Custom Audio Moderation Rules Filling Guide

To enable the AI audio moderation model to accurately identify specific risks in your business scenario, it is recommended that you follow the guidelines and tips below when filling in custom moderation rules. Clear, specific, and logically rigorous rule descriptions can significantly improve moderation accuracy.

I. Core Writing Principles

When writing rules, follow the logic of "Scenario + Subject + Specific Behavior" and clearly define the boundaries between "Prohibited" and "Allowed".

1. Specify Concrete Behaviors, Avoid Abstract Descriptions

AI struggles to understand vague adjectives (e.g., "prohibit violations," "prohibit vulgarity"). Translate rules into specific behavioral descriptions.

• Not Recommended: Prohibit streamers from randomly advertising.
• Recommended: In a live-streaming sales scenario, prohibit streamers from verbally guiding viewers to add personal WeChat, QQ, or join third-party fan groups for private transactions.

2. Set Exemption Conditions (Whitelist)

When certain sensitive words are part of normal communication in a specific context, be sure to state the "exceptions" to reduce false positives.

• Recommended: Prohibit the use of insulting language to abuse others. Exception: In game live streams, exclamatory interjections from a streamer due to a game loss (not targeting a specific individual) are not considered a violation.

3. Provide Typical Phrase Samples

Although AI has semantic understanding capabilities, providing typical industry violation phrases (samples) helps the AI lock onto targets faster.

• Recommended: Prohibit inducing minors to send gifts. Reference Phrases: "Kids, go get your parents' phones and scan this code," "Don't tell your parents," "Babies who haven't started school yet, send me a gift."

II. Scenario-Based Rule Template Library

You can directly copy the following rule templates into the input box based on your actual business type, and modify or supplement them according to your needs.

1. E-commerce Sales & Live Shopping Scenario

Key Controls: Prohibited traffic diversion, violations of advertising laws, false promises.

1. Prohibit Private Traffic Diversion (Lead Generation): Strictly forbid streamers from verbally sharing phone numbers, WeChat IDs, QQ numbers, or using reasons like "check my profile," "join the fan group," "get internal materials" to guide viewers off-platform for private transactions.
2. Prohibit Absolute Terms: Strictly forbid the use of terms violating advertising laws, such as "national level," "highest level," "number one," "exclusive," "unparalleled," etc., to describe products.
3. Prohibit False Promises: Strictly forbid promising "full refund if ineffective" (unless supported by official platform policy), exaggerating product efficacy (e.g., weight loss, medical effects), or implying the sale of high-end counterfeits or replicas.

2. Talent Show, Appearance & Social Dating Scenario

Key Controls: Pornography and vulgarity, soft-core ASMR, inducing non-standard transactions.

1. Prohibit Audio Pornography & ASMR: Strictly forbid streamers from making sounds with strong sexual implications, such as moaning, heavy breathing, ear licking, sucking, or performing ASMR. Focus on identifying rapid breathing sounds and sounds simulating sexual acts.
2. Prohibit Disseminating Pornographic Inducements: Strictly forbid using "benefits," "private photos," or "short videos" as bait to induce users to send gifts or add friends in exchange for pornographic or vulgar content.
3. Prohibit Vulgar Language Harassment: Strictly forbid discussing sexual organs, sexual positions, sexual experiences, or engaging in sexually suggestive verbal teasing of viewers.

3. Game Boosting & Esports Live Streaming Scenario

Key Controls: Abusive attacks, prohibited boosting transactions, political speech.

1. Prohibit Malicious Abuse: Strictly forbid personal attacks, character defamation, or regional discrimination against teammates, opponents, or viewers. Note: Normal tactical shouting during gameplay is not a violation.
2. Prohibit Prohibited Transaction Promotion: Strictly forbid promoting services like boosting, rank selling, cheats (scripts/hacks), or cheap top-ups via voice chat.
3. Prohibit Political or Terrorist Speech: Strictly forbid discussing sensitive political figures, national policies, or spreading terrorist or extremist ideologies.

4. Financial Management & Stock Securities Scenario

Key Controls: Romance scams, illegal stock recommendations, false return promises.

1. Prohibit Guaranteeing Returns: Strictly forbid making absolute return promises on financial products, such as "guaranteed principal and interest," "risk-free profit," or "100% annualized return."
2. Prohibit Illegal Stock Recommendations: Strictly forbid uncertified individuals from calling themselves teachers or experts to recommend specific stock codes, or guiding users to join "internal stock trading groups" or "VIP real-time trading groups."
3. Prohibit Inducing Loans: Strictly forbid promoting unlicensed lending platforms or inducing students or low-income individuals to take out loans for investment.

5. Emotional Counseling & Live Call-in Scenario

Key Controls: Negative values, feudal superstition, psychological control.

1. Prohibit Spreading Negative Values: Strictly forbid disseminating PUA techniques (psychological control of partners), "romance scam" techniques, or promoting extreme materialism, objectification of the opposite sex, etc.
2. Prohibit Feudal Superstitious Activities: Strictly forbid inducing users to pay for fortune-telling, physiognomy, divination, or rituals, or scaring users into paying money by claiming they will face disasters.

6. Minor Protection (Universal High-Risk Rules)

Key Controls: Inducing minors to tip, privacy infringement.

1. Prohibit Inducing Minors to Spend: Once a voice is clearly identified as belonging to a minor, or a user claims to be a student/minor, strictly forbid streamers from soliciting gifts, cash, or payment passwords.
2. Prohibit Infringing on Minors' Rights: Strictly forbid verbally intimidating, deceiving, or asking minors for private information such as home addresses and school names, or engaging in sexual harassment.

III. Advanced Tips: Identifying Industry Jargon and Implicit Expressions

If your business scenario involves a significant amount of "jargon" or "code words," create a dedicated [Jargon Dictionary] module in the rules to explicitly tell the AI the actual meaning of these words.

Filling Example:

... (Fill in regular rules here) ...

Please pay special attention to identifying the following industry jargon and implicit expressions, which are considered violations:

1. Terms for Money: Mi, W, Da Bu Liu, Ruan Mei Bi, Te Chan, Lollipop (in specific contexts referring to funds).
2. Terms for Contact Information: Green Software, V, Penguin, House Number, Secret Code.
3. Terms for Prohibited Categories: Spinach (gambling), Leaves (drugs), Balloon (laughing gas), Go Ashore (handling overdue online loans).

IV. Frequently Asked Questions (FAQ)

Q: Is a longer rule always better? A: No. Rules should be concise and accurate. Excessive redundant descriptions can interfere with the AI's judgment logic, but necessary definitions (e.g., clearly defining what a "competitor product" is) should not be omitted.

Q: How do I know if a rule is effective? A: It is recommended to monitor the subsequent moderation logs after saving the rule. If you find missed violations (violations not caught) or false positives (normal content flagged), fine-tune the rule description based on the actual case, adding or removing constraints.

Q: Can I write multiple rules together? A: It is recommended to use numbering (1. 2. 3.) to state rules separately, with each rule corresponding to an independent violation scenario. This helps the AI understand the logical relationships more clearly.