A design method for an intelligent copying system for official document approval based on artificial intelligence

Abstract

The application relates to a design method of an artificial intelligence-based official document approval intelligent transcription system, and belongs to the field of computer software.The application establishes a handwriting recognition algorithm model aiming at handwriting signature content recognition, first acquires a handwriting annotation area image, provides input for handwriting content recognition, extracts handwriting approval content, avoids manual transcription and input work, and reduces the difficulty and quantity of manual recognition work.In order to improve the correctness of handwriting approval final acquisition content, a content error correction link is added, a content error correction algorithm model is established, error detection is carried out from two aspects of word granularity and character granularity through word segmentation, an optimal correction word is obtained through a language model, error correction is completed, and the final accuracy of handwriting content recognition is greatly improved;through template engine technology, a transcription template is constructed, and a document is automatically generated, manual arrangement and document input work is saved, and the file circulation speed is greatly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of computer software, and particularly relates to a design method of a document review intelligent copying system based on artificial intelligence. BACKGROUND

[0002] With the rapid development of information technology, smart office has become an indispensable part of the office environment. The development of artificial intelligence technology has made intelligent office enter a new stage. By integrating various office software and tools, higher level of intelligence is realized, and man-machine cooperation is further enhanced, thereby helping office workers to better complete various tasks, greatly saving human resources and improving work efficiency. In order to meet the requirements of people's work and life, the research and development of smart office related software has become an important task.

[0003] In daily office work, a large number of documents such as review and handling documents are often generated. These documents need to be circulated at multiple levels and require review and signature by leaders at each level. After each level of signature, the secretary or relevant personnel needs to arrange the circle reading content and signature content and copy them into the system with the main text to be distributed to the lower level. First, the recognition of handwritten signatures and the copying of contents greatly increase the work burden of the copying personnel. Second, the manual copying and input work often consumes a lot of time, reduces the efficiency of document circulation, and further affects the progress of related work.

[0004] For the document review copying work, artificial intelligence technology is used for design and modification at different stages. First, the recognition problem of handwritten signature content is solved. Image cutting technology is used to identify the area of handwritten annotations, and then OCR and handwritten body recognition are performed on the area to obtain the relevant annotation content. Second, the recognized review content is processed by using Chinese word segmentation and error correction technology to improve the accuracy of the final content recognition. Third, the content input and arrangement problem is solved. Important information is retrieved from the database to obtain the leader's name, time and other necessary elements. A copying format template is constructed. The template engine is used to arrange the obtained element information and annotation content. Finally, a document file is output to realize the intelligentization of the whole process of document review copying. SUMMARY

[0005] (I) Technical problem to be solved

[0006] The technical problem to be solved by the present application is how to provide a design method of a document review intelligent copying system based on artificial intelligence to solve the problem that the recognition of handwritten signatures and the copying of contents greatly increase the work burden of the copying personnel, and the manual copying and input work often consumes a lot of time and reduces the efficiency of document circulation.

[0007] (II) Technical scheme

[0008] In order to solve the above technical problems, the present application proposes a design method of a document review and approval intelligent transcription system based on artificial intelligence, which is used for text view analysis and text document generation, wherein the text view analysis includes handwriting recognition and content error correction, and the service is divided into offline and online modes; the text document generation is supported by a template engine and automatically generates text content documents according to different templates; the steps used by the system include:

[0009] First, the original file is obtained and converted into a picture format, and region cropping is performed to obtain a region image containing only handwritten signature content, which is input into a handwriting recognition model to perform end-to-end text recognition on indefinite-length text sequences;

[0010] The recognized text content is input into a text error correction model, which includes error detection and error correction. The error detection obtains all suspected error positions, and then the error correction uses similar words to replace the words at the error positions. Finally, the language model calculates the sentence perplexity, compares and sorts all candidate set results, obtains the optimal correction word, and obtains the corrected text content;

[0011] When data splicing, relevant element data is obtained from the database; in online mode, after each handwriting signature is completed, the signature content is recognized through a service interface call, and the signature time and signature person name are obtained through the signature system, which are stored in the database and bound to the corresponding file for persistence; in offline mode, the file on the object storage server is selected for recognition, and the element data is spliced;

[0012] Before the file circulation and signature, the corrected text content is obtained, the transcription content document is generated according to the transcription format template, and the original file is circulated together. The document content is updated each time, and the signature opinion can be selected when viewing the original text to display the signature content set in the corresponding text format.

[0013] (Three) beneficial effects

[0014] The application proposes an intelligent copy system design method for official document approval based on artificial intelligence, compared with the prior art, the application proposes an intelligent system design for the whole process of official document approval copy based on artificial intelligence, and selects appropriate algorithms for each processing process of the task to optimize the framework. For handwritten signature content recognition, a handwriting recognition algorithm model is established, first, the handwritten annotation area image is obtained to provide input for handwriting content recognition, the handwritten content is extracted to avoid manual transcription and reduce the difficulty and amount of manual recognition. In order to improve the correctness of the final obtained content of the handwritten signature, a content error correction algorithm model is established, error detection is carried out from two aspects of word granularity and word granularity through word segmentation, the optimal correction word is obtained through the language model, the error correction is completed, and the final accuracy of the handwriting content recognition is greatly improved; through the template engine technology, the document is automatically generated according to the copy template constructed, the manual arrangement and document input work is saved, and the file circulation speed is greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 It is the overall architecture diagram of the application;

[0016] Figure 2 It is the system flowchart of the application;

[0017] Figure 3 It is the text recognition flowchart of the application;

[0018] Figure 4 It is the content error correction flowchart of the application;

[0019] Figure 5 It is the document generation flowchart of the application. DETAILED DESCRIPTION

[0020] In order to make the purpose, content and advantages of the application more clear, the specific embodiments of the application are further described in detail below in combination with the drawings and examples.

[0021] The purpose of the application is to propose an official document approval intelligent copy system design which can utilize artificial intelligence technology and realize the whole process of official document approval copy in the office field from signature content recognition to copy document generation.

[0022] 1. Overall architecture

[0023] Figure 1 The overall architecture diagram of the application is shown in Figure 1 The system technical architecture can be divided into four layers, namely the presentation layer, the transmission layer, the service layer and the data layer.

[0024] The presentation layer is mainly the user-oriented web terminal. The system implements web-based online handwriting content recognition, facilitates operations such as content display, export, and document generation, and is built using the React development framework.

[0025] The transport layer is responsible for data transmission between the presentation layer and the service layer. This system mainly uses HTTP network request related technologies for data transmission, while maintaining the RESTful style in interface design;

[0026] The service layer is responsible for the implementation of all business functions of this system. It adopts SpringBoot technology and uses SpringAOP to implement the system's logging, exception handling and permission control functions. MyBatis is used as the persistence layer framework, and TensorFlow and Opencv related technologies are used to implement handwritten text recognition, text error correction, and text generation algorithms. At the same time, the FlaskPythonWeb framework is used to deploy the algorithm online. Finally, ResTemplate is used to realize the communication between the business server and the algorithm server.

[0027] The data layer is responsible for storing data such as instruction records and template information. This system uses the MySQL database to store system-related data, and uses the OOS object storage service to store various types of documents, copy format templates, generated copy content documents and other files.

[0028] 2. System Usage

[0029] This system is used for text view parsing and text document generation. Text view parsing includes handwriting recognition and content correction, and the service is available in both offline and online modes. Text view files can be uploaded to a storage service using object storage technology for offline recognition. Alternatively, online recognition interfaces can be called to provide real-time capabilities through interfaced services. Text document generation is powered by a template engine, automatically generating text content documents based on different templates.

[0030] The steps used by the system are as follows Figure 2 shown.

[0031] First, obtain the original file, convert it into an image format, and crop the area to obtain an image of the area containing only the handwritten signature content. This image is input into the handwriting recognition model to perform end-to-end text recognition on text sequences of variable length. For details on the specific process, see the "Handwriting Recognition" section.

[0032] The recognized text content is input into the text correction model, which includes error detection and error correction. Error detection obtains all suspected error locations. Then, through error correction, the words in the incorrect location are replaced with similar words. Finally, the sentence perplexity is calculated through the language model. All candidate set results are compared and sorted to obtain the optimal correction word and the corrected text content. For details on the process, see the "Content Correction" section.

[0033] When splicing data, relevant element data is retrieved from the database. In online mode, after each handwritten approval is completed, the approval content is identified through a service interface call. The approval system retrieves the approval time and the name of the approver, stores them in the database, and binds them to the corresponding file for persistence. In offline mode, files already on the object storage server can be selected for identification and element data splicing.

[0034] Before a document is circulated for review and approval, the corrected text is captured and a clean copy document is generated based on a clean copy format template. This document is then circulated along with the original document, with the document content updated with each circulation. When viewing the original document, you can select "View Approval Comments" to display the collection of approved comments in the corresponding text format.

[0035] 3. Handwriting recognition

[0036] Handwriting recognition first requires obtaining the original document with handwritten signatures, converting it into an image, obtaining the handwritten signature area as the input of the handwriting recognition model, and obtaining the handwritten text content through recognition.

[0037] 3.1 Region Identification

[0038] Document text and handwritten signatures are distinguished by color, and color-based object recognition is implemented using OpenCV. First, the image is grayscaled and filtered to convert it to an HSV image. The target color is identified using OpenCV's internal function library, and the region outline is extracted. Finally, the target's bounding rectangle is drawn and cropped to obtain an image containing only the handwritten signature.

[0039] 3.2 Content Identification

[0040] a. Algorithm selection

[0041] Due to personal writing habits, handwritten signatures often have varying font sizes and line lengths. To avoid manual segmentation, a sequence learning model was introduced. Given the varying number of characters required to decode an image, decoding audio into Chinese characters is essentially the same as decoding an image into Chinese characters. Therefore, it is entirely feasible to build an algorithm network guided by LSTM and CTC, which are widely used in speech recognition.

[0042] b. Dataset

[0043] The training and validation sets used the open-source CASIA-HWD handwritten Chinese character dataset. When reading data, we first obtain the paths to all images, assign corresponding labels to each image, convert the paths and label lists into corresponding vectors, and then slice them. The images are then read and preprocessed (including image inversion and brightness changes), and the labels are uniquely encoded to form shuffle batches, which are then fed into the model.

[0044] c. Algorithm network construction

[0045] The CNN+RNN+CTC method is used to recognize handwritten Chinese characters. The CNN extracts image features, while the RNN uses a bidirectional LSTM network (BiLSTM) to extract text sequence features based on convolutional features. CTCLoss replaces the softmax loss, eliminating the need for alignment of training samples, resolving the issue of inconsistent output and label lengths and avoiding strict manual alignment.

[0046] d. Handwriting recognition process

[0047] Handwriting recognition first detects the text position, then defines the text block slicing function, and finally performs character recognition on the sliced ​​single-word image.

[0048] The steps of handwriting recognition are as follows Figure 4 As shown in the figure, first build a handwriting recognition algorithm model and call it. Input a text image, slice it into text blocks, perform a perspective transformation based on four points, and transform the quadrilateral represented by the four points into a horizontal rectangle. Grayscale the image to be recognized, perform predictive recognition, and obtain the recognition result.

[0049] 4. Content Correction

[0050] a. Algorithm selection

[0051] Compared with RNN and LSTM, BERT can read the entire text sequence at one time, extract text information at multiple levels, and interpret text semantics more comprehensively. Therefore, the Soft-MaskedBert algorithm is used for modeling.

[0052] b. Dataset

[0053] When training a model, using a self-built dataset can often achieve customization of the corpus. Compared with general public datasets, it has unique advantages and can relatively accurately detect the performance of the model in a specific field.

[0054] Therefore, we capture the text content of public documents in different fields. Using the Jieba word segmentation library and the hit_stopwords stop word list, we segment the original corpus based on word frequency to form a dictionary that can be used for deep network training. Through word segmentation, we detect errors at both the character and word granularity levels. We integrate the suspected error results at both granularities to form a candidate set of suspected error positions. Then, through error correction, we replace the words in the incorrect position with similar words. Finally, by calculating the sentence perplexity, we compare and sort all the candidate set results to obtain the optimal correction word and the text content after error recognition. The original corpus is divided into short sentences with punctuation marks as interruptions, forming 10,000 text sequences as the real input corpus for deep network training.

[0055] c. Algorithm network construction

[0056] The error correction model is constructed with reference to the Soft-MaskedBert algorithm. The main body is divided into two parts: the error detection network and the error correction network. The error detection network is a bidirectional GRU (Gate Recurrent Unit) network, that is, a bidirectional gated recurrent unit. Its function is to predict the probability of errors in characters at various positions. The error correction network is a correction network based on BERT, which corrects and replaces the characters at the positions where the errors are successfully detected.

[0057] The error detection network and the error correction network are connected into a whole through Soft-Masked technology, and the output information of the error detection network is the input information of the BERT correction network.

[0058] d. Model training

[0059] During training, a masking strategy based on a confusion set is employed. 15% of the tokens in the training set are subjected to glyph-based confusion word replacement, random replacement, and the original word remains unchanged. The original text sentence is converted into a word vector using BERT's embedding mechanism. This vector is then fed into the bidirectional GRU of the error detection network. This network treats each character of the text data as a word and uses the bidirectional GRU to capture contextual information, learn the error probability of each character, and mask the characters most likely to be incorrect. The error correction network uses BERT to encode the masked character sequence, capture contextual information, and generate appropriate outputs. Based on the predicted results, the corrected character sequence is generated. Training is repeated over multiple epochs to achieve the final trained model.

[0060] e. Error correction process

[0061] The steps for text correction are as follows: Figure 4 shown.

[0062] The text output in the handwriting recognition stage is subjected to word frequency segmentation processing to form a word sequence that can be used for deep network training. The input is subjected to error correction modeling, error detection by a detection network to obtain all suspected error positions, error correction by a correction network using similar words to replace the words in the error positions, and finally calculation of the sentence perplexity to compare and sort all candidate set results to obtain the optimal correction word and obtain the corrected text content.

[0063] 5. Text generation

[0064] A text file is generated according to a template using the freemarker technology of SpringBoot.

[0065] A word template is prepared, and placeholders are defined, such as the name of the reviewer, the review time, the review content, etc. The placeholders are occupied using the ${string} method, and the placeholder field name is consistent with the review record instance property name. Different templates are constructed and uploaded for different copy formats, and different review-related attribute data is spliced to form a standard copy statement, so different template files need to be constructed and uploaded. Different templates are applied for management, and a certain template can be selected for application in the review process of a certain type of file. Different versions of the same template can be saved and restored to previous versions.

[0066] After the template file is uploaded, the word file is converted into an xml document, and then the xml format file is converted into an ftl format. Because the signing process is multi-node, there are multiple pieces of review data under the same file, so the freemarker dependency is introduced in the project to support the cyclic display of review content by adding the list tag of freemarker.

[0067] The steps of document generation are shown in Figure 5 .

[0068] The template file is read, and freemarker parses the corresponding tags. After the handwritten signature is submitted, the review file identifier is bound, and the review record is formed in the database. The list data of the review record is obtained from the database, the data is assembled, the variable replacement is performed according to the ftl template file, the file stream is formed, the word document is generated, the document content is consistent with the selected template, and the placeholders of the fields are replaced with the corresponding form field content.

[0069] Compared with the prior art, the application proposes a system design for the whole process of document review and copying based on artificial intelligence, and selects appropriate algorithms and optimizes the framework for each processing process of the task. For handwritten signature content recognition, a handwritten recognition algorithm model is established. First, the handwritten annotation area image is obtained to provide input for handwritten content recognition, and the handwritten content is extracted to avoid manual copying and reduce the difficulty and amount of manual recognition. In order to improve the accuracy of the final obtained content of the handwritten signature, a content error correction algorithm model is established to detect errors from the word granularity and word granularity through word segmentation, and the optimal correction word is obtained through the language model to complete the error correction, greatly improving the final accuracy of the handwritten content recognition; through the template engine technology, the document is automatically generated according to the constructed copying template, which saves the manual arrangement and document input work, and greatly improves the file transfer speed.

[0070] The above only describes the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled persons in the art, without departing from the technical principles of the present application, a number of improvements and modifications can be made, and these improvements and modifications should be considered as the protection scope of the present application.

Claims

1. A design method for an intelligent copying system for official document approval based on artificial intelligence, characterized in that: This intelligent document copying and clearing system is used for text view analysis and text document generation. Text view analysis includes handwriting recognition and content correction, and the service is available in both offline and online modes. Text document generation is supported by a template engine, automatically generating text content documents based on different templates. The system uses the following steps: First, the original file is obtained and converted into an image format. The image is then cropped to obtain an image of the area containing only the handwritten signature. This image is then input into the handwriting recognition model to perform end-to-end text recognition on text sequences of variable length. The recognized text content is input into the text error correction model, which includes error detection and error correction. Error detection obtains all suspected error locations. Then, error correction replaces the words in the incorrect location with similar words. Finally, the sentence perplexity is calculated using the language model. All candidate set results are compared and sorted to obtain the optimal correction word and the corrected text content. When splicing data, relevant element data is obtained from the database. In the online mode, after each handwritten approval is completed, the approval content is identified through a service interface call, and the approval time and name of the approver are obtained through the approval system. These data are stored in the database and bound to the corresponding files for persistence. In the offline mode, files already on the object storage server are selected for identification and element data is spliced. Before the document is circulated for review and approval, the text content after error correction is obtained, and a copy-clear content document is generated based on the copy-clear format template. The document is circulated together with the original document, and the document content is supplemented and updated each time it is circulated. When viewing the original text, you can choose to view the approval opinions and display the collection of approval content in the corresponding text format.

2. The method for designing an intelligent document copying system based on artificial intelligence according to claim 1, characterized in that: The system uploads the text view file to the storage service and uses object storage technology to store it for offline recognition, or the system calls the online recognition interface to provide real-time related capabilities through interface services.

3. The design method of the intelligent copying system for official document approval based on artificial intelligence according to claim 1 is characterized in that: The system is divided into four layers: presentation layer, transport layer, service layer and data layer; The presentation layer is the user-facing web client. This system implements web-based online handwriting content recognition for content display, export, and document generation. It is built using the React development framework. The transport layer is responsible for data transmission between the presentation layer and the service layer. It uses HTTP network request related technologies for data transmission, while maintaining the RESTful style in interface design. The service layer is responsible for implementing all business functions of this system. It adopts Spring Boot technology and uses Spring AOP to implement system logging, exception handling, and permission control functions. MyBatis is used as the persistence layer framework. TensorFlow and OpenCV related technologies are used to implement handwritten text recognition, text error correction, and text generation algorithms. The Flask Python Web framework is used to deploy the algorithms online. Finally, ResTemplate is used to realize communication between the business server and the algorithm server. The data layer is responsible for storing data such as instruction records and template information. It uses the MySQL database to store system-related data and the OOS object storage service to store various types of documents, copy format templates, and generated copy content document files.

4. The method for designing an intelligent copying system for official document approval based on artificial intelligence according to any one of claims 1 to 3, characterized in that: When performing handwriting recognition, we first need to obtain the original document with the handwritten signature, convert it into an image, and obtain the handwritten signature area as the input of the handwriting recognition model. After recognition, we can obtain the text content of the handwritten signature. Document text and handwritten signatures are distinguished by different colors. OpenCV is used to implement color-based target recognition: the image is first processed into grayscale, then filtered and converted into an HSV image. The target color is identified using the OpenCV internal function library, and the region outline is extracted. Finally, the target's bounding rectangle is drawn and cropped to obtain an image containing only the handwritten signature. The handwriting recognition process is as follows: first detect the text position, then define the text block slicing function, and finally perform character recognition on the single-word image obtained by slicing; specifically, first build a handwriting recognition algorithm model and call it, input the text image, slice it into text blocks, perform perspective transformation based on four points, transform the quadrilateral represented by the four points into a horizontal rectangular image, grayscale the image that needs to be recognized, perform predictive recognition, and obtain the recognition result.

5. The method for designing an intelligent copying system for official document approval based on artificial intelligence according to claim 4, characterized in that: The handwriting recognition algorithm model uses the open-source CASIA-HWD handwritten Chinese character dataset for training and validation. When reading the data, we first obtain the paths of all images, assign corresponding labels to each image, convert the paths and label lists into corresponding vectors, and slice them. Then, we read and preprocess the images, and use the unique hot encoding representation of the labels to form shuffle batches, which are then fed into the model. The handwriting recognition algorithm model uses the CNN+RNN+CTC method to recognize handwritten Chinese characters. CNN is responsible for extracting image features, and RNN uses a bidirectional LSTM network to further extract text sequence features based on convolutional features. CTC Loss is used instead of Softmax Loss, so that training samples do not need to be aligned.

6. The method for designing an intelligent document copying system based on artificial intelligence according to claim 4, characterized in that: When correcting content, the text output from the handwriting recognition stage is segmented based on word frequency to form a word sequence that can be used for deep network training. The sequence is then input into the error correction model. Error detection is performed through the error detection network in the error correction model to obtain all suspected error locations. The error correction network then corrects the errors by replacing the words in the incorrect positions with similar words. Finally, the sentence perplexity is calculated, and all candidate set results are compared and sorted to obtain the optimal correction word and the corrected text content.

7. The method for designing an intelligent copying system for official document approval based on artificial intelligence according to claim 6, characterized in that: When training the error correction model, the use of self-built data sets can often achieve customization of the corpus set; the text content of public documents in different fields is captured, and the original corpus is segmented by word frequency using the jieba word segmentation library and the hit_stopwords stop word list to form a dictionary that can be used for deep network training. Error detection is performed from both character granularity and word granularity through word segmentation, and the suspected error results of the two granularities are integrated to form a candidate set of suspected error positions. Then, through error correction, the words in the wrong position are replaced by similar words. Finally, by calculating the sentence perplexity, all candidate set results are compared and sorted to obtain the optimal correction word and the text content after error recognition; the original corpus is divided into short sentences with punctuation marks as interruptions to form 10,000 text sequences as the real input corpus for deep network training.

8. The method for designing an intelligent document copying system based on artificial intelligence according to claim 7, characterized in that: The error correction model consists of two parts: an error detection network and an error correction network. The error detection network is a bidirectional GRU network, which is used to predict the probability of errors at various character positions. The error correction network is a BERT-based correction network that corrects and replaces characters at positions where errors are successfully detected. The error detection network and the error correction network are connected into a whole using soft-masked technology. The output information of the error detection network is the input information of the BERT correction network. The error correction model is trained using a mask strategy based on a confusion set. 15% of the tokens in the training set are replaced with obfuscated words based on glyphs, randomly replaced, and the original words remain unchanged. The original text sentence is converted into a word vector through BERT's embedding mechanism. This vector is then input into the bidirectional GRU of the error detection network. The network treats each character of the text data as a word and uses the bidirectional GRU to capture contextual information, learn the error probability of each character, and mask the characters most likely to be erroneous. The error correction network uses BERT to encode the masked character sequence, capture contextual information and generate appropriate outputs, and generates a corrected character sequence based on the prediction results; multiple epochs are performed continuously during training to obtain the final training model.

9. The method for designing an intelligent copying system for official document approval based on artificial intelligence according to any one of claims 6 to 8, characterized in that: When generating text documents, SpringBoot uses FreeMarker technology to generate text files based on templates. A Word template is developed, and placeholders are defined, including the reviewer's name, review time, and review content. Placeholders are represented using ${string}, and the placeholder field names are consistent with the attribute names of the review record instance. For different copying formats, different review-related attribute data needs to be spliced ​​to form standard copying statements, so different template files are constructed and uploaded. Application management is performed for different templates, and a specific template is selected for application in the review process of a certain type of file. Different versions of the same template are saved for restoration to the previous version. After uploading the template file, convert the Word file into an XML document, and then save the XML format file into FTL format; introduce FreeMarker dependency into the project, and support the circular display of the instruction content by adding the FreeMarker list tag.

10. The method for designing an intelligent document copying system based on artificial intelligence according to claim 9, characterized in that: The text document generation process is as follows: read the template file, parse the corresponding tags with FreeMarker, and after the handwritten signature is submitted, bind the review file identifier to form an instruction record in the database; obtain the list data of the instruction record from the database, assemble the data, replace the variables according to the FTL template file, form a file stream, and generate a Word document. The document content is consistent with the selected template, and the field placeholder will be replaced with the corresponding form field content.

Images