Abstract: In the past MIE tasks,researchers usually use data layer fusion to train neural network models for MIE.However,due to the heterogeneity among different modalities,this fusion approach can lead to issues such as feature redundancy,incompatibility,and lack of interpretability,which in turn affect the effectiveness of model training.In view of this,this paper proposes a decision-level fusion method based on the DS theory to solve the problems of feature redundancy,incompatibility,and lack of interpretability caused by data layer fusion.The evidence is generated by processing different modal features through neural networks and Dirichlet functions,and after evidence correction and weight assignment,the Shafer fusion rule is utilized to arrive at the final decision.This method effectively improves the accuracy of feature processing and the interpretability of the model.Using accuracy,recall,and F1 score as evaluation metrics,experiments on public and private datasets show an overall performance improvement of 0.22 to 1.87 percentage points compared to existing methods.

Key words: Key information extraction,Multimodality,Dempster-Shafer theory,Deep learning,Evidentiary amendments,Decision fusion