Prediction Of Fines In Customs False Trade Cases Based On Deep Learning

The expeditious growth of China's import and export economy and trade has been witnessed recently.Simultaneously,the violations in imports and export are increased.In order to ensure the healthy and orderly development of import and export trade,customs legal officers need to increase efforts to crack down on false trade cases and predict the corresponding fines based on the case description text.However,customs legal officer currently mainly rely on manual methods to analyze case texts and predict the number of fines based on this,which is heavy and inefficient.Because of the above analysis,this thesis proposes a deep learning-based model for predicting the number of fines in customs false trade cases.The main contents of the thesis as followed:(1)In the process of text analysis of false trade cases,this thesis uses the traditional text analysis model Quick-thoughts to encode the case text.Considering that the Quick-thoughts model uses GRU to encoder the text,and the case text is usually longer,GRU can not capture the long-distance dependence problem in the case text,which will affect the accuracy of the case analysis.In response to this problem,this thesis uses the Transformer-XL model to replace the GRU module in the Quick-thought model to solve the problem that GRU cannot model dependencies that exceed a fixed length,to generate text with high accuracy and sufficient semantic expression Vector Representation.(2)In order to fully learn semantic information from the text vector to accurately predict the number of fines,a suitable prediction model based on the Encoder-Decoder framework has been designed and built in this thesis.Specifically,this thesis uses the Bi-LSTM model as the Encoder to better learn semantic information from the case text vector;The attention mechanism module has been introduced between the Encoder and Decoder to achieve the semantic expression of each sentence,to classify the influential caused by the final forfeit.Finally,this thesis uses the CNN model as the Decoder to process high-dimensional vector representations,to achieve the final fine amount prediction.(3)Multiple comparison experiments based on real customs trade data show that the model proposed in this thesis can effectively improve the RMSE,MAPE,and consistency rate indicators of the fine amount prediction task on the real data set.The work of this thesis has effectively improved the accuracy in customs false trade cases in the customs judicial field,which can be regarded as a novel idea and method for handling fines in customs false trade cases,also a new exploration of deep learning applied in the customs justice.