FIELD: content moderation.
SUBSTANCE: method includes the following steps: performing, using a high-precision verification model and a high-memory verification model that are cascaded, a preliminary violation check in the current live frame from the studio live broadcast to be verified; and if the current live frame has been pre-tested for violation, inputting into the pre-built behaviour verification model a first violation score in the current live frame from the high-precision verification model, a second violation score in the current live frame from the high-retention verification model, and a multidimensional feature of the behaviour in the current live frame from the above studio live stream to obtain a target violation score in the current live frame.
EFFECT: elimination of erroneous rejection or erroneous approval during the live broadcast process.
10 cl, 8 dwg
Authors
Dates
2023-11-20—Published
2021-12-27—Filed