Asset registration map construction method and device, computer device, and medium

Abstract

The present disclosure relates to an asset registration graph construction method and device, a computer device and a medium to solve the problem that it is inconvenient to verify asset registration information. The method comprises: for the obtained asset registration table, extracting the text information of the asset registration table; for each sentence in the text information, preprocessing the sentence, and inputting the combined vector matrix of the sentence obtained after preprocessing into a natural language processing model to obtain the label type result of each character of the sentence output by the natural language processing model; constructing asset registration information according to each label type result, and querying the historical registration information corresponding to the historical registration number in the asset registration information from an asset information library; and constructing an asset registration graph according to the historical registration information and the asset registration information. In this way, the asset registration graph constructed based on the historical registration information and the asset registration information can improve the convenience of verifying asset registration information.

Description

Technical Field

[0001] This disclosure relates to the field of asset verification technology, and in particular to methods, apparatus, computer devices and media for constructing asset registration maps. Background Technology

[0002] In areas such as movable asset financing, the Central Bank Credit Reference Center's Unified Registration Platform for Movable Asset Financing (hereinafter referred to as the CSRC Platform) serves as a crucial risk management platform. Before lending, businesses can use this platform to check the registration information of the financing targets, preventing duplicate financing and identifying risks in advance. Furthermore, during the lending process, financing and registration information can be recorded, indicating asset ownership, prioritizing asset repayment, and avoiding legal disputes. Currently, searches can only be conducted using the registration number on the platform. After retrieving all registration information, manual verification of the relevant registration details is required. Summary of the Invention

[0003] To overcome the problems existing in related technologies, this disclosure provides a method, apparatus, computer device and medium for constructing an asset registration map, so as to solve the problem of inconvenience in verifying asset registration information, thereby reducing the pressure of manpower to verify asset registration information and improving the efficiency of asset registration information verification.

[0004] According to a first aspect of the present disclosure, a method for constructing an asset registration map is provided, comprising:

[0005] For the obtained asset registration form, extract the text information from the asset registration form;

[0006] For each sentence in the text information, the sentence is preprocessed, and the combined vector matrix of the preprocessed sentence is input into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model.

[0007] Based on the results of each of the aforementioned tag types, asset registration information is constructed, and historical registration information corresponding to the historical registration number in the asset registration information is queried from the asset information database;

[0008] An asset registration map is constructed based on the historical registration information and the asset registration information.

[0009] Optionally, the step of extracting text information from the asset registration form includes:

[0010] The target style template for the asset registration form is determined from the style template library;

[0011] By comparing the table style of the target style template, determine the key feature characters of the asset registration form and their corresponding value, as well as the feature character coordinates of the key feature characters and the value coordinates of the value.

[0012] The asset registration form is reconstructed based on the key feature characters, the coordinates of the feature characters, the value, and the coordinates of the value, and the text information is extracted from the key feature characters of the reconstructed asset registration form.

[0013] Optionally, the style template is obtained in the following way:

[0014] Extract text and table lines from the asset registration form;

[0015] Based on preset key feature characters, match template key feature characters from the text extraction results; and determine the template character coordinates of the template key feature characters;

[0016] The coordinates of the template character value corresponding to the key feature character of the template are determined based on the template character coordinates.

[0017] Based on the table line extraction results, the template character coordinates, and the template character value coordinates, template table lines are constructed to obtain the initial style template.

[0018] Using the coordinates of the template characters as the target coordinates, the key feature characters of the template are filled into the initial style template to obtain the style template.

[0019] Optionally, preprocessing of each sentence in the text information includes:

[0020] Each sentence in the text information is split into characters and words to obtain the character splitting results and word splitting results;

[0021] The word segmentation results are aligned using the length of the character segmentation results as the target length.

[0022] The word vector layer converts the character segmentation result index into character vectors of a preset dimension, and constructs a character vector matrix based on the length of the character segmentation result and the dimension of the character vectors; and,

[0023] The word vector layer converts the aligned word segmentation result index into word vectors of a preset dimension, and constructs a word vector matrix based on the length of the aligned word matrix segmentation result and the dimension of the word vectors.

[0024] The combined vector matrix is ​​obtained by indexing and combining the character vector matrix and the word vector matrix.

[0025] Optionally, the natural language processing model includes a convolutional neural network model and a conditional random field model, and the training of the natural language processing model includes:

[0026] The combined vector matrix corresponding to each sentence in the sample text information is used as the training sample for the convolutional layer to perform convolution calculation, thereby obtaining the effective information of the sentence.

[0027] The effective information is pooled through a pooling layer, and the pooled effective information is calculated through a fully connected layer to obtain the tag type score for each character in the sentence.

[0028] A matrix is ​​constructed based on the label type score of each character in the sentence to obtain the emission probability matrix;

[0029] The label type score is backpropagated, the gradient of the error at each weight in the convolutional network is calculated, and the parameters of the filters in the convolutional neural network model are adjusted using gradient descent; and,

[0030] The transition probability matrix is ​​obtained through random initialization, and the emission probability matrix and the transition probability matrix are backpropagated to correct the model parameters of the conditional random field model.

[0031] Optionally, the step of extracting text information from the asset registration form includes:

[0032] Text information is extracted from the original code of the asset registration form; and,

[0033] Determine whether the extracted text information includes escape characters;

[0034] If the text information includes escape characters, the text information is extracted after the escape characters are made readable.

[0035] Optionally, the asset registration information includes any one of the following: initial asset registration information, asset change registration information, and asset extension registration information.

[0036] According to a second aspect of the present disclosure, an asset registration map construction apparatus is provided, comprising:

[0037] The extraction module is configured to extract text information from the acquired asset registration form.

[0038] The input module is configured to preprocess each sentence in the text information, and input the combined vector matrix of the preprocessed sentence into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model.

[0039] The query module is configured to construct asset registration information based on the result of each of the aforementioned tag types, and to query historical registration information corresponding to the historical registration number in the asset registration information from the asset information database.

[0040] The construction module is configured to construct an asset registration map based on the historical registration information and the asset registration information.

[0041] Optionally, the extraction module is configured as follows:

[0042] The target style template for the asset registration form is determined from the style template library;

[0043] By comparing the table style of the target style template, determine the key feature characters of the asset registration form and their corresponding value, as well as the feature character coordinates of the key feature characters and the value coordinates of the value.

[0044] The asset registration form is reconstructed based on the key feature characters, the coordinates of the feature characters, the value, and the coordinates of the value, and the text information is extracted from the key feature characters of the reconstructed asset registration form.

[0045] Optionally, the style template is obtained in the following way:

[0046] Extract text and table lines from the asset registration form;

[0047] Based on preset key feature characters, match template key feature characters from the text extraction results; and determine the template character coordinates of the template key feature characters;

[0048] The coordinates of the template character value corresponding to the key feature character of the template are determined based on the template character coordinates.

[0049] Based on the table line extraction results, the template character coordinates, and the template character value coordinates, template table lines are constructed to obtain the initial style template.

[0050] Using the coordinates of the template characters as the target coordinates, the key feature characters of the template are filled into the initial style template to obtain the style template.

[0051] Optionally, the input module is configured as follows:

[0052] Each sentence in the text information is split into characters and words to obtain the character splitting results and word splitting results;

[0053] The word segmentation results are aligned using the length of the character segmentation results as the target length.

[0054] The word vector layer converts the character segmentation result index into character vectors of a preset dimension, and constructs a character vector matrix based on the length of the character segmentation result and the dimension of the character vectors; and,

[0055] The word vector layer converts the aligned word segmentation result index into word vectors of a preset dimension, and constructs a word vector matrix based on the length of the aligned word matrix segmentation result and the dimension of the word vectors.

[0056] The combined vector matrix is ​​obtained by indexing and combining the character vector matrix and the word vector matrix.

[0057] Optionally, the natural language processing model includes a convolutional neural network model and a conditional random field model, and the training of the natural language processing model includes:

[0058] The combined vector matrix corresponding to each sentence in the sample text information is used as the training sample for the convolutional layer to perform convolution calculation, thereby obtaining the effective information of the sentence.

[0059] The effective information is pooled through a pooling layer, and the pooled effective information is calculated through a fully connected layer to obtain the tag type score for each character in the sentence.

[0060] A matrix is ​​constructed based on the label type score of each character in the sentence to obtain the emission probability matrix;

[0061] The label type score is backpropagated, the gradient of the error at each weight in the convolutional network is calculated, and the parameters of the filters in the convolutional neural network model are adjusted using gradient descent; and,

[0062] The transition probability matrix is ​​obtained through random initialization, and the emission probability matrix and the transition probability matrix are backpropagated to correct the model parameters of the conditional random field model.

[0063] Optionally, the extraction module is configured as follows:

[0064] Text information is extracted from the original code of the asset registration form; and,

[0065] Determine whether the extracted text information includes escape characters;

[0066] If the text information includes escape characters, the text information is extracted after the escape characters are made readable.

[0067] Optionally, the asset registration information includes any one of the following: initial asset registration information, asset change registration information, and asset extension registration information.

[0068] According to a third aspect of the present disclosure, a computer apparatus is provided, comprising:

[0069] processor;

[0070] Memory used to store processor-executable instructions;

[0071] The processor is configured as follows:

[0072] For the obtained asset registration form, extract the text information from the asset registration form;

[0073] For each sentence in the text information, the sentence is preprocessed, and the combined vector matrix of the preprocessed sentence is input into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model.

[0074] Based on the results of each of the aforementioned tag types, asset registration information is constructed, and historical registration information corresponding to the historical registration number in the asset registration information is queried from the asset information database;

[0075] An asset registration map is constructed based on the historical registration information and the asset registration information.

[0076] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided that stores computer program instructions thereon, which, when executed by a processor, implement the steps of the asset registry map construction method according to any one of the first aspects.

[0077] The technical solutions provided by the embodiments of this disclosure can include the following beneficial effects: For the obtained asset registration form, text information is extracted; for each sentence in the text information, the sentence is preprocessed, and the combined vector matrix of the preprocessed sentence is input into a natural language processing model to obtain the label type result for each character in the sentence output by the natural language processing model; asset registration information is constructed based on each label type result, and historical registration information corresponding to the historical registration number in the asset registration information is queried from the asset information database; an asset registration map is constructed based on the historical registration information and the asset registration information. Thus, the asset registration map constructed based on the historical registration information and the asset registration information can improve the convenience of verifying asset registration information, and at the same time, improve the accuracy of verifying asset registration information, thereby reducing the pressure of manual verification of asset registration information and improving the efficiency of asset registration information verification.

[0078] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0079] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0080] Figure 1 This is a flowchart illustrating an asset registration map construction method according to an exemplary embodiment.

[0081] Figure 2 This is a schematic diagram of an asset registration map according to an exemplary embodiment.

[0082] Figure 3 This is an implementation illustrated according to an exemplary embodiment. Figure 1 The flowchart for step S11.

[0083] Figure 4 This is a flowchart illustrating a method for generating a style template according to an exemplary embodiment.

[0084] Figure 5 This is a flowchart illustrating a preprocessing method for text information according to an exemplary embodiment.

[0085] Figure 6 This is a flowchart illustrating the training of a natural language processing model according to an exemplary embodiment.

[0086] Figure 7 This is a schematic diagram illustrating the structure of a neural network model according to an exemplary embodiment.

[0087] Figure 8 This is a block diagram illustrating an asset registry map construction apparatus according to an exemplary embodiment.

[0088] Figure 9 This is a block diagram illustrating an apparatus for constructing an asset registry map according to an exemplary embodiment. Detailed Implementation

[0089] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0090] Furthermore, it is worth noting that, for the sake of simplicity, the method embodiments provided in this disclosure are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions. For example, terms such as "S53" and "S54" are used to distinguish method steps and should not be construed as describing a specific execution order. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to this disclosure.

[0091] In relevant scenarios, manually downloading relevant registration information from the website provided by the relevant platform and manually verifying it is cumbersome and inconvenient. Furthermore, due to the unstructured nature of the data on the China Securities Depository and Clearing Corporation (CSDC), the verification process is lengthy, fragmented, and inefficient with low accuracy. In addition, the number of asset transfers is high, making it difficult for manual verification to check every single registration.

[0092] To address this issue, this disclosure provides a method, apparatus, computer device, and medium for constructing an asset registration map, thereby reducing the burden of manual verification of asset registration information and improving the efficiency of asset registration information verification.

[0093] Figure 1 This is a flowchart illustrating an asset registration map construction method according to an exemplary embodiment. This method can be applied to a computer terminal or a server. This disclosure uses an application to a computer terminal as an example for illustration. Figure 1 As shown, it includes the following steps.

[0094] In step S11, text information is extracted from the obtained asset registration form.

[0095] In step S12, for each sentence in the text information, the sentence is preprocessed, and the combined vector matrix of the preprocessed sentence is input into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model.

[0096] In step S13, asset registration information is constructed based on the result of each tag type, and historical registration information corresponding to the historical registration number in the asset registration information is queried from the asset information database.

[0097] In step S14, an asset registration map is constructed based on historical registration information and asset registration information.

[0098] Optionally, the asset register may include a movable property register and a real property register.

[0099] Specifically, in step S11, before extracting text information from the obtained asset registration form, the following steps are included:

[0100] Decompress the obtained asset archive; and,

[0101] Based on semantic analysis, the asset registration form is obtained from the decompressed asset package according to the name information of the asset compressed package.

[0102] For example, the system obtains the user's asset registration information from the unified registration platform for movable property financing at the central bank's credit reference center. Upon receiving the asset registration information, it sends the relevant registration number to the unified registration platform to obtain the asset compressed package corresponding to that registration number. For instance, by sending an initial registration number to the unified registration platform, it obtains the asset compressed package corresponding to that initial registration number. Based on semantic analysis, it determines that the name of the asset compressed package is "Initial Registration Information and Supporting Documents," and after decompressing the asset compressed package, it analyzes the names of the files within the package. Then, based on the "Initial Registration Information" included in the asset compressed package name, it obtains the "Initial Registration Form."

[0103] Alternatively, the asset registration form can be read using a PDF (Portable Document Format) reader via OCR (Optical Character Recognition) to obtain the text information. Alternatively, the original PDF code can be obtained, and the text information extracted from the original code.

[0104] Based on the above embodiments, in step S11, the extraction of text information from the asset registration form includes:

[0105] Text information is extracted from the original code of the asset registration form; and,

[0106] Determine whether the extracted text information includes escape characters;

[0107] If the text information includes escape characters, the text information is extracted after the escape characters are made readable.

[0108] In one implementation, if the OCR result includes unrecognizable text, the text information can be extracted by obtaining the original code of the text. Furthermore, if the text information extracted by the original code includes escape characters, the escape characters are processed to make them readable before the text information extraction is completed.

[0109] The natural language processing model uses the combined vector matrix corresponding to each sentence in the sample text information as the convolution training sample to perform convolution training, and obtains the label type score of each character in the sentence. The matrix constructed based on the label type score of each character in the sentence is used as the emission probability matrix to train the conditional random field model.

[0110] Based on the above embodiments, the asset registration information includes any one of the following: initial asset registration information, asset change registration information, and asset extension registration information.

[0111] It is understandable that when the asset registration information is for the initial registration of the asset, the historical registration number of the asset registration information will not exist in the asset information database. When the asset registration information is for the change of asset registration, the asset information database may contain the initial registration number, the change of asset registration number, and the extension of asset registration number.

[0112] In one approach, the asset registry graph includes graph nodes and corresponding edges, such as... Figure 2 As shown, the asset registration graph is constructed by using the registration time of historical registration information and asset registration information as graph nodes and single arrows as corresponding edges to represent the time flow of registration information.

[0113] For example, when historical registration information includes initial asset registration information, one asset change registration information, and the corresponding asset extension registration information, and the asset registration information is the second asset change registration information, the initial asset registration time of the initial asset registration information is used as the first graph node, where the first graph node corresponds to the enterprise name, credit code, legal person, and other attribute information of the initial asset registration information. The asset change registration time of the asset change registration information is used as the second graph node, where the second graph node corresponds to the transferor information, transferee information, and other attribute information of the asset change registration information. The asset extension registration time of the asset extension registration information is used as the third graph node, where the third graph node corresponds to the extension registration time, extension expiration time, and other attribute information of the asset extension registration information. Finally, the asset change registration time of the asset registration form is used as the fourth graph node, where the fourth graph node corresponds to the transferor information, transferee information, and other attribute information of the second asset change registration information.

[0114] The above technical solution involves extracting textual information from the acquired asset registration form; preprocessing each sentence within the textual information and inputting the resulting combined vector matrix of the sentence into a natural language processing (NLP) model to obtain the label type results for each character in the sentence; constructing asset registration information based on each label type result and querying historical registration information corresponding to historical registration numbers from the asset information database; and constructing an asset registration graph based on the historical registration information and the asset registration information. This asset registration graph, constructed based on historical and asset registration information, improves the convenience and accuracy of verifying asset registration information, thereby reducing the workload of manual verification and increasing the efficiency of asset registration information verification.

[0115] Based on the above embodiments, Figure 3 This is an implementation illustrated according to an exemplary embodiment. Figure 1 The flowchart of step S11 shows that the text information extraction of the asset registration form in step S11 includes the following steps.

[0116] In step S111, the target style template for the asset registration form is determined from the style templates in the style template library.

[0117] In step S112, by referring to the table style of the target style template, the key feature characters of the asset registration form and their corresponding value values ​​are determined, as well as the feature character coordinates of the key feature characters and the value coordinates of the value values.

[0118] In step S113, the asset registration form is reconstructed based on the key feature characters, feature character coordinates, value values, and value value coordinates, and text information is extracted from the key feature characters of the reconstructed asset registration form.

[0119] In one possible implementation, if the target style template for the asset register is not matched from the style template library, it can be achieved through methods such as... Figure 4 In this way, a corresponding style template is generated based on the asset registration form.

[0120] When the target style template for the asset registration form is matched from the style template library, the text information is structured by comparing it with the table style of the target style template to obtain the key feature characters of the asset registration form and the corresponding value of the key feature characters.

[0121] The coordinates of key feature characters in the asset registration form are determined based on the coordinates of template characters in the target style template. For example, the coordinates of the template characters of the name in the transferor information of the target style template are used as the coordinates of the key feature characters of the name in the transferor information of the asset registration form. The coordinates of the value of the name in the transferor information of the target style template are used as the coordinates of the value of the name in the transferor information of the asset registration form.

[0122] In this way, key feature characters and their values ​​can be located through style templates to complete the reconstruction of the asset registration form, converting the obtained asset registration form into a readable asset registration form, which facilitates the recognition of text information, thereby improving the accuracy of asset registration information verification, reducing the pressure of manual verification of asset registration information and improving the efficiency of asset registration information verification.

[0123] Based on the above embodiments, Figure 4 This is a flowchart illustrating a method for generating a style template according to an exemplary embodiment, such as... Figure 4 As shown, it includes the following steps.

[0124] In step S31, text extraction and table line extraction are performed on the asset registration form.

[0125] In step S32, based on the preset key feature characters, the template key feature characters are matched from the text extraction results, and the template character coordinates of the template key feature characters are determined.

[0126] In step S33, the coordinates of the template character value corresponding to the key feature character of the template are determined based on the template character coordinates.

[0127] In step S34, template table lines are constructed based on the table line extraction results, template character coordinates, and template character value coordinates to obtain the initial style template.

[0128] In step S35, the key feature characters of the template are filled into the initial style template using the template character coordinates as the target coordinates to obtain the style template.

[0129] The method for extracting text has been described in the aforementioned embodiments and will not be repeated here.

[0130] For example, based on the preset key feature character "transferor information - name", the template key feature character "transferor information" and its template character coordinates can be matched from the text information, as well as the template key feature character "name" and its template character coordinates. For instance, using the lower left corner of the asset registration form as the origin, the left side of the asset registration form as the y-axis, and the bottom edge as the x-axis, and selecting an appropriate unit length, a two-dimensional rectangular coordinate system is constructed. This allows the determination of the coordinates of each character, and consequently, the template character coordinates.

[0131] Furthermore, based on the coordinates of the template characters in the "Name", the coordinates of its template character value are determined. Typically, the value and the template key feature characters have a positional correspondence. For example, the value and the template key feature characters may be on the same line, meaning they have the same vertical coordinate; or the value and the template key feature characters may be on adjacent lines, with the template key feature characters on the previous line and the value on the next line, meaning they have the same horizontal coordinate.

[0132] In one implementation, an initial table can be preset. The initial table is adjusted based on the table line extraction results, template character coordinates, and template character value coordinates to obtain an initial style template. Specifically, the initial table can be preset according to different versions and types. For example, initial tables can be preset according to three types: initial asset registration form, asset change registration form, and asset extension registration form, and corresponding initial tables can be preset according to versions released by the central bank on different dates.

[0133] By adopting the above technical solution, the template key feature characters can be matched by preset key feature characters, and the template character coordinates of the template key feature characters can be determined. The value coordinates of the template character corresponding to the template key feature characters can be determined based on the template character coordinates, so as to obtain an initial style template. Then, the style template can be constructed in the template library, which facilitates the reconstruction of the asset registration form based on the style template.

[0134] Based on the above embodiments, Figure 5 This is a flowchart illustrating preprocessing text information according to an exemplary embodiment, such as... Figure 5 As shown, it includes the following steps.

[0135] In step S41, each sentence in the text information is split into characters and words to obtain the character splitting results and word splitting results.

[0136] In step S42, the word segmentation results are aligned using the length of the character segmentation result as the target length.

[0137] In step S43, the character splitting result indices are converted into character vectors of a preset dimension through the word vector layer, and a character vector matrix is constructed according to the length of the character splitting result and the dimension of the character vectors.

[0138] In step S44, the word splitting result indices after alignment processing are converted into word vectors of a preset dimension through the word vector layer, and a word vector matrix is constructed according to the length of the word matrix splitting result after alignment processing and the dimension of the word vectors.

[0139] In step S45, index combination is performed on the character vector matrix and the word vector matrix to obtain a combined vector matrix.

[0140] The following description takes the name "Jia Technology Co., Ltd." in the assignor information as an example. The sentence is subjected to character splitting, and the character splitting result is obtained as "Jia", "Ji", "Shu", "You", "Xian", "Gong" and "Si". The sentence is subjected to word splitting, and the word splitting result is obtained as "Jia", "Technology", "Limited" and "Company".

[0141] Obviously, the length of the character splitting result is 7, while the length of the word splitting result is 4. With the 7 characters of the character splitting result as the target length, alignment processing is performed on the word splitting result to expand the 4-length word splitting result to 7 in length.

[0142] Further, the character splitting result indices are converted into character vectors of a preset 300 dimensions through the word vector layer, so as to construct a 7*300 character vector matrix; and the word splitting result indices after alignment processing are converted into word vectors of a preset 300 dimensions through the word vector layer, so as to construct a 7*300 word vector matrix.

[0143] On the basis of the above embodiments, the natural language processing model includes a convolutional neural network model and a conditional random field model, Figure 6 is a flowchart of training a natural language processing model according to an exemplary embodiment, as Figure 6 shown, including the following steps.

[0144] In step S51, the combined vector matrix corresponding to each sentence in the sample text information is used as the training sample of the convolutional layer for convolution calculation, to obtain valid information of the sentence.

[0145] In step S52, pooling processing is performed on the valid information through a pooling layer, and calculation is performed on the pooled valid information through a fully connected layer, to obtain the label type score of each character in the sentence.

[0146] In step S53, a matrix is constructed according to the label type score of each character in the sentence, to obtain an emission probability matrix.

[0147] In step S54, the label type score is backpropagated, the gradient of the error on each weight in the convolutional network is calculated, and the parameters of the filters in the convolutional neural network model are adjusted using gradient descent.

[0148] In step S55, the transition probability matrix is ​​randomly initialized and backpropagated to obtain the emission probability matrix and the transition probability matrix, thereby correcting the model parameters of the conditional random field model.

[0149] Specifically, see Figure 7 The convolutional neural network shown includes: an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input to the input layer is a combined vector matrix corresponding to each sentence, with dimensions equal to the length of the sentence's character-level splitting result by a preset dimension. Then, the combined vector matrix is ​​passed forward to calculate the score for each category.

[0150] The forward propagation of a convolutional neural network (CNN) consists of two steps: convolution in the convolutional layer and pooling in the pooling layer. The convolutional layer is the most important layer in a CNN. The following formula shows the calculation method for the convolution operation, which involves multiplying two variables and then summing them. The convolutional layer can extract features from the textual information in the combined vector matrix using its internal convolutional kernels, outputting a matrix constructed based on the label type score of each character in each sentence. Each element of the convolutional kernel has a weight coefficient and a bias.

[0151]

[0152] Pooling layers primarily include max-pooling and average-pooling. Pooling layers reduce the dimensionality of the output, making network layer computation more efficient. Simultaneously, pooling operations can extract abstract information. In summary, convolutional layers, pooling layers, and fully connected layers are designed together to construct the basic structure of a convolutional neural network.

[0153] Furthermore, the convolutional neural network calculates the gradient of the error at each weight in the convolutional network by backpropagating the label type score and comparing it with the labeled label type score. It then uses gradient descent to adjust the parameters of the filters in the convolutional neural network model to minimize the output error.

[0154] It can be explained that, due to the advantages of Conditional Random Fields (CRFs) in solving model decoding problems, a conditional probability distribution model P(Y|X) can be used to represent a random field of output random variable Y given a set of output random variables X. Training a CRF model also includes forward propagation and backpropagation. The loss value is calculated based on the loss function, and the loss value is backpropagated to correct the model parameters. Typically, the loss value can include the emission probability matrix and the transition probability matrix.

[0155] The forward propagation process of a Conditional Random Field (CRF) model involves calculating the most probable label combination based on the emission probability matrix using the following formula. Define the matrix P∈R. N×L Let N be the emission probability matrix. Here, N represents the number of characters in each sentence, L represents the number of tag types, and P represents the emission probability matrix. ij This represents the probability that the i-th character in the sentence has a tag type result of j.

[0156]

[0157] Where S(X,Y) is the loss value, A yi,yi+1 Let be the transition probability matrix, which is obtained through random initialization.

[0158] For a given pair of sentences X, the predicted label sequence Y = {y1, y2, ..., y...} n The probability value of} is:

[0159]

[0160] The loss function for the label sequence during the training of the Conditional Random Field model is:

[0161]

[0162] Among them, Y A For all possible combinations of labels in the set, output the set of label type scores with the highest probability according to the following formula, as the most likely label type score output by the model.

[0163] y * =argmax Y S(X,Y)

[0164] Among them, y * Score based on label type.

[0165] Overall, the training model primarily uses a convolutional neural network (CNN) model. The CNN is used to extract n-gram features from textual information, performing both feature extraction and encoding. Simultaneously, a conditional random field (CRF) model is used as the decoding layer to provide label type results with constraints. Thus, the input to the natural language processing model is a combined vector matrix consisting of the indices of character and word segmentation results in each sentence, and the output is a label type score for each character.

[0166] During training, the data needs to be manually labeled beforehand. In this disclosure, five label types, {B-ORG, I-ORG, B-INVOICE, I-INVOICE, O}, are used to label each character. Specifically, the starting character of entities such as "company" is labeled with "B-ORG", and the middle character is labeled with "I-ORG". The starting character of entities such as "invoice" is labeled with "B-INVOICE", and the middle character is labeled with "I-INVOICE". All other characters are labeled with "O".

[0167] Based on the same inventive concept, this disclosure also provides an asset registration map construction device, which can implement all or part of the steps of the asset registration map construction method in the form of software, hardware or a combination of both. Figure 8 This is a block diagram illustrating an asset registry map construction apparatus 100 according to an exemplary embodiment, such as... Figure 8 As shown, the device 100 includes: an extraction module 110, an input module 120, a query module 130, and a construction module 140.

[0168] The extraction module 110 is configured to extract text information from the acquired asset registration form.

[0169] The input module 120 is configured to preprocess each sentence in the text information, and input the combined vector matrix of the preprocessed sentence into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model.

[0170] The query module 130 is configured to construct asset registration information based on the result of each of the aforementioned tag types, and to query historical registration information corresponding to the historical registration number in the asset registration information from the asset information database.

[0171] The building module 140 is configured to build an asset registration map based on the historical registration information and the asset registration information.

[0172] The aforementioned device, based on the asset registration map constructed from historical registration information and asset registration information, can improve the convenience of verifying asset registration information, and at the same time, improve the accuracy of verifying asset registration information, thereby reducing the pressure of manual verification of asset registration information and improving the efficiency of asset registration information verification.

[0173] Optionally, the extraction module 110 is configured to:

[0174] The target style template for the asset registration form is determined from the style template library;

[0175] By comparing the table style of the target style template, determine the key feature characters of the asset registration form and their corresponding value, as well as the feature character coordinates of the key feature characters and the value coordinates of the value.

[0176] The asset registration form is reconstructed based on the key feature characters, the coordinates of the feature characters, the value, and the coordinates of the value, and the text information is extracted from the key feature characters of the reconstructed asset registration form.

[0177] Optionally, the style template is obtained in the following way:

[0178] Extract text and table lines from the asset registration form;

[0179] Based on preset key feature characters, match template key feature characters from the text extraction results; and determine the template character coordinates of the template key feature characters;

[0180] The coordinates of the template character value corresponding to the key feature character of the template are determined based on the template character coordinates.

[0181] Based on the table line extraction results, the template character coordinates, and the template character value coordinates, template table lines are constructed to obtain the initial style template.

[0182] Using the coordinates of the template characters as the target coordinates, the key feature characters of the template are filled into the initial style template to obtain the style template.

[0183] Optionally, the input module 120 is configured to:

[0184] Each sentence in the text information is split into characters and words to obtain the character splitting results and word splitting results;

[0185] The word segmentation results are aligned using the length of the character segmentation results as the target length.

[0186] The word vector layer converts the character segmentation result index into character vectors of a preset dimension, and constructs a character vector matrix based on the length of the character segmentation result and the dimension of the character vectors; and,

[0187] The word vector layer converts the aligned word segmentation result index into word vectors of a preset dimension, and constructs a word vector matrix based on the length of the aligned word matrix segmentation result and the dimension of the word vectors.

[0188] The combined vector matrix is ​​obtained by indexing and combining the character vector matrix and the word vector matrix.

[0189] Optionally, the natural language processing model includes a convolutional neural network model and a conditional random field model, and the training of the natural language processing model includes:

[0190] The combined vector matrix corresponding to each sentence in the sample text information is used as the training sample for the convolutional layer to perform convolution calculation, thereby obtaining the effective information of the sentence.

[0191] The effective information is pooled through a pooling layer, and the pooled effective information is calculated through a fully connected layer to obtain the tag type score for each character in the sentence.

[0192] A matrix is ​​constructed based on the label type score of each character in the sentence to obtain the emission probability matrix;

[0193] The label type score is backpropagated, the gradient of the error at each weight in the convolutional network is calculated, and the parameters of the filters in the convolutional neural network model are adjusted using gradient descent; and,

[0194] The transition probability matrix is ​​obtained through random initialization, and the emission probability matrix and the transition probability matrix are backpropagated to correct the model parameters of the conditional random field model.

[0195] Optionally, the extraction module 110 is configured to:

[0196] Text information is extracted from the original code of the asset registration form; and,

[0197] Determine whether the extracted text information includes escape characters;

[0198] If the text information includes escape characters, the text information is extracted after the escape characters are made readable.

[0199] Optionally, the asset registration information includes any one of the following: initial asset registration information, asset change registration information, and asset extension registration information.

[0200] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0201] Furthermore, it is worth noting that the modules in the above embodiments can be independent devices or the same device in specific implementations. For example, the query module 130 and the construction module 140 can be the same module or two modules. This disclosure does not limit this.

[0202] According to this disclosure, a computer device is also provided, comprising:

[0203] processor;

[0204] Memory used to store processor-executable instructions;

[0205] The processor is configured as follows:

[0206] For the obtained asset registration form, extract the text information from the asset registration form;

[0207] For each sentence in the text information, the sentence is preprocessed, and the combined vector matrix of the preprocessed sentence is input into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model.

[0208] Based on the results of each of the aforementioned tag types, asset registration information is constructed, and historical registration information corresponding to the historical registration number in the asset registration information is queried from the asset information database;

[0209] An asset registration map is constructed based on the historical registration information and the asset registration information.

[0210] This disclosure also provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the steps of the asset registration map construction method provided in this disclosure.

[0211] Figure 9 This is a block diagram illustrating an apparatus 800 for constructing an asset registration map according to an exemplary embodiment. The apparatus 800 can be configured as a computer device to perform all or part of the steps of the aforementioned asset registration map construction method. For example, the apparatus 800 can be a computer, a digital broadcasting terminal, a message transceiver device, etc.

[0212] Reference Figure 9The device 800 may include one or more of the following components: a processing component 802, a memory 804, a power component 806, a multimedia component 808, an audio component 810, an input / output (I / O) interface 812, a sensor component 814, and a communication component 816.

[0213] Processing component 802 typically controls the overall operation of device 800, including operations related to display, data communication, receiving data packets including asset registers, and recording operations. Processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps of the asset register map construction method described above. Furthermore, processing component 802 may include one or more modules to facilitate interaction between processing component 802 and other components. For example, processing component 802 may include a multimedia module to facilitate interaction between multimedia component 808 and processing component 802.

[0214] Memory 804 is configured to store various types of data to support operation on device 800. Examples of this data include instructions for any application or method operating on device 800, such as decompression, invoking OCR, etc. Memory 804 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0215] The power supply component 806 provides power to the various components of the device 800. The power supply component 806 may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device 800.

[0216] Multimedia component 808 includes a screen that provides an output interface between the device 800 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may sense not only the boundaries of the touch or swipe action but also the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 808 includes a front-facing camera and / or a rear-facing camera for acquiring user facial data. When the device 800 is in an operating mode, such as a shooting mode or a video mode, the front-facing camera and / or the rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0217] Audio component 810 is configured to output and / or input audio signals. For example, audio component 810 includes a microphone (MIC) configured to receive external audio signals when device 800 is in an operating mode, such as voice recognition mode. The received audio signals may be further stored in memory 804 or transmitted via communication component 816. In some embodiments, audio component 810 also includes a speaker for outputting audio signals.

[0218] I / O interface 812 provides an interface between processing component 802 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0219] Sensor assembly 814 includes one or more sensors for providing status assessments of various aspects of device 800. For example, sensor assembly 814 can detect the on / off state of device 800, the relative positioning of components such as the display and keypad of device 800, changes in the position of device 800 or a component of device 800, the presence or absence of user contact with device 800, the orientation or acceleration / deceleration of device 800, and temperature changes of device 800. Sensor assembly 814 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 814 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications.

[0220] Communication component 816 is configured to facilitate wired or wireless communication between device 800 and other devices. Device 800 can access wireless networks based on communication standards, such as WiFi, 2G, or 3G, or combinations thereof. In one exemplary embodiment, communication component 816 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 816 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0221] In an exemplary embodiment, the apparatus 800 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described asset registration map construction method.

[0222] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 804 including instructions, which can be executed by a processor 820 of the device 800 to complete the asset registration map construction method described above. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0223] In another exemplary embodiment, a computer program product is also provided, comprising a computer program executable by a programmable device, the computer program having a code portion for performing the above-described asset registry map construction method when executed by the programmable device.

[0224] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of this disclosure. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0225] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for constructing an asset registration map, characterized in that, include: For the obtained asset registration form, extract the text information from the asset registration form; For each sentence in the text information, the sentence is preprocessed, and the combined vector matrix of the preprocessed sentence is input into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model. Based on the results of each of the aforementioned tag types, asset registration information is constructed, and historical registration information corresponding to the historical registration number in the asset registration information is queried from the asset information database; Constructing an asset registration graph based on the historical registration information and the asset registration information includes: using the registration time of the historical registration information and the asset registration information as graph nodes, and using single arrows as corresponding edges to represent the time flow of the registration information, to construct the asset registration graph.

2. The method according to claim 1, characterized in that, The extraction of text information from the asset registration form includes: The target style template for the asset registration form is determined from the style template library; By comparing the table style of the target style template, determine the key feature characters of the asset registration form and their corresponding values, as well as the feature character coordinates of the key feature characters and the value coordinates of the value. The asset registration form is reconstructed based on the key feature characters, the coordinates of the feature characters, the value, and the coordinates of the value, and the text information is extracted from the key feature characters of the reconstructed asset registration form.

3. The method according to claim 2, characterized in that, The style template was obtained in the following way: Extract text and table lines from the asset registration form; Based on preset key feature characters, match template key feature characters from the text extraction results; and determine the template character coordinates of the template key feature characters; The coordinates of the template character value corresponding to the key feature character of the template are determined based on the template character coordinates. Based on the table line extraction results, the template character coordinates, and the template character value coordinates, template table lines are constructed to obtain the initial style template. Using the coordinates of the template characters as the target coordinates, the key feature characters of the template are filled into the initial style template to obtain the style template.

4. The method according to claim 1, characterized in that, Preprocessing of each sentence in the text information includes: Each sentence in the text information is split into characters and words to obtain the character splitting results and word splitting results; The word segmentation results are aligned using the length of the character segmentation results as the target length. The word vector layer converts the character segmentation results into character vectors of a preset dimension, and constructs a character vector matrix based on the length of the character segmentation results and the dimension of the character vectors; and, The word vector layer converts the aligned word segmentation result index into word vectors of a preset dimension, and constructs a word vector matrix based on the length of the aligned word matrix segmentation result and the dimension of the word vectors. The combined vector matrix is ​​obtained by indexing and combining the character vector matrix and the word vector matrix.

5. The method according to claim 1, characterized in that, The natural language processing model includes a convolutional neural network model and a conditional random field model. The training of the natural language processing model includes: The combined vector matrix corresponding to each sentence in the sample text information is used as the training sample for the convolutional layer to perform convolution calculation, thereby obtaining the effective information of the sentence. The effective information is pooled through a pooling layer, and the pooled effective information is calculated through a fully connected layer to obtain the tag type score for each character in the sentence. A matrix is ​​constructed based on the label type score of each character in the sentence to obtain the emission probability matrix; The label type score is backpropagated, the gradient of the error at each weight in the convolutional network is calculated, and the parameters of the filters in the convolutional neural network model are adjusted using gradient descent; and, The transition probability matrix is ​​obtained through random initialization, and the emission probability matrix and the transition probability matrix are backpropagated to correct the model parameters of the conditional random field model.

6. The method according to claim 1, characterized in that, The extraction of text information from the asset registration form includes: Text information is extracted from the original code of the asset registration form; and it is determined whether the extracted text information includes escape characters. If the text information includes escape characters, the text information is extracted after the escape characters are made readable.

7. The method according to any one of claims 1-6, characterized in that, The asset registration information includes any one of the following: initial asset registration information, asset change registration information, and asset extension registration information.

8. An asset registration map construction device, characterized in that, include: The extraction module is configured to extract text information from the acquired asset registration form. The input module is configured to preprocess each sentence in the text information, and input the combined vector matrix of the preprocessed sentence into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model. The query module is configured to construct asset registration information based on the result of each of the aforementioned tag types, and to query historical registration information corresponding to the historical registration number in the asset registration information from the asset information database. The construction module is configured to construct an asset registration graph based on the historical registration information and the asset registration information, including: using the registration time of the historical registration information and the asset registration information as graph nodes, and using single arrows as corresponding edges to represent the time flow of the registration information, to construct the asset registration graph.

9. A computer device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured as follows: For the obtained asset registration form, extract the text information from the asset registration form; For each sentence in the text information, the sentence is preprocessed, and the combined vector matrix of the preprocessed sentence is input into the natural language processing model to obtain the label type result of each character in the sentence output by the natural language processing model. Based on the results of each of the aforementioned tag types, asset registration information is constructed, and historical registration information corresponding to the historical registration number in the asset registration information is queried from the asset information database; Constructing an asset registration graph based on the historical registration information and the asset registration information includes: using the registration time of the historical registration information and the asset registration information as graph nodes, and using single arrows as corresponding edges to represent the time flow of the registration information, to construct the asset registration graph.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When executed by a processor, the program instructions implement the steps of the method described in any one of claims 1 to 7.

Images