Attributes and languages
What openspective scores, in which languages, with which model.
Supported
Section titled “Supported”| Attribute | Languages | Model output |
|---|---|---|
TOXICITY |
en, fr, es, it, pt, tr, ru | toxicity (English), toxic (others) |
SEVERE_TOXICITY |
en | severe_toxicity |
IDENTITY_ATTACK |
en | identity_attack |
INSULT |
en | insult |
PROFANITY |
en | obscene |
THREAT |
en | threat |
SEXUALLY_EXPLICIT |
en | sexual_explicit |
The _EXPERIMENTAL names Perspective used (INSULT_EXPERIMENTAL, TOXICITY_EXPERIMENTAL and so on) and TOXICITY_FAST are accepted as aliases. The response keeps the name you asked for.
A running server lists what it supports at GET /v1alpha1/attributes.
Models
Section titled “Models”| Model | Used for | License |
|---|---|---|
| unitary/unbiased-toxic-roberta | Every attribute in English | Apache-2.0 |
| unitary/multilingual-toxic-xlm-roberta | TOXICITY in fr, es, it, pt, tr, ru |
Apache-2.0 |
Both come from the Detoxify project and were trained on data from Jigsaw’s Kaggle competitions, the same team behind Perspective. The English model was trained to reduce bias against identity terms. The multilingual model was trained for overall toxicity only, so openspective doesn’t use it for the other attributes: a score from a head that was never trained in that language would look precise and mean nothing.
Not supported yet
Section titled “Not supported yet”- Other languages for TOXICITY. Perspective also scored German, Arabic, Japanese, Korean, Chinese and others. Adding them is the top item on the roadmap.
- Non-English INSULT, THREAT, PROFANITY and IDENTITY_ATTACK.
- FLIRTATION, the New York Times attributes (
ATTACK_ON_AUTHOR,INFLAMMATORY,LIKELY_TO_REJECT, …) and the bridging attributes (AFFINITY_EXPERIMENTAL,REASONING_EXPERIMENTAL, …). No open model covers them.
Requesting any of these returns the error Perspective returns for an unsupported language, so dropUnsupportedAttributes works as before.