Semantic Preservation Loss
A regularization loss function that constrains the difference between evolved text embeddings and the original pretrained language model embeddings, preventing text representations from deviating excessively from the general semantic space during co-evolution.
Timeline (last 90 days)
The framework introduces a language model maintenance objective that anchors the evolving text embeddings to the original semantic structure to prevent semantic distortion.
The contrastive pretraining loss and the semantic maintenance loss act together to achieve controlled co-evolution.
All Facts (2)
The framework introduces a language model maintenance objective that anchors the evolving text embeddings to the original semantic structure to prevent semantic distortion.
50%UnverifiedThe contrastive pretraining loss and the semantic maintenance loss act together to achieve controlled co-evolution.
50%