done if a recommended post contains conflicting text and visual information?
What should be done if a recommended post contains conflicting text and visual information?
Signup and view all the answers
What formula is used to calculate overall accuracy?
What formula is used to calculate overall accuracy?
Signup and view all the answers
Study Notes
TikTok Recommendation Feature
- TikTok is developing an interactive feature that allows users to customize video recommendations based on their interests.
- Users can provide their interests through surveys or freeform text input.
Surveys
- Surveys are triggered after user interactions with a video, with two types – positive and negative – based on the nature of interaction.
-
Four options in the survey:
- Option 1: Fixed – See more/less content similar to the current video.
- Options 2 & 3: Generated based on the video – View more of this content.
- Option 4: Fixed – Allows users to express specific likes or dislikes.
Freeform Text Input
- Users enter their desired content preferences in freeform text.
- The AI model processes the user's text, translating it into recommendations the recommendation system can understand.
- The model’s interpretation is shown to the user for confirmation before refreshing the feed.
Relevant Content Retrieval
- After input, relevant content based on the prompt is retrieved.
- Two types of prompts:
-
i2i (Item-to-Item): Users seek “more posts like this” or “fewer posts like this.”
- Retrieved content is similar to the current video or the one associated with the freeform input.
-
Entity: Users specify a desired topic, like "cats," "Lady Gaga," etc.
- Content matching the user's entity prompt is retrieved.
-
i2i (Item-to-Item): Users seek “more posts like this” or “fewer posts like this.”
Labeling Guidelines
- Evaluation focuses on the relevancy of recommended content to the user prompt or the original post.
Discard Rules
- User prompts are discarded if they meet any of the following conditions:
- Non-target language: Most words are not in the target language.
- Unintelligible: User intent is unclear, even if most words are in the target language.
-
Personal data: User prompt contains any of these:
- Full name (excluding single names/characters)
- Individual phone number (excluding business numbers)
- Individual email address (excluding business addresses)
- Individual postal address (excluding business addresses)
- ID numbers
- Account numbers for personal/business accounts
- Usernames (including public usernames)
- Passwords/passcodes
Prompt Type Classification
-
i2i (Item-to-Item) Prompt:
- Users request "more posts like this" or "fewer posts like this," without specifying a specific topic.
- Examples: “More”, "Love", "Relatable", "Plus", "More posts like this lifeguard"
-
Entity Prompt:
- Users ask for specific content types, languages, formats, creators, etc.
- Examples: "More basketball videos", "less news", "More yoga videos in Spanish"
-
Entity Prompt Aspects:
- What: Specific theme/topic (e.g., Cats, Basketball videos)
- Action: Boost (more), Deboost (fewer), or remove entirely (no)
- Format: Video, photo, live streaming, ads, or unknown (if not specified)
- Content Language: English (en), Spanish (es), etc., or unknown (if not specified)
- Creator: Friends, follow, or unknown (if not specified)
Content Labeling Procedure
- Step 1: Label Prompt Type: Identify prompt as either "i2i" or "entity."
-
Step 2 (for i2i only): Label Original Video Clarity: Determine if the original video is clear and conveys a clear meaning.
- "Yes" if clear, "No" otherwise.
-
Step 3: Label Recalled Content Relevancy: Evaluate the relevancy of each recommended post to the original post or user prompt.
- "Relevant" if both text and visual information match, "Not relevant" otherwise.
- "Uncertain" if the video is blurry, unclear, or conflicting.
- "Not available" if the video link is broken.
- Step 4: Calculate Overall Accuracy: Calculate the proportion of relevant content to total recommended content: "Overall Accuracy = No. of relevant / (No. of relevant + No. of irrelevant)"
Additional Notes
- The labeling process involves evaluating 6 recommended posts in most cases, although fewer posts may be available.
- The platform automatically categorizes overall accuracy into ranges, like 0%, 1-25%, 6-49%, 50%, 51-75%, 76-99%, or 100%.
Studying That Suits You
Use AI to generate personalized quizzes and flashcards to suit your learning preferences.
Related Documents
Description
Explore TikTok's new interactive feature that personalizes video recommendations based on user interests. Learn how surveys and freeform text inputs work to tailor content suggestions effectively. This overview covers the methods TikTok employs to engage users in customizing their viewing experience.
