Corpus knowledge base self-learning method and device based on comparative learning of gradient descent method

Abstract

The embodiment of the present application provides a corpus knowledge base self-learning method and device based on comparative learning using a gradient descent method, and relates to the technical field of archive digitization. The method comprises: obtaining scanned archive image data and corresponding recorded voice data; performing font recognition on the archive image data and converting it into first text information, and adding it to the record item field to obtain the record item field; performing voice recognition on the recorded voice and converting it into second text information to form an artificial directory; performing intelligent recognition on the record item field to obtain a machine directory; performing comparative analysis on the artificial directory and the machine directory to form a learning result; performing semantic retrieval analysis on the learning result; and applying artificial intelligence technology to the archive digitization operation process by means of big data processing and intelligent learning, so as to effectively improve the efficiency and output quality of the archive digitization method and accelerate the development of archive digitization.

Description

Technical Field

[0001] The present application relates to the technical field of archive digitization, and in particular to a corpus knowledge base self-learning method and device based on comparative learning using a gradient descent method. Background Art

[0002] With the advancement of technology, the process of archive digitization has been significantly accelerated. However, there are also certain problems in archive digitization, which has become a bottleneck restricting the development of archive digitization. At present, in the archive digitization workflow, most of the work, including scanning, cataloging, and proofreading, relies on manual operation by staff, which makes it easy to produce accidental errors in archive digitization, and the work efficiency is low. There are certain differences in the archive digitization resources completed by different undertakers. The quality of archive digitization depends to a large extent on the experience of the operator. There are differences in the quality of archive digitization practitioners, and the quality of archive digitization output is not uniform. Many archive work is both professional and diverse, and has the characteristics of programming and repetitiveness. It takes a long period of time to train a qualified archive professional. At the same time, there is a shortage of professional talents in archive digitization, which leads to the slow progress of the overall digital development of the industry. Summary of the invention

[0003] The purpose of the embodiments of the present application is to provide a corpus knowledge base self-learning method and device based on comparative learning using a gradient descent method, so as to solve the problems of low efficiency, inconsistent output quality, and slow development progress of current archive digitization methods.

[0004] In a first aspect, an embodiment of the present application provides a corpus knowledge base self-learning method based on comparative learning using a gradient descent method, comprising:

[0005] Obtain scanned archive image data and corresponding recorded voice data;

[0006] Based on a set font recognition rule, the file image data is subjected to font recognition and converted into first text information, and the first text information is added to the entry field to obtain the entry field;

[0007] Based on the set speech recognition rules, the recorded speech is recognized and converted into second text information, and combined with the recorded item fields to form an artificial directory, or the artificial directory is formed by the recognition result of the framed area of ​​the recorded item fields;

[0008] Based on the set catalog recognition rules, the catalog fields are intelligently identified to obtain the machine catalog;

[0009] Compare and analyze the manual catalog and the machine catalog to form learning results;

[0010] Conduct semantic retrieval analysis on the learning results.

[0011] In the above implementation process, the scanned archive image data and the corresponding recorded voice data are obtained; the font recognition of the archive image data is performed based on the set font recognition rules and converted into first text information, and the first text information is added to the record field to obtain the record field; the recorded voice is recognized based on the set voice recognition rules and converted into second text information, and an artificial directory is formed in combination with the record field, or an artificial directory is formed by identifying the result of the framed area of ​​the record field; the record field is intelligently identified based on the set record recognition rules to obtain a machine directory; the manual directory and the machine directory are compared and analyzed to form a learning result; the learning result is subjected to semantic retrieval analysis; and artificial intelligence technology is applied to the archive digitization operation process by means of big data processing and intelligent learning, so as to effectively improve the efficiency and output quality of the archive digitization method and accelerate the development of archive digitization.

[0012] Furthermore, the step of obtaining the scanned archive image data and the corresponding recorded voice data includes:

[0013] Use a scanner to scan the archive pages and obtain archive images containing traditional Chinese characters;

[0014] Use the image processing library to grayscale the archive image and convert the color image into a single-channel grayscale image;

[0015] The median filtering algorithm is used to remove the salt and pepper noise in the archival image, and the contrast of the archival image is enhanced by the histogram equalization method to obtain the final archival image data.

[0016] A microphone array is used to collect recorded speech signals; wherein the microphone array uses beamforming technology to perform weighted summation on the recorded speech signals collected by multiple microphones;

[0017] The collected recorded voice signal is converted into digital form, and then high-pass filtered by a digital signal processing algorithm to remove low-frequency noise, and then automatically gain controlled to obtain the final recorded voice data.

[0018] In the above implementation process, archival images are obtained and preprocessed to provide a good image basis for subsequent recognition work; recorded voice signals are obtained and preprocessed to obtain high-quality voice data, providing a good basis for subsequent voice recognition.

[0019] Furthermore, the font recognition of the archive image data is performed based on the set font recognition rule and converted into first text information, and the first text information is added to the entry field to obtain the entry field, including:

[0020] Based on the deep learning convolutional neural network model, the structure and stroke features of traditional Chinese characters in the archival image data are extracted and recognized. In the recognition process, the parameters of the neural network are automatically adjusted according to the image features.

[0021] Add the recognized first text information to a custom database field, and combine it with the full text recognition result to obtain a bibliographic item field;

[0022] Using a pre-trained language model based on a recurrent neural network, combined with a corpus containing a large number of traditional Chinese characters, error detection and correction of the entry fields;

[0023] The catalog fields and correction results are compared word by word, the difference features and error patterns are extracted, and then fed back to the deep learning model for retraining.

[0024] In the above implementation process, the convolutional neural network model based on deep learning can accurately extract and recognize the structure and stroke features of traditional Chinese characters; make corrections to ensure the accuracy of the data, optimize the model based on the comparison results, and continuously improve the recognition accuracy of traditional Chinese characters, achieving self-learning and continuous improvement.

[0025] Furthermore, the voice recognition of the recorded voice data based on the set voice recognition rules is converted into second text information, and the manual directory is formed by combining the recorded item fields, or the manual directory is formed by the recognition result of the framed area of ​​the recorded item fields, including:

[0026] Based on the deep learning speech recognition model, the recorded speech data is recognized; wherein, during the recognition process, the speech data is analyzed and recognized frame by frame according to the pre-trained language model and acoustic model, and converted into the second text information;

[0027] Using part-of-speech tagging and syntactic analysis techniques in natural language processing, a preliminary grammar and semantic check is performed on the identified second text information;

[0028] According to the preset recording rules, the designated recording item fields corresponding to the speech recognition results are determined and recorded in the log file, where the designated recording item fields include recording time, operator ID and file number;

[0029] Using speech recognition technology to convert the updated voice instructions into text instructions, analyze the text instructions, find the corresponding field in the database of the record field, add the new content to the field, and update the record log to record the detailed information of the editing operation, where the detailed information includes the editing time, operator, content before modification and content after modification;

[0030] Combine the fields of the entry to form an artificial catalogue;

[0031] Or, a manual catalogue can be formed by identifying the results of selecting the area in the description field.

[0032] In the above implementation process, the acoustic features and language features in the speech signal are effectively extracted and recognized, and the recognized information is preliminarily checked for grammar and semantics to improve the accuracy and logic of the text information; the recording operation is recorded in a detailed log file for subsequent data management and auditing; new content is added to the field, and the recording log is updated to record the detailed information of the editing operation, so as to realize real-time and accurate editing of the recording content and traceability of the data.

[0033] Furthermore, the intelligent identification of the entry fields based on the set entry identification rules to obtain the machine directory includes:

[0034] Obtaining the input manual entry information, verifying and storing it, and obtaining the first text information;

[0035] Use deep learning algorithms to conduct detailed comparative analysis of the first text information and the manual entry information, extract difference features and error patterns, and automatically adjust the parameters of the optical character recognition (OCR) model;

[0036] Integrate the recognition results of the optimized OCR model with the manual catalog information, remove redundant and erroneous information, generate and record archival information;

[0037] Store the integrated archival information in the database and classify, index and archive it according to the standards of archival management;

[0038] The pre-trained semantic analysis model based on natural language processing technology performs semantic understanding of archival information;

[0039] The grammatical structure tree of the sentence is constructed through syntactic analysis to determine the syntactic relationship between the components, and the semantic role of each word in the sentence is determined by using semantic role labeling technology to extract archive element information; wherein the archive element information includes key elements and their corresponding semantic relationships;

[0040] Automatically match the pre-set record field according to the extracted archive element information;

[0041] A semantic induction algorithm based on the combination of rules and machine learning is used to automatically summarize and organize the archival element information, and the summarized and organized information is accurately entered into the corresponding catalog items to complete the automatic cataloging of the archives and obtain a mechanical catalog.

[0042] In the above implementation process, the recognition results of the optimized OCR model are integrated with the manual cataloging information to remove redundant and erroneous information, generate archival information, and classify, index and archive it for subsequent machine self-learning; semantic understanding technology is used to extract the elements of the archival content, automatically summarize the archival content semantically, complete the automatic cataloging of the archives, and achieve efficient and accurate intelligent cataloging.

[0043] Furthermore, the comparison and analysis of the manual catalog and the machine catalog to form a learning result includes:

[0044] Use the gradient descent method to calculate the loss function to find the weight coefficient;

[0045] The loss function is used to express the error between the machine catalog and the manual catalog. In the gradient descent method, the weight parameter is calculated as follows:

[0046] ;

[0047] In the formula, lr represents the learning rate, which is a hyperparameter. When the learning rate is large, the update amplitude of the weight coefficient is faster, and when the learning rate is small, the update amplitude of the weight coefficient is small; Floss represents the loss function;

[0048] After several iterations, a weight coefficient is obtained that minimizes the output loss function.

[0049] The loss function uses mean square error:

[0050] ;

[0051] Among them, H represents the amount of data in the height direction, W represents the amount of data in the width direction, Y is the output value, I is the true value, and MSE calculates the sum of the differences between each data point between Y and I and finally averages them;

[0052] Among them, in the process of back propagation, the algorithm calculates the partial derivatives of the parameters of each layer layer by layer from front to back through the loss function, and updates all the parameters of each layer according to the results to form a learning result.

[0053] In the above implementation process, the optimal learning result is achieved.

[0054] Furthermore, the semantic retrieval analysis of the learning results includes:

[0055] Construct grammatical vectors for learning results, set weight parameters, set special case processing rules and set similarity calculation rules.

[0056] In the above implementation process, all archival resources related to the search phrase can be realized to improve the recording quality of Ming and Qing archival titles.

[0057] In a second aspect, the embodiment of the present application provides a corpus knowledge base self-learning device based on gradient descent method comparison learning, comprising:

[0058] A data acquisition module, used to acquire scanned archive image data and corresponding recorded voice data;

[0059] A font recognition module, used for performing font recognition on the archive image data based on a set font recognition rule and converting the font into first text information, and adding the first text information to the entry field to obtain the entry field;

[0060] A speech recognition module, used to perform speech recognition on the recorded speech based on a set speech recognition rule and convert it into a second text message, and form an artificial directory by combining the recorded item field, or form an artificial directory by recognizing the result of the framed area of ​​the recorded item field;

[0061] The directory recognition module is used to intelligently recognize the fields of the entry based on the set entry recognition rules to obtain the machine directory;

[0062] Comparison and analysis module, used to compare and analyze the manual catalog with the machine catalog to form learning results;

[0063] The semantic analysis module is used to perform semantic retrieval analysis on the learning results.

[0064] In a third aspect, an embodiment of the present application provides an electronic device, including:

[0065] A processor, a memory and a bus, wherein the processor is connected to the memory via the bus, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, they are used to implement the corpus knowledge base self-learning method based on comparative learning using the gradient descent method as described above.

[0066] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a server, implements the corpus knowledge base self-learning method based on comparative learning using the gradient descent method as described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0068] Figure 1A flowchart of a corpus knowledge base self-learning method based on comparative learning using a gradient descent method provided in an embodiment of the present application;

[0069] Figure 2 It is a structural schematic diagram of a corpus knowledge base self-learning device based on gradient descent method comparison learning provided in an embodiment of the present application;

[0070] Figure 3 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0072] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0073] Please see Figure 1 , Figure 1 A flowchart of a corpus knowledge base self-learning method based on gradient descent method comparison learning provided in an embodiment of the present application. The corpus knowledge base self-learning method based on gradient descent method comparison learning includes:

[0074] 100. Obtain the scanned archival image data and the corresponding recorded voice data.

[0075] It can be understood that the embodiment of the present application uses a corpus knowledge base self-learning system to perform data collection and various recognition analyses.

[0076] 110. Use a scanner to scan the archive pages and obtain archive images containing traditional Chinese characters.

[0077] 120. Use the image processing library to grayscale the archival image and convert the color image into a single-channel grayscale image.

[0078] 130. The median filtering algorithm is used to remove the salt and pepper noise in the archival image, and the contrast of the archival image is enhanced through the histogram equalization method to obtain the final archival image data.

[0079] Specifically, a high-precision scanner is used to scan ancient books or archive pages to obtain archive images containing traditional Chinese characters. Then, the archive images are grayed out using an image processing library (such as OpenCV) to convert color images into single-channel grayscale images, reducing the amount of data and computational complexity. Next, the median filter algorithm is used to remove salt and pepper noise in the archive images, and the image contrast is enhanced through the histogram equalization method, making the strokes of traditional Chinese characters clearer and more prominent, providing a good image foundation for subsequent recognition work.

[0080] 140. Use a microphone array to collect recorded voice signals; wherein the microphone array uses beamforming technology to perform weighted summation on the recorded voice signals collected by multiple microphones.

[0081] 150. After the collected recorded voice signal is converted into digital form, it is high-pass filtered through a digital signal processing algorithm to remove low-frequency noise, and then automatic gain control is performed to obtain the final recorded voice data.

[0082] Specifically, a microphone array is used to collect the voice of the recording processor. The microphone array uses beamforming technology to weight and sum the signals collected by multiple microphones to enhance the target voice signal while suppressing noise interference from other directions. After the collected voice signal is converted to digital, it is high-pass filtered through a digital signal processing algorithm to remove low-frequency noise, and then automatic gain control is performed to ensure the amplitude of the voice signal is stable and obtain high-quality voice data, providing a good foundation for subsequent voice recognition.

[0083] 200. Perform font recognition on the archive image data based on a set font recognition rule and convert it into first text information, and add the first text information to the entry field to obtain the entry field.

[0084] 210. Based on the deep learning convolutional neural network model, the structure and stroke features of traditional Chinese characters in archival image data are extracted and recognized; in the recognition process, the parameters of the neural network are automatically adjusted according to the image features.

[0085] Exemplarily, the preprocessed archival image data is input into an identification module integrated with a professional traditional Chinese character recognition SDK package. The recognition interface in this SDK package is based on a convolutional neural network model of deep learning, which accurately extracts and recognizes the structure and stroke features of traditional Chinese characters. Among them, during the recognition process, according to the characteristics of the archival image such as resolution, text size, font style, etc., the parameters of the neural network are automatically adjusted, such as the size of the convolutional kernel, the number of layers, the activation function, etc., to adapt to different traditional Chinese character recognition requirements. For example, for traditional Chinese characters with larger fonts and thicker strokes, a larger convolutional kernel is used for feature extraction to capture more macroscopic text structure information; for traditional Chinese characters with delicate strokes and unique font styles, the number of layers of the neural network is increased to improve the feature expression ability of the model, thereby improving the recognition accuracy.

[0086] It should be noted that the SDK package, that is, the Software Development Kit package, is a collection of development tools provided by a software provider or other organizations.

[0087] 220. Add the recognized first text information to the custom database field, and combine with the full-text recognition result to obtain the bibliographic item field.

[0088] Specifically, the system records the recognized traditional Chinese character text information into a custom database field, which is specifically used to store the text information extracted from the archival image; in addition, combining with the recognition result of the complete area of the archival image, the bibliographic item field is obtained. It can be understood that the full-text text recognition result of the embodiment of the present application includes the recognition of simplified Chinese characters.

[0089] 230. Use a pre-trained language model based on a recurrent neural network, combined with a corpus containing a large number of traditional Chinese characters, to perform error detection and correction on the bibliographic item field.

[0090] Exemplarily, the system records the recognized traditional Chinese character text information into a custom database field, which is specifically used to store the text information extracted from the image. At the same time, start the correction module, which uses a pre-trained language model based on a recurrent neural network and a corpus containing a large number of traditional Chinese characters for error detection and correction. For example, when the recognized result shows that "發" is misrecognized as "髪", the correction module analyzes the common contexts and usages of the character "發" in the corpus, and combines the understanding of the context semantics by the language model to determine that it should be "發" here, and corrects it. The corrected text information will replace the original misrecognized result to ensure the accuracy of the data.

[0091] 240. Compare the bibliographic item field and the correction result word by word, extract the difference features and error patterns, and feedback them to the deep learning model for retraining.

[0092] Exemplarily, the self-learning module obtains the recognition results (i.e., bibliographic item fields) and correction results, compares them word by word, and extracts the differential features and error patterns therefrom. For example, if it is found that the character "車" is repeatedly misrecognized as a wrong writing of the traditional Chinese character "车", these error cases are used as training samples and fed back to the deep learning model for retraining; through the backpropagation algorithm, the weights and parameters of the model are adjusted so that the model can learn the features and rules for correctly recognizing the character "車" and avoid similar errors from occurring again. With continuous comparison and learning of a large number of recognition results and correction results, the recognition model of the OCR engine will be gradually optimized, and the recognition accuracy for traditional Chinese characters will continue to improve, realizing the self-learning and continuous improvement of the system.

[0093] It should be noted that OCR (Optical Character Recognition) is a technology that can convert the text in an image into editable and searchable digital text. The OCR technology converts the characters in the image into a computer-readable format, such as TXT, DOC, PDF, etc., by recognizing the character shapes and patterns in the image.

[0094] 300. Perform speech recognition on the bibliographic speech based on the set speech recognition rules and convert it into second text information, combine it with the bibliographic item fields to form an artificial catalog, or form an artificial catalog through the recognition result of the selected area of the bibliographic item fields.

[0095] Optionally, the system calls the speech recognition interface in the speech recognition SDK package to automatically convert the speech information of the bibliographic processing staff into second text information; obtains the second text information returned by the speech recognition interface, records this information in the specified bibliographic item fields, and supports adding, modifying, and deleting the bibliographic content in real time through the speech recognition technology. In addition, combine the recognition result of the preset selected area of the bibliographic item fields to form an artificial catalog, where the preset selected area can be a traditional Chinese character text area or an area set according to requirements.

[0096] 310. Use a deep learning-based speech recognition model to recognize the bibliographic speech data; wherein, during the recognition process, perform frame-by-frame analysis and recognition on the speech data according to the pre-trained language model and acoustic model, and convert it into second text information.

[0097] 320. Use the part-of-speech tagging and syntactic analysis techniques in natural language processing to perform preliminary grammar and semantic checks on the recognized second text information.

[0098] Exemplarily, the preprocessed speech data is input into the speech recognition module, in which the speech recognition SDK package is based on a deep learning speech recognition model, such as a combination of a long short-term memory network (LSTM) and a convolutional neural network (CNN) architecture, to effectively extract and recognize the acoustic and language features in the speech signal. During the recognition process, the model analyzes and recognizes the speech data frame by frame based on the pre-trained language model and acoustic model, and converts it into the second text information. At the same time, using technologies such as part-of-speech tagging and syntactic analysis in natural language processing, the recognized text is subjected to a preliminary grammatical and semantic check to correct some common errors, such as word order reversal and misuse of part-of-speech, so as to improve the accuracy and logic of the text information.

[0099] 330. According to a preset recording rule, determine the designated recording item field corresponding to the speech recognition result and record it in a log file, wherein the designated recording item field includes recording time, operator ID and file number.

[0100] Exemplarily, the information recording module determines the designated recording item field corresponding to the speech recognition result according to the metadata and the pre-set recording rules. For example, for a personnel file, when the text information obtained by speech recognition is "Zhang San, male, born in May 1985", according to the recording specification of the personnel file, "Zhang San" is recorded in the "name" field, "male" is recorded in the "gender" field, and "born in May 1985" is recorded in the "date of birth" field, and these recording operations are recorded in a detailed log file, including the recording time, operator ID, file number and other information, so as to facilitate subsequent data management and auditing.

[0101] 340. Use speech recognition technology to convert the updated voice instructions into text instructions, analyze the text instructions, find the corresponding field in the database of the catalog field, add the new content to the field, and update the catalog log to record the detailed information of the editing operation, where the detailed information includes the editing time, operator, content before modification and content after modification.

[0102] For example, when the recording operator says the voice command "add XX position in XX company from 2010 to 2015 in work experience", the real-time editing module first converts it into a text command using voice recognition technology, and then analyzes the text command through the semantic parsing module to determine that the operation type is "add", the target recording item is "work experience", and the specific content is "XX position in XX company from 2010 to 2015". Then, the real-time editing module interacts with the information recording module, finds the corresponding archive record and "work experience" field in the database, adds the new content to the field, and updates the recording log to record the detailed information of this editing operation, including editing time, operator, content before modification, content after modification, etc., to achieve real-time and accurate editing of recording content and data traceability.

[0103] 350. Combine the entry fields to form an artificial directory.

[0104] 360. Or, by identifying the results of the selected area of ​​the entry field, an artificial directory is formed.

[0105] 400. Based on the set catalog recognition rules, intelligent recognition is performed on the catalog field to obtain the machine catalog.

[0106] Specifically, the system calls the selected item interface in the intelligent cataloging SDK package to determine all the items that need to be intelligently cataloged, and then calls the intelligent cataloging interface in the intelligent cataloging SDK package. The system transmits the archive image directory path parameters to be identified to the interface, and automatically records the cataloging information automatically identified and semantically analyzed by the interface into the selected items, which must include the file number information and the cataloging information of each item, and finally forms a machine directory.

[0107] 410. Obtain input manual entry information, verify and store it, and obtain first text information.

[0108] Specifically, in the training phase: the Traditional Chinese OCR recognition interface is called to recognize the scanned image, and the OCR engine automatically compares and learns the recognition results with the results of manually recorded items such as title, time, and responsible person.

[0109] Among them, manual catalog information input: archive management personnel enter the archive title, time, responsible person and other key catalog information in accordance with the prescribed format and requirements through the manual catalog information input module, and the system verifies and stores the input information.

[0110] 420. Use a deep learning algorithm to perform detailed comparative analysis on the first text information and the manual catalog information, extract difference features and error patterns, and automatically adjust the parameters of the optical character recognition (OCR) model.

[0111] Among them, comparison learning: The comparison learning module obtains OCR recognition results and manual catalog information, uses deep learning algorithms to perform detailed comparison analysis, extracts difference features and error patterns, and automatically adjusts the parameters of the OCR recognition model based on this information to achieve model optimization and self-learning.

[0112] 430. Integrate the recognition results of the optimized OCR model with the manual catalog information, remove redundant and erroneous information, generate archival information and record it.

[0113] 440. Store the integrated archival information in the database and classify, index and archive it according to the standards of archival management.

[0114] Among them, information integration and storage: integrate the optimized OCR recognition results with the manual catalog information, remove redundant and erroneous information, and generate complete and accurate archival information records; store the integrated archival information in the database, classify, index and archive it according to the standards of archival management, so as to facilitate subsequent machine self-learning.

[0115] 450. The pre-trained semantic analysis model based on natural language processing technology provides semantic understanding of archival information.

[0116] Specifically, after the comparison learning effect reaches the standard, semantic understanding technology is used to extract the elements of the archive content, automatically semantically summarize the responsible persons and titles of the archive content, match the catalog items, and automatically complete the cataloging.

[0117] Exemplarily, the pre-trained semantic analysis model in natural language processing technology (such as a language model based on the Transformer architecture) is used to perform semantic understanding on the archive content that has reached the standard after comparison learning. First, the lexical analysis tool is used to segment and tag the text. For example, "Zhang San wrote the "XX Research Report" in 1980" is annotated as "Zhang San / name in / preposition 1980 / time writing / verb / auxiliary word "XX Research Report" / book title". Then, the grammatical structure tree of the sentence is constructed through syntactic analysis to determine the syntactic relationship between the components. Finally, the semantic role labeling technology is used to determine the semantic role of each word in the sentence, such as "Zhang San" is the executor of the action of "writing", that is, the person in charge, and "XX Research Report" is the object of "writing", that is, the title; through these technical means, the key elements and their semantic relationships in the archive content are extracted.

[0118] 460. Construct a grammatical structure tree of the sentence through syntactic analysis to determine the syntactic relationship between the components, and use semantic role labeling technology to determine the semantic role of each word in the sentence to extract file element information; wherein, the file element information includes key elements and their corresponding semantic relationships.

[0119] 470. Automatically match pre-set catalog entry fields based on the extracted archival element information.

[0120] 480. Use a semantic induction algorithm based on a combination of rules and machine learning to automatically summarize and organize archival element information, accurately enter the summarized and organized information into the corresponding catalog items, complete the automatic cataloging of the archives, and obtain a mechanical catalog.

[0121] Exemplarily, according to the extracted archival element information, the pre-set catalog field is automatically matched. For example, for the extracted responsible person "Zhang San", it is matched to the "Author" field in the catalog; for the title "XX Research Report", it is matched to the "Document Name" field. At the same time, a semantic induction algorithm based on a combination of rules and machine learning is used to automatically summarize and organize key elements such as responsible persons and titles. For example, for multiple responsible persons "Zhang San, Li Si, Wang Wu", through cluster analysis and named entity recognition technology, they are summarized as "Zhang San, etc." to make them conform to the standard format of archival recording. Then, the summarized and organized information is accurately entered into the corresponding catalog, the automatic recording of the archives is completed, efficient and accurate intelligent recording is achieved, and a mechanical catalog is obtained.

[0122] 500. Compare and analyze the manual catalog and the machine catalog to form learning results.

[0123] Specifically, the system forms learning results by comparing the manual catalog with the machine catalog, analyzing the differences between intelligent cataloging and manual cataloging.

[0124] 510. Use the gradient descent method to calculate the loss function to find the weight coefficient.

[0125] Among them, the comparative learning method is to use the gradient descent method to calculate the loss function to find the appropriate weight coefficient W so that the output value of the loss function is minimized, thereby achieving the optimal learning result.

[0126] 520. Use the loss function to express the error between the machine catalog and the manual catalog. In the gradient descent method, the weight parameter is calculated as follows:

[0127] ;

[0128] In the formula, lr represents the learning rate, which is a hyperparameter. When the learning rate is large, the update amplitude of the weight coefficient is faster, and when the learning rate is small, the update amplitude of the weight coefficient is small; Floss represents the loss function;

[0129] The calculation of the output error is quantified through the loss function, so that the system can automatically perform reverse correction.

[0130] 530. After several iterations, a weight coefficient is obtained that minimizes the output loss function. The loss function uses the mean square error:

[0131] ;

[0132] Among them, H represents the amount of data in the height direction, W represents the amount of data in the width direction, Y is the output value, I is the true value, and MSE calculates the sum of the differences between each data point between Y and I and finally averages them;

[0133] Among them, in the process of back propagation, the algorithm calculates the partial derivatives of the parameters of each layer layer by layer from front to back through the loss function, and updates all the parameters of each layer according to the results to form a learning result.

[0134] 600. Perform semantic retrieval analysis on the learning results.

[0135] Specifically, a grammatical vector is constructed for the learning results, weight parameters are set, special case processing rules are set, and similarity calculation rules are set; all archival resources related to the search phrase can be realized to improve the cataloging quality of Ming and Qing archive titles.

[0136] Exemplary, grammatical vector construction: For example, for the sentence "beautiful flowers bloom in the garden", the words therein are first disambiguated. The word "beautiful" is determined to have its corresponding five-level code after disambiguation in the synonym forest. Similarly, the words "flowers", "garden", "bloom" and so on also obtain their corresponding five-level codes. These five-level codes are arranged in the order of the words in the sentence to construct the grammatical vector representation of the sentence, such as [the five-level code of "beautiful", the five-level code of "of", the five-level code of "flowers", the five-level code of "in", the five-level code of "garden", the five-level code of "in", the five-level code of "bloom"]. In this way, the sentence is presented in the form of a more semantically accurate grammatical vector, avoiding the interference that may be caused by the ambiguity of words.

[0137] Exemplary weight parameter setting: Analyze different parts of speech and grammatical components. For example, for noun components, based on their characteristics of usually carrying key semantic information in sentences, combined with a large number of text test analyses, a relatively high weight parameter (such as 1.2) is assigned; while for grammatical function words such as auxiliary words, since they mainly play the role of assisting the construction of sentence structure, a lower weight parameter (such as 0.1) is assigned. At the same time, set corresponding threshold parameters to distinguish dimensions of different importance. For example, when the ratio of the weight parameter of a certain dimension to the weight parameter of other dimensions is lower than a certain threshold parameter, its influence can be appropriately reduced in the similarity calculation. In this way, through reasonable parameter settings, the similarity calculation is more in line with the actual semantics and grammatical structure logic of the language.

[0138] Exemplary, special case handling: Assuming that when comparing the sentences "He is running" and "She is dancing", if a grammatical vector is constructed based on a certain feature extraction method, the position dimension corresponding to "running" in the second sentence is empty (Null). At this time, according to the rules of the present invention, the empty word is assigned the number "Ma01A01". When the similarity of the corresponding dimensions of the two sentences is subsequently calculated, the five-level codes corresponding to "Ma01A01" and "running" are calculated according to the specified similarity value (0.1), and when the two "Ma01A01" numbers appear in their respective sentences, the similarity value between them is zero. Through such unified processing, it is ensured that in the complex situation of the existence of empty words, the calculation of the similarity of the entire sentence can be carried out smoothly and reasonably, and the appearance of empty words will not cause confusion in the calculation logic or excessive deviation in the results.

[0139] For example, similarity calculation: Taking the sentences "sunlight shines on the earth" and "sunlight sprinkles all over the earth" as examples, first divide the sentences into sentence stems and sentence leaves. For example, consider "sunlight" and "earth" as the sentence stems, and "shining" and "sprinkled all over" as sentence leaves. According to the set sentence stem similarity (3), sentence leaf similarity (0.05), weight parameter (3.5), screening parameter (1) and other parameters, substitute them into the formula (sentence similarity = (3+0.05) / 3.51×1) to calculate and obtain the similarity value between the two sentences. This value can more accurately reflect the similarity between the two sentences at the semantic and grammatical levels, and provide a reliable similarity measurement basis for subsequent applications such as text classification and information retrieval.

[0140] As described above, the embodiment of the present application obtains scanned archive image data and corresponding recorded voice data; performs font recognition on the archive image data based on the set font recognition rules and converts it into first text information, and adds the first text information to the record field to obtain the record field; performs voice recognition on the recorded voice based on the set voice recognition rules and converts it into second text information, and forms an artificial directory in combination with the record field, or forms an artificial directory by identifying the result of the framed area of ​​the record field; performs intelligent recognition on the record field based on the set record recognition rules to obtain a machine directory; compares and analyzes the artificial directory with the machine directory to form a learning result; performs semantic retrieval analysis on the learning result; uses big data processing and intelligent learning and other means to apply artificial intelligence technology to the archive digitization operation process, effectively improves the efficiency of the archive digitization method and the output quality, and accelerates the development of archive digitization.

[0141] The above steps are not to be performed in a strict order as described in the numbers, but should be understood as an overall solution.

[0142] In the second aspect, based on the above embodiments, Figure 2 A schematic diagram of the structure of a corpus knowledge base self-learning device based on gradient descent method comparison learning provided in an embodiment of the present application. Figure 2 The corpus knowledge base self-learning device based on gradient descent method comparison learning provided in this embodiment specifically includes: a data acquisition module 201, a font recognition module 202, a speech recognition module 203, a directory recognition module 204, a comparison analysis module 205 and a semantic analysis module 206.

[0143] Among them, the data acquisition module 201 is used to acquire the scanned archival image data and the corresponding recorded voice data; the font recognition module 202 is used to perform font recognition on the archival image data based on the set font recognition rules and convert it into a first text information, and add the first text information to the record field to obtain the record field; the voice recognition module 203 is used to perform voice recognition on the recorded voice based on the set voice recognition rules and convert it into a second text information, and form an artificial directory in combination with the record field, or form an artificial directory by identifying the result of the framed area of ​​the record field; the directory recognition module 204 is used to perform intelligent recognition of the record field based on the set record recognition rules to obtain a machine directory; the comparison and analysis module 205 is used to compare and analyze the artificial directory with the machine directory to form a learning result; the semantic analysis module 206 is used to perform semantic retrieval analysis on the learning result.

[0144] As described above, the embodiment of the present application obtains scanned archive image data and corresponding recorded voice data; performs font recognition on the archive image data based on the set font recognition rules and converts it into first text information, and adds the first text information to the record field to obtain the record field; performs voice recognition on the recorded voice based on the set voice recognition rules and converts it into second text information, and forms an artificial directory in combination with the record field, or forms an artificial directory by identifying the result of the framed area of ​​the record field; performs intelligent recognition on the record field based on the set record recognition rules to obtain a machine directory; compares and analyzes the artificial directory with the machine directory to form a learning result; performs semantic retrieval analysis on the learning result; uses big data processing and intelligent learning and other means to apply artificial intelligence technology to the archive digitization operation process, effectively improves the efficiency of the archive digitization method and the output quality, and accelerates the development of archive digitization.

[0145] The corpus knowledge base self-learning device based on comparative learning by gradient descent method provided in the embodiment of the present application can be used to execute the corpus knowledge base self-learning method based on comparative learning by gradient descent method provided in the above embodiment, and has corresponding functions and beneficial effects.

[0146] In a third aspect, an embodiment of the present application further provides an electronic device, which can integrate the corpus knowledge base self-learning device based on gradient descent method comparison learning provided in an embodiment of the present application. Figure 3 Schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Figure 3 The electronic device includes: an input device 33, an output device 34, a memory 32 and one or more processors 31; the memory 32 is used to store one or more programs; when the one or more programs are executed by the one or more processors 31, the one or more processors 31 implement the corpus knowledge base self-learning method based on gradient descent method comparison learning as provided in the above embodiment. The input device 33, the output device 34, the memory 32 and the processor 31 can be connected by a bus or other means, Figure 3 The example of connecting through bus is taken in the following.

[0147] The processor 31 executes various functional applications and data processing of the device by running the software programs, instructions and modules stored in the memory 32, that is, realizes the above-mentioned corpus knowledge base self-learning method based on comparative learning of the gradient descent method.

[0148] The electronic device provided above can be used to execute the corpus knowledge base self-learning method based on gradient descent method comparison learning provided in the above embodiment, and has corresponding functions and beneficial effects.

[0149] In a fourth aspect, an embodiment of the present application also provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the corpus knowledge base self-learning method based on gradient descent method comparison learning as described above, and can achieve the same beneficial effects.

[0150] Of course, the storage medium containing computer executable instructions provided in an embodiment of the present application is not limited to the corpus knowledge base self-learning method based on comparative learning by gradient descent method as described above, and can also execute related operations in the corpus knowledge base self-learning method based on comparative learning by gradient descent method provided in any embodiment of the present application.

[0151] In a fifth aspect, the embodiments of the present application also provide a computer program product, and the methods described in the various embodiments of the present application can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instruction is loaded and executed on a computer, the processes or functions described in the various embodiments of the present application are executed in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user device, a core network device, an OAM (Open Application Model) or other programmable device.

[0152] The computer program or instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer program or instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired or wireless means. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; it may also be an optical medium, such as a digital video disk; it may also be a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include both volatile and non-volatile types of storage media.

[0153] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0154] In addition, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0155] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for enabling an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0156] The above description is only an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0157] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0158] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

Claims

1. A corpus knowledge base self-learning method based on gradient descent method comparison learning, characterized in that: The method comprises: Obtain scanned archive image data and corresponding recorded voice data; Based on a set font recognition rule, the file image data is subjected to font recognition and converted into first text information, and the first text information is added to the entry field to obtain the entry field; Based on the set speech recognition rules, the recorded speech is recognized and converted into second text information, and combined with the recorded item fields to form an artificial directory, or the artificial directory is formed by the recognition result of the framed area of ​​the recorded item fields; Based on the set catalog recognition rules, the catalog fields are intelligently identified to obtain the machine catalog; Compare and analyze the manual catalog and the machine catalog to form learning results; Conduct semantic retrieval analysis on learning results; The intelligent identification of the entry fields based on the set entry identification rules to obtain the machine directory includes: Obtaining the input manual entry information, verifying and storing it, and obtaining the first text information; Use deep learning algorithms to compare and analyze the first text information and the manual entry information, extract difference features and error patterns, and automatically adjust the parameters of the optical character recognition (OCR) model; Integrate the recognition results of the optimized OCR model with the manual catalog information, remove redundant and erroneous information, generate and record archival information; Store the integrated archival information in the database and classify, index and archive it according to the standards of archival management; The pre-trained semantic analysis model based on natural language processing technology performs semantic understanding of archival information. After the comparison and analysis results meet the standards, semantic understanding technology is used to extract the elements of the archival content, automatically perform semantic induction on the responsible persons and titles of the archival content, match the catalog items, and automatically complete the cataloging.

2. The corpus knowledge base self-learning method based on gradient descent method comparison learning according to claim 1 is characterized in that: The step of obtaining the scanned archive image data and the corresponding recorded voice data includes: Use a scanner to scan the archive pages and obtain archive images containing traditional Chinese characters; Use the image processing library to grayscale the archive image and convert the color image into a single-channel grayscale image; The median filtering algorithm is used to remove the salt and pepper noise in the archival image, and the contrast of the archival image is enhanced by the histogram equalization method to obtain the final archival image data. A microphone array is used to collect recorded speech signals; wherein the microphone array uses beamforming technology to perform weighted summation on the recorded speech signals collected by multiple microphones; The collected recorded voice signal is converted into digital form, and then high-pass filtered by a digital signal processing algorithm to remove low-frequency noise, and then automatically gain controlled to obtain the final recorded voice data.

3. The corpus knowledge base self-learning method based on gradient descent method comparison learning according to claim 1 is characterized in that: The method of performing font recognition on the archive image data based on the set font recognition rule and converting the font into first text information, and adding the first text information to the entry field to obtain the entry field, includes: Based on the deep learning convolutional neural network model, the structure and stroke features of traditional Chinese characters in the archival image data are extracted and recognized. In the recognition process, the parameters of the neural network are automatically adjusted according to the image features. Add the recognized first text information to a custom database field, and combine it with the full text recognition result to obtain a bibliographic item field; Using a pre-trained language model based on a recurrent neural network, combined with a corpus containing a large number of traditional Chinese characters, error detection and correction of the entry fields; The catalog fields and correction results are compared word by word, the difference features and error patterns are extracted, and then fed back to the deep learning model for retraining.

4. The corpus knowledge base self-learning method based on gradient descent method comparison learning according to claim 1 is characterized in that: The performing speech recognition on the recorded speech based on the set speech recognition rules and converting it into second text information, and forming an artificial directory in combination with the recorded item fields, or forming an artificial directory by recognizing the result of the framed area of ​​the recorded item fields, includes: Based on the deep learning speech recognition model, the recorded speech data is recognized; wherein, during the recognition process, the speech data is analyzed and recognized frame by frame according to the pre-trained language model and acoustic model, and converted into the second text information; Using part-of-speech tagging and syntactic analysis techniques in natural language processing, a preliminary grammar and semantic check is performed on the identified second text information; According to the preset recording rules, the designated recording item fields corresponding to the speech recognition results are determined and recorded in the log file, where the designated recording item fields include recording time, operator ID and file number; Using speech recognition technology to convert the updated voice instructions into text instructions, analyze the text instructions, find the corresponding field in the database of the record field, add the new content to the field, and update the record log to record the detailed information of the editing operation, where the detailed information includes the editing time, operator, content before modification and content after modification; Combine the fields of the entry to form an artificial catalogue; Or, a manual catalogue can be formed by identifying the results of selecting the area in the description field.

5. The corpus knowledge base self-learning method based on gradient descent method comparison learning according to claim 3 is characterized in that: The matching entry automatically completes the recording, including: The grammatical structure tree of the sentence is constructed through syntactic analysis to determine the syntactic relationship between the components, and the semantic role of each word in the sentence is determined by using semantic role labeling technology to extract archive element information; wherein the archive element information includes key elements and their corresponding semantic relationships; Automatically match the pre-set record field according to the extracted archive element information; A semantic induction algorithm based on a combination of rules and machine learning is used to automatically summarize and organize archival element information, and the summarized and organized information is accurately entered into the corresponding catalog items to complete the automatic cataloging of the archives and obtain a machine catalog.

6. The corpus knowledge base self-learning method based on gradient descent method comparison learning according to claim 1 is characterized in that: The comparison and analysis of the manual catalog and the machine catalog to form a learning result includes: Use the gradient descent method to calculate the loss function to find the weight coefficient; The loss function is used to express the error between the machine catalog and the manual catalog. In the gradient descent method, the weight parameter is calculated as follows: ； In the formula, lr represents the learning rate, which is a hyperparameter. When the learning rate is large, the update amplitude of the weight coefficient is faster, and when the learning rate is small, the update amplitude of the weight coefficient is small; Floss represents the loss function; After several iterations, a weight coefficient is obtained that minimizes the output loss function. The loss function uses mean square error: ； Among them, H represents the amount of data in the height direction, W represents the amount of data in the width direction, Y is the output value, I is the true value, and MSE calculates the sum of the differences between each data point between Y and I and finally averages them; Among them, in the process of back propagation, the algorithm calculates the partial derivatives of the parameters of each layer from front to back through the loss function layer by layer, and updates all the parameters of each layer according to the results to form a learning result.

7. The corpus knowledge base self-learning method based on gradient descent method comparison learning according to claim 1 is characterized in that: The performing semantic retrieval analysis on the learning results includes: Construct grammatical vectors for learning results, set weight parameters, set special case handling rules and set similarity calculation rules.

8. A corpus knowledge base self-learning device based on gradient descent method comparison learning, characterized in that: include: A data acquisition module, used to acquire scanned archive image data and corresponding recorded voice data; A font recognition module, used for performing font recognition on the archive image data based on a set font recognition rule and converting the font into first text information, and adding the first text information to the entry field to obtain the entry field; A speech recognition module, used to perform speech recognition on the recorded speech based on a set speech recognition rule and convert it into a second text message, and form an artificial directory by combining the recorded item fields, or form an artificial directory by recognizing the result of the framed area of ​​the recorded item fields; The directory recognition module is used to intelligently recognize the fields of the entry based on the set entry recognition rules to obtain the machine directory; The comparison and analysis module is used to compare and analyze the manual catalog with the machine catalog to form learning results; Semantic analysis module, used to perform semantic retrieval analysis on learning results; The intelligent identification of the entry fields based on the set entry identification rules to obtain the machine directory includes: Obtaining the input manual entry information, verifying and storing it, and obtaining the first text information; Use deep learning algorithms to compare and analyze the first text information and the manual entry information, extract difference features and error patterns, and automatically adjust the parameters of the optical character recognition (OCR) model; Integrate the recognition results of the optimized OCR model with the manual catalog information, remove redundant and erroneous information, generate and record archival information; Store the integrated archival information in the database and classify, index and archive it according to the standards of archival management; The pre-trained semantic analysis model based on natural language processing technology performs semantic understanding of archival information. After the comparison and analysis results meet the standards, semantic understanding technology is used to extract the elements of the archival content, automatically perform semantic induction on the responsible persons and titles of the archival content, match the catalog items, and automatically complete the cataloging.

9. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the processor is connected to the memory via the bus, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the method is used to implement the corpus knowledge base self-learning method based on gradient descent method comparison learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a server, implements the corpus knowledge base self-learning method based on gradient descent method comparison learning as described in any one of claims 1-7.

Images