Attribute Mapping

This page separates three different cases:

• attributes that map directly
• Google names we preserve as aliases
• attributes where Respectify offers only an approximation

Direct Equivalents

These map directly to a Respectify attribute with the same main purpose.

Perspective name	Respectify attribute	Notes
TOXICITY	toxicity	Hostility, contempt, demeaning tone
SEVERE_TOXICITY	severe_toxicity	More extreme abusive content
IDENTITY_ATTACK	identity_attack	Attacks tied to identity
INSULT	insult	Personal put-downs and demeaning language
PROFANITY	profanity	Swearing and crude language
THREAT	threat	Threatening language
SEXUALLY_EXPLICIT	sexually_explicit	Explicit sexual content
INCOHERENT	incoherent	Meaning is hard to recover
INFLAMMATORY	inflammatory	Provocative, escalation-seeking rhetoric
SPAM	spam	Promotional, junk, or irrelevant content
FLIRTATION	flirtation	Flirtatious language

Preserved Aliases

These Google names are accepted on the compatibility endpoint and echoed back in compatibility responses (ie preserved in the response) when you request them. Internally, we use the Respectify attribute when scoring them.

Perspective name	Respectify attribute	Notes
ATTACK_ON_AUTHOR	identity_attack	Preserved alias
ATTACK_ON_COMMENTER	identity_attack	Preserved alias
TOXICITY_EXPERIMENTAL	toxicity	Preserved alias
SEVERE_TOXICITY_EXPERIMENTAL	severe_toxicity	Preserved alias
IDENTITY_ATTACK_EXPERIMENTAL	identity_attack	Preserved alias
INSULT_EXPERIMENTAL	insult	Preserved alias
PROFANITY_EXPERIMENTAL	profanity	Preserved alias
THREAT_EXPERIMENTAL	threat	Preserved alias
OBSCENE	profanity	Preserved alias with a close semantic match

Approximations

These are accepted for migration convenience, but they are not exact one-to-one replacements.

Perspective name	Respectify attribute	Notes
LIKELY_TO_REJECT	toxicity	Approximation only. Google's original signal was publication-policy specific.
UNSUBSTANTIAL	incoherent	Approximation only. Low-value content and incoherence overlap, but they are not identical concepts.

If you relied heavily on LIKELY_TO_REJECT or UNSUBSTANTIAL, please review your thresholds after migration instead of assuming parity.

Additional Attributes In Respectify

These are available in Respectify's scoring model.

Some of them overlap with Google/Jigsaw's experimental bridging work around constructive conversation. They are not part of the core Perspective attribute set, and migration code that only used the usual moderation attributes will not depend on them.

Respectify attribute	What it tries to capture
reasoning	Structured thinking, evidence, justification
curiosity	Openness, questions, willingness to learn
nuance	Complexity, qualification, more than one side
compassion	Empathy and care for other people
constructiveness	Whether the comment helps move discussion forward
respect	Dignity and basic regard, even in disagreement
personal_story	Relevant lived experience
affinity	Shared interests, rapport, common ground

Practical Advice

• If your existing system only uses TOXICITY, INSULT, and PROFANITY, the most common attributes, the compatibility path should feel straightforward.

• If you used experimental Google names, keep using those names when migrating.

• If you depended on less common attributes, treat the first migration as something to review, not just a swap in APIs or endpoints. That is, test the responses you get and see if they are indeed measuring what you need to measure, or if the resulting scores need some adjustment to match your previous scores. Respectify aims to give approximately similar values in its returned results, but as a different system, its scores will not match exactly. Thus if you have thresholds, say, requiring >= 0.75 for one attribute, review those using test data once you have switched.

• Migration steps

• Compatibility details

• or back to the Perspective Compatibility Overview