Research And Implementation Of Financial Report Automatic Generation System

As the mother of all industries,finance is deeply intertwined with them and produces a large number of financial documents.Although information technology continues to evolve,the work of data collection,sorting,and writing still relies on humans.If we can use automation technology to do the tedious and time-consuming financial report,we can help financial practitioners to focus on more valuable work,improve their work efficiency,and reduce repetitive manual works.However,there are still many problems in applying deep learning technology to report generation.First,financial report generation is a problem that belongs to the category of text generation.The template method is used traditionally,but this method lacks flexibility.And while there are so many different types of reports,saving these templates is very expensive.Second,because data-to-text generation needs to ensure the originally data appear in the generated text,it cannot directly use text-to-text methods such as dialogue generation and poetry generation to solve the problem.Third,text generation is a task with simple input but complex output,and the complexity of the problem is high.The generating methods that currently used need to be improved.Fourth,after de-template in report automation,how to evaluate the generated text is also a problem.Therefore,this article shows the following work which can help to solve these problems:(1)Based on the requirement analysis on the financial reports automatic generation task,this article designs a practical financial report automatic generation system.The system uses a deep learning model on text sequence to realize report generation,including data preprocessing,report generation model realization,model training,model evaluation and other functions.(2)This article optimizes the current model for the automated generation of financial reports.In the traditional text generation model,the input text is regarded as a timeseries text sequence.The time-series is considered as the feature to generate expected text.However,the input of the data-to-text model is based on structured data rather than the time-series data.Based on the selection of a suitable text generation model,this article optimizes the model according to the characteristics of financial reports to adapt to structured natural language input to generate high quality and accuracy text.(3)To improve the performance of the model.With the Bert model achieving good results in the field of natural language processing,the high performance of Transformer neural network unit has been proven.This thesis tries to use Transformer to replace RNN neural network unit and so it can help traditional text generation model to generate financial report with higher quality.In summary,this thesis realizes the financial report automatic generation system.Through these works,this thesis improves the performance of data-to-text model.Research indicates that 1)Compared with the traditional neural network text generation model,the improved data-totext text generation model in this article has higher generation accuracy.2)This text generation model has certain practicability in financial report automation tasks.