| With the rapid development of the financial industry,information in the financial field is produced in a spurt,and the amount of data that the banking system needs to process is increasing day by day.In the process of bank information processing,it is very important to extract and picket the bills used in international trade.Traditionally relying on manual identification to extract key information of bills is time-consuming and labor-intensive,and the efficiency is too low.It is urgent to use machines to replace manual extraction of bill information.At present,most of the common bill recognition software on the market can only identify fixed-format invoices,tickets,etc.The content and format of these bills are relatively simple,and the identification difficulty is relatively low.However,the financial instruments used in international trade do not have fixed typesetting rules and there are still a lot of irrelevant information,so it is difficult to meet the identification requirements using general instrument identification technology.Named entity recognition technology is a hot spot in the field of artificial intelligence research.Information extraction tasks.Therefore,this subject designs an improved BiLSTM +CRF model with self-attention mechanism added to complete the extraction of key information of financial instruments in international trade.The specific work is as follows:(1)Aiming at the problems of different text formats and irregular grammar of financial bills in international trade,a named entity recognition model for financial bills incorporating part-ofspeech information is proposed.The word segmentation tool NLTK is selected to extract the partof-speech information of the text,and the part-of-speech information is used as a feature and The word embedding vector is combined as the input of the BiLSTM+CRF model;at the same time,based on the BiLSTM+CRF model,a self-attention mechanism is introduced to generate the BiLSTM+ATT+CRF model,and the self-attention mechanism pays attention to important feature information and assigns more information to the key features.High weights,so as to further improve the feature engineering of bill text.(2)In the acquisition of training data,a complete training set generation and enhancement scheme is proposed.Scan the bill image provided by the bank and input it into the computer in the form of a grayscale image.First,the image is preprocessed,then the Google open source tool Tesseract is used to identify the text in the bill image,and finally the bill named entity is constructed using the Yedda tool developed based on the Tkinter package.The identified training and test sets.In addition,due to the privacy of bank bill information and other reasons,there are few original bill images that can be directly obtained,resulting in insufficient training set samples.This paper proposes a variety of targeted data enhancement methods to expand the training set,so that the model can obtain better recognition effect.(3)The model hyperparameter control experiment,the training set data enhancement effect control experiment and the multi-model control experiment were carried out,and the model was finally applied to the actual project.The hyperparameter control experiments compared different learning rates,batch samples and drop out parameters,and gave the optimal hyperparameters suitable for the follow-up experiments of this topic;multiple sets of control experiments explored the use of various individual data augmentation methods to expand the training set and The effect of augmenting the training set on the model effect was carried out using a multi-combination data augmentation method;the multi-model control experiment referred to other named entity recognition studies to investigate the effect of different deep learning models on the named entity recognition task of financial bills in international trade.After obtaining the optimal recognition model,this paper applies the model to practical engineering projects to verify the feasibility and practicability of the model.The experimental results show that the data augmentation method proposed in this paper can effectively improve the model performance,and the BiLSTM+ATT+CRF model with selfattention layer is also better than the original BiLSTM+CRF model.The bank bill named entity recognition model finally used in this paper can achieve the experimental effect of F1 value exceeding 68 in the application background with few bill image samples and insufficient image definition,and can effectively identify and extract entity information in bills. |
