---

“Kill him! Kill him NOW!”
“I will kick your a*!”
“Touch your chest. Then push the ball harder.”

When processing chat in computer games and sports-related communities, moderation APIs face a dilemma. On one hand, the classifier can’t just ignore utterances that are normally considered policy violations. On the other hand, as Dr Evil’s support group members would say, “we don’t really want to kill each other in here.” And then, there are real threats that are not to be ignored either.

Some moderation APIs “solve” the issue by assigning “severity levels” and recommending their users to dismiss anything under a specific severity threshold. But game violence may be severe, and it’s still nothing to do with the real world. So even the most severe utterances are to be ignored. On the other hand, reliance on severity means that actual threats may have to be discarded.

Tisane offers two methods of handling this dilemma:

Method 1: ignore all “criminal_activity” type occurrences that have tags like “violence” and “death” on the client app side.

Method 2 (recommended): using a special “game violence” flag to ignore game & sporting competitive language.

Example:


{"language":"en",
"content":"Shoot him!!!",
"settings":{"snippets":true, "memory":{"flags":["game_violence_ok"]}}}


The game_violence_ok flag ensures that all alerts related to sports competitive language and game violence are silenced. Note that the flag won’t silence actual threats like “I know where you live” or anything clearly not related to gaming.

Output:

 

For sports, you may also supply expected domain of discourse IDs: "expected_domains":[12838,12572,12579,12553,154300,135033,43220].