Document intelligent evaluation method and system based on auto-encoder
By optimizing the structure of the autoencoder and combining the SVM algorithm, using the coded representation of the intermediate hidden layer for document classification identification, the problem that intelligent document evaluation method in the existing technology cannot cope with the insufficient accuracy of massive data and feature extraction, and achieve high efficiency , accurate document processing.
Patent Information
- Application Number
- CN202411229279.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-03
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing intelligent document evaluation methods cannot cope with the explosive growth of massive document data, and the feature extraction accuracy is limited, which cannot meet the efficient and accurate needs of enterprises for document processing.
Using the method based on the autoencoder, the structure optimization of the autoencoder is used to extract the feature information of the document evaluation data, and a classification model is constructed in combination with the support vector machine (SVM) algorithm, and the encoding representation of the intermediate hidden layer is used for document classification identification. Generate a document evaluation report.
It improves the efficiency and accuracy of document feature extraction, overcomes the shortcomings of traditional methods in dimensional disasters and data recognition scale accuracy, provides high-quality evaluation vectors, improves the training speed of intelligent models and reduces computational costs.
Smart Images

Figure CN119090662A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of document data processing, and in particular to a document intelligent evaluation method and system based on an autoencoder. Background Art
[0002] In the trend of economic globalization and digital transformation, many companies have established their own financial shared centers, which are platforms dedicated to processing and sharing financial information. Financial shared centers receive a large number of documents of various types every day. How to process them efficiently, accurately and quickly is crucial for capital allocation, inventory management and supply chain optimization. However, traditional manual processing methods are time-consuming and error-prone, and can no longer meet the requirements of rapid development of enterprises. Therefore, many companies have begun to seek solutions to use computers instead of manpower to process documents, and optimizing the document processing process by extracting the evaluation vector information of the processed documents is a crucial step in this solution.
[0003] Currently, the information technology for extracting and processing the evaluation vector of documents mostly relies on manual coding or statistical feature extraction technology. Among them, the most widely used methods are principal component analysis (PCA) and feature selection. These technologies can simply process sample data and have low technical difficulty. They can quickly extract evaluation information vectors and analyze them. Although simple and easy to use, facing the explosive growth of data volume, the above methods can no longer cope with the massive amount of document data, and the feature extraction accuracy is limited. Summary of the invention
[0004] The embodiments of the present invention provide a method and system for intelligent evaluation of documents based on an autoencoder, which are used to solve the following technical problems: In the existing intelligent evaluation methods of documents, the extraction technology of document evaluation vector information cannot cope with the explosive growth of document data, and the feature extraction accuracy is limited, which is not conducive to the intelligent evaluation of documents.
[0005] The embodiment of the present invention adopts the following technical solutions:
[0006] On the one hand, an embodiment of the present invention provides a document intelligent evaluation method based on an autoencoder, the method comprising: collecting document evaluation data in a financial shared center system, and performing preprocessing to obtain a document evaluation data set;
[0007] The structure of the autoencoder is optimized to obtain a new autoencoder;
[0008] By using the novel autoencoder, feature extraction is performed on the document evaluation data set to obtain evaluation data feature information;
[0009] Training a classification model through the evaluation data feature information;
[0010] The target document evaluation data is classified and identified through the novel autoencoder and the trained classification model to generate a document evaluation report.
[0011] In a feasible implementation, the document evaluation data is collected in the financial shared center system and preprocessed to obtain a document evaluation data set, which specifically includes:
[0012] Extracting document evaluation data of various documents on a large scale in the financial shared center system; wherein the various documents at least include reimbursement documents, sales orders, and purchase orders;
[0013] The extracted document evaluation data is converted into numerical data through one-hot encoding;
[0014] The numerical data are standardized by a z-score standardization method;
[0015] The deviation standardization method is used to normalize the numerical data after standardization;
[0016] The normalized data is expanded and converted into two-dimensional data to obtain the document evaluation data set.
[0017] In a feasible implementation, the structure of the autoencoder is optimized to obtain a new autoencoder, which specifically includes:
[0018] Set the input vector of the autoencoder to x∈Rl;
[0019] Add a preset encoding function h in each hidden layer of the autoencoder i =σ(ω i h i-1 +b i ), perform structural optimization to encode and map the input vector learned by each hidden layer of the autoencoder to x i ∈Rl; where i = (1, 2, ..., n), n is the number of hidden layers, σ is the activation function, ω is the parameter of the hidden layer convolution kernel, b is the bias of the convolution kernel, h i is the encoding representation of the input vector after the i-th hidden layer, l is the dimension of the input vector, and Rl is the original data sample set;
[0020] Outputting the feature reconstruction error of the original data sample set through the structurally optimized autoencoder;
[0021] Based on the feature reconstruction error, the autoencoder is error adjusted to obtain the new autoencoder.
[0022] In a feasible implementation manner, outputting the feature reconstruction error of the original data sample set by the structurally optimized autoencoder specifically includes:
[0023] Inputting the original data sample set into the autoencoder after structural optimization, and outputting the corresponding feature reconstruction result set;
[0024] according to Get the feature reconstruction error of the autoencoder after structural optimization Among them, m is the number of training samples, x i is the i-th original data sample, y i is the i-th original data sample x i The feature reconstruction results.
[0025] In a feasible implementation manner, based on the feature reconstruction error, the error of the autoencoder is adjusted to obtain the novel autoencoder, which specifically includes:
[0026] Back-propagating the feature reconstruction error to the input layer of the autoencoder while updating the weights and biases of each hidden layer;
[0027] After the error adjustment, the feature reconstruction error of the autoencoder is recalculated until the feature reconstruction error is less than a preset threshold, and the autoencoder at this time is saved to obtain the new autoencoder.
[0028] In a feasible implementation manner, training a classification model through the evaluation data feature information specifically includes:
[0029] A classification model is constructed by using a support vector machine (SVM) algorithm; the classification model includes multiple classifiers for respectively identifying different types of document evaluation data;
[0030] Taking the evaluation data feature information as the input of the classification model, and selecting the radial basis function as the kernel function of the SVM algorithm to train the classification model;
[0031] The classification model is connected after the output layer of the novel autoencoder.
[0032] In a feasible implementation manner, after training the classification model through the evaluation data feature information, the method further includes:
[0033] The evaluation data vectors in the test set are reduced in dimension through t-SNE dimensionality reduction technology and visualized;
[0034] Applying the processed test set to the trained classification model and calculating the cosine distance between the output value and the actual value to evaluate the classification performance of the classification model;
[0035] If the classification performance does not meet the preset requirements, the classification model is further optimized.
[0036] In a feasible implementation, the target document evaluation data is classified and identified by using the novel autoencoder and the trained classification model to generate a document evaluation report, which specifically includes:
[0037] Inputting the target document evaluation data into the novel autoencoder for feature extraction to obtain a corresponding target feature reconstruction result;
[0038] The target feature reconstruction result output by the output layer of the novel autoencoder is directly input into the classification model to perform classification recognition to obtain a classification result;
[0039] The document evaluation report is generated according to the classification result.
[0040] In a feasible implementation manner, generating the document evaluation report according to the classification result specifically includes:
[0041] The document type, evaluation type and question type of the target document evaluation data are extracted from the classification results, and filled into a preset evaluation report template to generate the document evaluation report.
[0042] On the other hand, an embodiment of the present invention further provides a document intelligent evaluation system based on an autoencoder, the system comprising:
[0043] The data preparation module is used to collect document evaluation data in the financial shared center system and perform preprocessing to obtain a document evaluation data set;
[0044] A feature extraction module is used to optimize the structure of the autoencoder to obtain a new autoencoder; through the new autoencoder, feature extraction is performed on the document evaluation data set to obtain evaluation data feature information;
[0045] The document intelligent evaluation module is used to train the classification model through the characteristic information of the evaluation data; classify and identify the target document evaluation data through the new autoencoder and the trained classification model to generate a document evaluation report.
[0046] Compared with the prior art, the document intelligent evaluation method and system based on the autoencoder provided in the embodiment of the present invention has the following beneficial effects:
[0047] The present invention firstly proposes an asymmetric deep convolutional autoencoder method that only retains the encoder part. By retaining the encoder part and combining it with the SVM classifier, it can effectively extract and utilize the encoding representation of the intermediate hidden layer to identify the evaluation information, which is expected to improve the efficiency and accuracy of feature extraction. The asymmetric stacked multi-layer encoding model of this method also provides a novel and effective way to process evaluation data.
[0048] Compared with the prior art, the present invention does not focus on what the output of the autoencoder is, but focuses on the encoding representation of the intermediate hidden layer, in other words, it focuses on the mapping process from the input layer to the hidden layer. Therefore, the present invention is more suitable for processing intelligent evaluation information based on documents, can complete more sophisticated vectorization processing, retain the unique characteristics of the evaluation information, provide high-quality evaluation vector references for subsequent document processing procedures, improve the training speed of the intelligent model and reduce the calculation cost of the model, and overcome the shortcomings of traditional feature extraction technology in terms of dimensional disaster, data recognition scale accuracy, etc. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:
[0050] Figure 1 A flow chart of a method for intelligent document evaluation based on an autoencoder provided in an embodiment of the present invention;
[0051] Figure 2 A detailed implementation process flow chart of a method for intelligent document evaluation based on an autoencoder provided in an embodiment of the present invention;
[0052] Figure 3 A schematic diagram of the structure of an intelligent invoice evaluation system based on an autoencoder provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0053] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0054] The embodiment of the present invention provides a method for intelligently evaluating a document based on an autoencoder, such as Figure 1 As shown, the document intelligent evaluation method based on the autoencoder specifically includes steps S101-S105:
[0055] S101. Collect document evaluation data in the financial shared center system and perform preprocessing to obtain a document evaluation data set.
[0056] Specifically, document evaluation data of various documents are extracted on a large scale in the financial shared center system; wherein the various documents at least include reimbursement documents, sales orders and purchase orders.
[0057] Furthermore, the extracted document evaluation data is converted into numerical data through one-hot encoding, and then the numerical data is standardized through the z-score standardization method.
[0058] Furthermore, the deviation normalization method is used to normalize the numerical data after the normalization process. The normalized data is expanded and converted into two-dimensional data to obtain the document evaluation data set.
[0059] S102. Optimize the structure of the autoencoder to obtain a new autoencoder.
[0060] Specifically, the input vector of the autoencoder is set to x∈Rl. Then, a preset encoding function h is added to each hidden layer of the autoencoder. i =σ(ω i h i-1 +b i ), perform structural optimization to encode and map the input vector learned by each hidden layer of the autoencoder to x i ∈Rl; where i = (1, 2, ..., n), n is the number of hidden layers, σ is the activation function, ω is the parameter of the hidden layer convolution kernel, b is the bias of the convolution kernel, h i is the encoded representation of the input vector after the i-th hidden layer, l is the dimension of the input vector, and Rl is the original data sample set.
[0061] Furthermore, the autoencoder with optimized structure outputs the feature reconstruction error of the original data sample set. Based on the feature reconstruction error, the autoencoder is error-adjusted to obtain a new autoencoder.
[0062] As a feasible implementation method, the feature reconstruction error of the original data sample set is output by the autoencoder after structural optimization, specifically including:
[0063] The original data sample set is input into the structure-optimized autoencoder, and the corresponding feature reconstruction result set is output. Get the feature reconstruction error of the autoencoder after structural optimization Among them, m is the number of training samples, x i is the i-th original data sample, y i is the i-th original data sample x i The feature reconstruction results.
[0064] As a feasible implementation method, based on the feature reconstruction error, the error of the autoencoder is adjusted to obtain a new autoencoder, which specifically includes:
[0065] The feature reconstruction error is back-propagated to the input layer of the autoencoder, and the weights and biases of each hidden layer are updated at the same time. After the error is adjusted, the feature reconstruction error of the autoencoder is recalculated until the feature reconstruction error is less than the preset threshold, and the autoencoder at this time is saved to obtain a new autoencoder.
[0066] S103. Using a novel autoencoder, feature extraction is performed on the document evaluation data set to obtain feature information of the evaluation data.
[0067] Specifically, the preprocessed document evaluation data set is input into the new autoencoder for feature extraction to obtain the corresponding evaluation data feature information.
[0068] S104: training a classification model by evaluating data feature information.
[0069] Specifically, a classification model is constructed by using a support vector machine (SVM) algorithm; the classification model includes multiple classifiers for respectively identifying different types of document evaluation data.
[0070] Furthermore, the evaluation data feature information is used as the input of the classification model, and the radial basis function is selected as the kernel function of the SVM algorithm to train the classification model. The classification model is connected after the output layer of the novel autoencoder.
[0071] Furthermore, the evaluation data vectors in the test set are reduced in dimension by t-SNE dimensionality reduction technology and visualized. The processed test set is then applied to the trained classification model, and the cosine distance between the output value and the actual value is calculated to evaluate the classification performance of the classification model. If the classification performance does not meet the preset requirements, the classification model is further tuned.
[0072] S105. Classify and identify the target document evaluation data through the new autoencoder and the trained classification model to generate a document evaluation report.
[0073] Specifically, the target document evaluation data is input into the new autoencoder for feature extraction to obtain the corresponding target feature reconstruction result. Then, the target feature reconstruction result output by the output layer of the new autoencoder is directly input into the classification model for classification recognition to obtain the classification result.
[0074] Furthermore, a document evaluation report is generated based on the classification results, including:
[0075] The document type, evaluation type and problem type of the target document evaluation data are extracted from the classification results, and filled into a preset evaluation report template to generate a document evaluation report.
[0076] To describe the technical solution of the present application in detail, a specific embodiment is used as an example to explain in detail:
[0077] Figure 2 A detailed implementation process flow chart of a method for intelligent document evaluation based on an autoencoder provided in an embodiment of the present invention, such as Figure 2 As shown, the main steps of the present invention are as follows:
[0078] 1. Data Preparation
[0079] Collect and prepare the document evaluation data set for training, and perform preprocessing operations such as normalization and denoising on the data so that the autoencoder can better learn the data features. First, a large amount of document evaluation data is found from the financial shared center system, including evaluation information of different types of documents such as reimbursement documents, sales orders, and purchase orders. All of these data come from the company's internal system and are classified and sorted according to the type of document. Then, the data samples are converted into numerical types using the one-hot encoding method. Secondly, the data after numerical processing is standardized using the z-score standardization method. The deviation standardization method is used again to normalize the standardized data set, so that the data samples are comparable between different dimensions. Finally, for the normalized data set, the original 1-dimensional feature attribute sample set is expanded by filling the number 0 and converted into 2-dimensional data, thereby obtaining the standard input of the feature extraction model.
[0080] When integrating the contents of the document, you need to pay attention to the following points:
[0081] 1. Remove noise, errors or invalid evaluation information to improve the quality of the data set
[0082] 2. Pay attention to the size of the dataset. Larger datasets usually provide more comprehensive and extensive feature coverage, which helps improve the accuracy of the model. If the data is insufficient, consider using data augmentation techniques or introducing external data sources.
[0083] 2. Feature extraction
[0084] The feature extraction method based on the improved autoencoder is used to extract features from the preprocessed evaluation data information to obtain the optimal feature representation of the original data. It mainly includes the following steps:
[0085] 1. Set the encoder input vector to x∈Rl, encode the input learned in each layer and map it to x i ∈Rl, where l refers to the dimension of the input data and Rl is the original data sample set. The encoding function is determined as h i =σ(ω i h i-1 +b i ), where i = (1, 2, ..., n), n is the number of hidden layers, σ is the activation function, ω is the parameter of the hidden layer convolution kernel, b is the bias of the convolution kernel, h i It is the encoded representation of the input vector after passing through the i-th hidden layer.
[0086] 2. After the conversion of the n-layer encoder hidden layer and activation function, it can be expressed as h n =σ(ω n h n-1 +b n ),i=(1,2,...,n) feature data, where ω n is the parameter of the convolution kernel of the nth encoder hidden layer, b n is the bias of the convolution kernel of the nth encoder hidden layer, h n is the encoded representation of the sample after passing through the nth encoder hidden layer, h n-1 is the encoding representation of the sample after passing through the n-1th encoder hidden layer. When processing nonlinear functions, the gradient of the function value in the non-negative interval is taken as a function, which means that the gradient will not approach 0 when the back propagation algorithm gradient is updated, ensuring that the convergence speed of the model remains in a stable state.
[0087] 3. Adjust the error by back propagation and express the reconstruction error of the final feature extraction model as Among them, m is the number of training samples, x i is the i-th original data sample, y i is the i-th original data sample x i The feature reconstruction result is is the Euclidean distance between the original data and the reconstructed result.
[0088] The encoder hidden layer is specifically as follows: the model adopts an 8-layer hidden layer structure, in which the first convolution layer uses 64 convolution kernels, each of which is set to a 8×8 two-dimensional matrix; the second convolution layer uses 32 convolution kernels, and the convolution kernel size is set to 6×6; the third convolution layer also uses 32 convolution kernels, and the convolution kernel size is set to 6×6; the fourth convolution layer uses 16 convolution kernels, and the size of each convolution kernel is set to 4×4; at the same time, after each convolution layer is processed, a pooling layer is used to process the data to further extract features, and the size of the pooling layer convolution kernel is set to 4×4; after the preprocessed data is processed by 4 convolution layers and 4 pooling layers, all connection layers are used to process the data.
[0089] 3. Training Model
[0090] The extracted features are used to train the classification model and classify and identify different document evaluation information. The main steps include:
[0091] 1. Use the feature extraction model to extract features from the preprocessed training set samples to obtain the reduced-dimensional traffic feature data;
[0092] 2. Use the support vector machine (SVM) algorithm to train the feature data of the evaluation data after dimensionality reduction to build a classification model. Use the above evaluation information feature data as the input of the SVM algorithm, and select the radial basis function as the kernel function of the SVM algorithm. Then adjust the parameters continuously to make the trained classifier maintain a better classification effect. Since multiple categories of evaluations need to be identified, multiple classifiers need to be trained.
[0093] 3. Connect the SVM classifier after the output layer of the autoencoder to classify and identify the data samples to obtain the classification results.
[0094] 4. Generate Report
[0095] Based on the recognition results obtained in the third step, a document evaluation information analysis report is drawn up according to document category, problem type, etc.
[0096] 5. Model Evaluation
[0097] 1. Apply the classification model to the test set to evaluate the model performance.
[0098] 2. Use t-SNE and other dimensionality reduction techniques to reduce the dimensionality of the evaluation data vector, and then visualize it to more intuitively view the performance of the classification model.
[0099] 3. Objectively evaluate the performance of the classification model by calculating quantitative indicators such as the distance and cosine value between evaluation data vectors.
[0100] 6. Application and Tuning
[0101] According to actual needs and application scenarios, the trained classification model is applied to specific tasks. According to task requirements, the evaluation data vector may need to be further tuned or fine-tuned.
[0102] 7. Data and model updates
[0103] The model can be trained and evaluated in a cycle every day based on the evaluation information of newly generated documents, ensuring that the generated model will not affect the accuracy of similarity classification of new documents in terms of continuous improvement.
[0104] The convolutional autoencoder fully combines the advantages of autoencoders and convolutional neural networks, and realizes efficient feature extraction and conversion of data through convolution and pooling operations. It effectively solves the problem of too many training parameters caused by the fully connected structure in traditional autoencoders, and reduces the dependence on large-scale labeled data, making the model more suitable for practical application scenarios. Using the deep training method to construct multiple hidden layers, it is possible to extract the best low-dimensional representation from the original high-dimensional data, providing a more efficient solution for data analysis and processing. The model realizes feature extraction and conversion of data by constructing multiple hidden layers. The output of each hidden layer will be used as input for higher levels. Therefore, generally speaking, the input data is used for learning in the first layer, and then the first-order features are passed as output to the second layer to learn the second-order features related to the first-order features, and so on, and finally the best low-dimensional representation of the original high-dimensional data is obtained.
[0105] As for extracting high-quality features of evaluation data, the present invention does not focus on what the output of the autoencoder is, but focuses on the coding representation of the middle hidden layer, in other words, it focuses on the mapping process from the input layer to the hidden layer. Therefore, the present invention proposes a feature extraction and evaluation information intelligent recognition method based on an improved autoencoder. First, an asymmetric deep convolutional autoencoder method that only retains the encoder part is proposed. By retaining the encoder part and combining it with the SVM classifier, the coding representation of the middle hidden layer can be effectively extracted and utilized to identify the evaluation information, which is expected to improve the efficiency and accuracy of feature extraction. The asymmetric stacked multi-layer coding model of this method also provides a novel and effective way to process evaluation data. It overcomes some shortcomings of traditional feature extraction technology in terms of dimensionality disaster, data recognition scale accuracy, etc.
[0106] In addition, the embodiment of the present invention also provides a document intelligent evaluation system based on an autoencoder, such as Figure 3 As shown, the document intelligent evaluation system 300 based on the autoencoder specifically includes:
[0107] The data preparation module 310 is used to collect document evaluation data in the financial shared center system and perform preprocessing to obtain a document evaluation data set;
[0108] The feature extraction module 320 is used to optimize the structure of the autoencoder to obtain a new autoencoder; the new autoencoder is used to extract features from the document evaluation data set to obtain evaluation data feature information;
[0109] The document intelligent evaluation module 330 is used to train a classification model through the evaluation data feature information; classify and identify the target document evaluation data through the new autoencoder and the trained classification model to generate a document evaluation report.
[0110] Each embodiment of the present invention is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.
[0111] The above describes specific embodiments of the present invention. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0112] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for intelligent document evaluation based on autoencoder, characterized in that: The method comprises: Collect document evaluation data in the financial shared center system and pre-process it to obtain a document evaluation data set; The structure of the autoencoder is optimized to obtain a new autoencoder; By using the novel autoencoder, feature extraction is performed on the document evaluation data set to obtain evaluation data feature information; Training a classification model through the evaluation data feature information; The target document evaluation data is classified and identified through the novel autoencoder and the trained classification model to generate a document evaluation report.
2. According to the method for intelligent document evaluation based on autoencoder according to claim 1, it is characterized in that: The document evaluation data is collected in the financial shared center system and preprocessed to obtain the document evaluation data set, which includes: Extracting document evaluation data of various documents on a large scale in the financial shared center system; wherein the various documents at least include reimbursement documents, sales orders, and purchase orders; The extracted document evaluation data is converted into numerical data through one-hot encoding; The numerical data are standardized by a z-score standardization method; The deviation standardization method is used to normalize the numerical data after standardization; The normalized data is expanded and converted into two-dimensional data to obtain the document evaluation data set.
3. According to the method of intelligent document evaluation based on autoencoder in claim 1, it is characterized in that: The structure of the autoencoder is optimized to obtain a new autoencoder, which includes: Set the input vector of the autoencoder to x∈Rl; Add a preset encoding function h in each hidden layer of the autoencoder i =σ(ω i h i-1 +b i ), perform structural optimization to encode and map the input vector learned by each hidden layer of the autoencoder to x i ∈Rl; where i = (1, 2, ..., n), n is the number of hidden layers, σ is the activation function, ω is the parameter of the hidden layer convolution kernel, b is the bias of the convolution kernel, h i is the encoding representation of the input vector after the i-th hidden layer, l is the dimension of the input vector, and Rl is the original data sample set; Outputting the feature reconstruction error of the original data sample set through the structurally optimized autoencoder; Based on the feature reconstruction error, the autoencoder is error adjusted to obtain the new autoencoder.
4. The method for intelligent document evaluation based on an autoencoder according to claim 3, characterized in that: The feature reconstruction error of the original data sample set outputted by the structurally optimized autoencoder specifically includes: Inputting the original data sample set into the autoencoder after structural optimization, and outputting the corresponding feature reconstruction result set; according to Get the feature reconstruction error of the autoencoder after structural optimization Among them, m is the number of training samples, x i is the i-th original data sample, y i is the i-th original data sample x i The feature reconstruction results.
5. The method for intelligent document evaluation based on an autoencoder according to claim 3, characterized in that: Based on the feature reconstruction error, the error of the autoencoder is adjusted to obtain the novel autoencoder, which specifically includes: Back-propagating the feature reconstruction error to the input layer of the autoencoder while updating the weights and biases of each hidden layer; After the error adjustment, the feature reconstruction error of the autoencoder is recalculated until the feature reconstruction error is less than a preset threshold, and the autoencoder at this time is saved to obtain the new autoencoder.
6. The method for intelligent document evaluation based on an autoencoder according to claim 1, characterized in that: Training a classification model through the evaluation data feature information specifically includes: A classification model is constructed by using a support vector machine (SVM) algorithm; the classification model includes multiple classifiers for respectively identifying different types of document evaluation data; Taking the evaluation data feature information as the input of the classification model, and selecting the radial basis function as the kernel function of the SVM algorithm to train the classification model; The classification model is connected after the output layer of the novel autoencoder.
7. The method for intelligent document evaluation based on an autoencoder according to claim 1, characterized in that: After training the classification model through the evaluation data feature information, the method further includes: The evaluation data vectors in the test set are reduced in dimension through t-SNE dimensionality reduction technology and visualized; Applying the processed test set to the trained classification model and calculating the cosine distance between the output value and the actual value to evaluate the classification performance of the classification model; If the classification performance does not meet the preset requirements, the classification model is further optimized.
8. The method for intelligent document evaluation based on an autoencoder according to claim 1, characterized in that: Through the novel autoencoder and the trained classification model, the target document evaluation data is classified and identified to generate a document evaluation report, which specifically includes: Inputting the target document evaluation data into the novel autoencoder for feature extraction to obtain a corresponding target feature reconstruction result; The target feature reconstruction result output by the output layer of the novel autoencoder is directly input into the classification model to perform classification recognition to obtain a classification result; The document evaluation report is generated according to the classification result.
9. The method for intelligent document evaluation based on an autoencoder according to claim 8, characterized in that: According to the classification result, the document evaluation report is generated, which specifically includes: The document type, evaluation type and question type of the target document evaluation data are extracted from the classification results, and filled into a preset evaluation report template to generate the document evaluation report.
10. A document intelligent evaluation system based on autoencoder, characterized in that: The system comprises: The data preparation module is used to collect document evaluation data in the financial shared center system and perform preprocessing to obtain a document evaluation data set; A feature extraction module is used to optimize the structure of the autoencoder to obtain a new autoencoder; through the new autoencoder, feature extraction is performed on the document evaluation data set to obtain evaluation data feature information; The document intelligent evaluation module is used to train the classification model through the characteristic information of the evaluation data; classify and identify the target document evaluation data through the new autoencoder and the trained classification model to generate a document evaluation report.
Citation Information
Patent Citations
Liver disease evaluation method based on autoencoder
CN110364255A
Feature extraction and vulnerability utilization attack detection method based on improved auto-encoder
CN113468537A
Automatic reimbursement receipt identification and verification system based on image identification
CN118570825A
