Differential soft editing distance loss calculation method and device based on semantic prior

Abstract

The invention provides a differential soft editing distance loss calculation method and device based on semantic prior, and belongs to the technical field of data processing. The method comprises the following steps: extracting context features of an obtained prediction sequence to generate a context vector; inputting the vector into a cost prediction network to obtain a normalized original score, and performing linear mapping to generate a dynamic basic transposition cost; based on a preset semantic exchangeability matrix, calculating semantic exchange items between adjacent elements of the prediction sequence and cross gating items between adjacent elements of the prediction sequence and the real tag sequence; constructing a dynamic planning state transition equation according to the dynamic basic transposition cost, the semantic exchange item and the cross gating item; and carrying out smooth aggregation on the path cost in the equation by adopting a softening operator to obtain a final alignment loss value, and realizing end-to-end gradient back propagation. Based on the method, the invention also provides corresponding equipment. According to the method, adaptive decoupling and characterization of causal logic and random arrangement of the sequence elements are realized.

Images