Research On Automatic Generation Of Financial Research Report Based On Deep Learning

In the domain of finance,researchers often need to write a large number of financial reports,such as macroeconomic research report,investment strategy report and prospectus.The writing of such reports involves tedious work load collecting,sorting and analyzing massive data and requires tracking of real-time financial including event news.At present,financial investment companies in the market adopt a pre-defined report template to personalize customized content for different types of reports.After processing and analyzing massive structured data,updated data will be filled in the corresponding position of the report.This is not flexible as data is combined with fixed-format templates,although it can reduce the amount of data calculation and replication.How to extract financial knowledge and key logic from massive real-time financial event news information and then automatically generate relevant analysis research reports,is an important research issue.For the automatic generation of financial research report,this thesis adopts natural language generation technology based on deep learning technique to intelligently extract effective information of unstructured data such as financial event news.The data collection is obtained through a web crawler to obtain a macro analysis report of the financial portal and corresponding financial event news is extracted to generate a training data set.The financial portals are Oriental Fortune Network,Sina Finance Network and Flushing Network.We investigate the text-generated text model in recent years and focus more on the sequence-to-sequence model,attention mechanism,pointer generation model and algorithmic ideas of the variational autoencoder model.This thesis proposes two deep learning models for this specific text-to-text generation tasks of financial event news with fewer words and generates longer research reports.It can realize short text information extraction and analysis,and then quickly and automatically generates long text function.The first model is a text generation model based on keyword extraction.Another model is a text generation model based on multiple edits.The text-generating model based on keyword extraction adopts the unsupervised Text Rank algorithm in the first stage and selects top words from financial news words according to importance order as keywords.At the second stage,it uses keyword semantic generation model to generate models text.The text generation model based on multiple edits uses financial news as input during the first stage of decoding and uses the pointer god generated network model to generate a coarse-grained outline of research report.In the second stage of decoding,it uses the outline of research report and financial news generated in the first stage as input and uses variational autoencoder model to generate a fine-grained research report final.In this thesis,comparative experiments have been conducted on dataset obtained by web crawler.The comparison model includes the commonly used natural language generation models.The evaluation indicators are BLEU indicator,ROUGE indicator and manual indicator.The effectiveness of the deep learning-based text generation model proposed in this thesis is verified by the promising experimental results.