基于语义先验的可微分软性编辑距离损失计算方法和设备
By introducing dynamic transpose cost and softening operator, a dynamic programming state transition equation is constructed, which solves the problem of distinguishing between causal time series and random permutations in sequence prediction models, and improves the model's generalization ability and training stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH AT WEIHAI
- Filing Date
- 2026-03-25
- Publication Date
- 2026-07-17
AI Technical Summary
In existing sequence prediction training methods, the hard-matching loss function causes the model to be unable to distinguish between causal time series and random permutations, resulting in overfitting and poor generalization ability.
We employ a semantic prior-based differentiable soft edit distance loss calculation method. By introducing dynamic transpose cost and softening operator, we construct a dynamic programming state transition equation to achieve end-to-end gradient backpropagation and adaptively learn the causal constraints and random permutations between sequence elements.
It improves the model's generalization ability and training stability for nondeterministic sequences, and can maintain logical rigor while allowing reasonable changes in order, thus avoiding overfitting of the model to the training data.
Smart Images
Figure CN121901541B_ABST