Case filing pre-auditing system of intelligent governance platform
The case text is preprocessed and iteratively trained through the TextMap algorithm and recurrent memory network to build a comprehensive data weight matrix, which solves the problems of slow keyword refining speed and high typos replacement error rate in the existing technology, and improves the filing of the smart governance platform Pre-examination efficiency.
Patent Information
- Application Number
- CN202411105068.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-04-30
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-13
Smart Images

Figure CN119027052A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing technology, and in particular to a case filing pre-examination system for a smart governance platform. Background Art
[0002] With the development of society and the advancement of information technology, grassroots governance is facing more and more challenges. How to improve the efficiency of grassroots governance and achieve refined and intelligent management has become an urgent problem to be solved.
[0003] With the rapid development of modern information technologies such as artificial intelligence, big data, and cloud computing, the government has begun to explore new modes of information technology application. Reasonable prediction and judgment of whether grassroots case information meets the conditions for filing a case is crucial to improving the scientificity and accuracy of government administrative decision-making.
[0004] Patent No. CN2022106050526 discloses a smart grassroots governance system, which involves the field of computer technology, including: application layer, business layer, platform layer and support layer; the application layer includes: problem reflection APP, problem handling APP and analysis and decision-making APP; the platform layer includes: information service platform, problem handling platform and analysis and decision-making platform; the support layer includes: data management module and permission management module; the data management module includes: workflow management module, content management module and data analysis module; the permission management module provides permission support services for the platform layer. The above application can avoid the constraints of space and time on problem collection and processing, thereby improving the efficiency of problem collection and processing, and facilitating the tracking and supervision of the process of problem handling.
[0005] Patent No. CN2023107391062 discloses a grassroots smart governance management system and method, including an application terminal layer, a business application layer, a business support layer, a data resource layer, an infrastructure layer and a security layer; the application terminal layer is used to open a unified service port and information display; the business application layer is used to provide access and scope of grassroots governance applications; the business support layer is used to provide the technical capabilities involved in each process in the grassroots governance application; the data resource layer is used to collect, store, integrate and call various types of data; the infrastructure layer is a government cloud network; the security layer is used to ensure the data security and stable operation of the system. The above invention has the advantages of high service efficiency, high service capability and good service quality.
[0006] However, the above patents have great deficiencies in data processing, which leads to slow keyword extraction of case texts, high error rate in replacing typos, and low efficiency in pre-trial filing of case texts. Summary of the invention
[0007] The purpose of the present invention is to provide a case filing pre-examination system for a smart governance platform, which can complete the keyword extraction of case texts through the TextMap algorithm, and iteratively train the case texts through a learning and training layer to construct a comprehensive data weight matrix; thereby improving the keyword extraction speed of case texts and improving the case filing pre-examination efficiency of case texts.
[0008] The present invention utilizes the following technical solutions: A case filing pre-examination system for a smart governance platform, comprising a processing conversion layer, a matching filtering layer, a learning training layer, and an identification pre-examination layer; The processing conversion layer is used to pre-process the case text, including text cleaning, typo replacement and information extraction; The matching filter layer is used to perform category matching, length filtering and vector compilation on the pre-processed case text; The learning and training layer is used to iteratively train the case texts that have completed matching and filtering to build a comprehensive data weight matrix; The recognition pre-screening layer is used to mask the messy information in the case text and then import it into the comprehensive data weight matrix for text recognition and screening.
[0009] Preferably, the processing conversion layer includes a text cleaning unit, a typo replacement unit and an information extraction unit; the text cleaning unit uses regular expressions to remove special characters, punctuation marks and blank chapters in the case text, and eliminates duplicate case texts based on the location of the event; the typo replacement unit uses a preset recurrent memory network to convert similar-looking characters and / or similar-sounding characters in the case text into correct fonts; the information extraction unit uses the TextMap algorithm to extract keywords from the case text with typos replaced, thereby removing the time of the event, the location of the event and explanatory sentences.
[0010] Preferably, the workflow of the TextMap algorithm is: a: Convert the case text into an undirected graph, and use the words as nodes and the sentences as edges of the undirected graph; b: Calculate the weights and information entropy of sentence edges and word nodes according to the frequency of occurrence of sentences and words in the case text; c: Abstract the obtained weight values into a directed graph, and construct the transition probability to randomly walk in the directed graph according to the information entropy of word nodes and sentence edges to generate a global sequence; d: hierarchically encode the global sequence and perform clustering local optimization on the hierarchical encoding to generate local optimization results; e: Extract keywords from the case text based on the local optimization results.
[0011] Preferably, the recurrent memory network first sorts similar characters to form sequence labels; at the same time, the similar characters, similar sounding characters and sequence labels are respectively input into three network branches for feature extraction; wherein, the similar characters use the first network branch to obtain the depth feature values of the similar characters; the first network branch includes 1 bidirectional LSTM layer, 4 hidden state layers, 2 fully connected layers and PRELU activation function; the similar sounding characters use the second network branch to obtain the depth feature values of the similar sounding characters; the second network branch includes 2 residual blocks, 5 maximum pooling layers, 2 fully connected layers and LeakyReLU activation function; the sequence labels use the third network branch to extract the depth feature values; the third network branch includes 3 Inception layers, 2 auxiliary classification layers and RELU activation function; then respectively The deep feature values output by the first network branch and the third network branch are concatenated and fused using the feature fusion layer, and then the fused deep feature values are input into the first branch of the recognition converter to obtain the form matrix; the first branch includes 1 average pooling layer, 2 fully connected layers, random inactivation function and ELU activation function to obtain the form matrix; at the same time, the deep feature values output by the second network branch and the third network branch are concatenated and fused using the feature fusion layer, and then the fused deep feature values are input into the second branch of the recognition converter to obtain the phonetic matrix; the second branch includes 1 LSTM layer, 2 hidden state layers, 1 fully connected layer and MISH activation function to obtain the phonetic matrix; finally, the conversion process of similar-sounding characters and similar-looking characters is completed by concatenating the phonetic matrix and the form matrix.
[0012] Preferably, the matching filtering layer includes a category matching unit, a text filtering unit, a labeling division unit and a vector compilation unit; the category matching unit uses a text matching algorithm to match the keywords extracted from the case text according to the functions of each department; the text filtering unit uses a length filtering algorithm to filter and screen the matched case text according to the character length within 200; the labeling division unit uses LabeImg software to label the case text according to the functions of each department and the keywords extracted from the case text that has completed filtering and screening, so as to form a training set and a test set; the vector compilation unit uses a global bag-of-words algorithm to compile the training set and the test set into vector matrices respectively.
[0013] Preferably, the learning training layer first inputs the vector matrix of the training set into the feature extractor for channel expansion to divide the vector matrix of the training set into a functional matrix and a keyword matrix; the feature extractor includes 4 3X3 convolutional layers with a step size of 2 and 1 batch standard layer; then the functional matrix and the keyword matrix are respectively input into two branches of the backbone network for multiple iterative training to obtain a comprehensive data weight matrix; the first branch includes 1 CBL block, 2 residual blocks, 3 segmentation layers, 3 Inception blocks and a Silu activation function; the CBL block includes 1 3X3 convolutional layer with a step size of 1, 1 batch standard layer and a LeakyRelu activation function; the second branch includes 2 Transformer encoding blocks, 1 batch standard layer, a fusion attention mechanism and a MISH function; finally, the input network output of the test set is used to infer and verify the comprehensive data weight matrix, thereby optimizing the comprehensive data weight matrix; the network output includes 3 normalization layers, 2 fully connected layers and a Softmax activation function.
[0014] Preferably, the identification pre-examination layer includes an information masking unit; the information masking unit uses a masking language network to replace the time of event occurrence, the location of event occurrence and the explanatory language in the case text.
[0015] Preferably, the workflow of the masked language network is: S1: Use a token encoder to segment sentences of the case text into individual words, and encode the words to form a case text code; S2: query the coded words according to the pre-set taboo coding table; the taboo coding table includes the time of the event, the place of the event and the explanatory language; S3: Use a masker to perform bilinear interpolation masking and replacement on the word codes that meet the taboo code.
[0016] Preferably, the identification pre-examination layer also includes a case filing pre-examination unit; the case filing pre-examination unit uses a causal language network to input the masked case text encoding into a comprehensive data weight matrix for classification and filing, and outputs the case filing pre-examination result.
[0017] Preferably, the workflow of the causal language network is: A: The masked case text encoding is enhanced through the spatial channel attention mechanism; B: Encode the case text after feature enhancement using 4 fully connected layers and import the comprehensive data weight matrix for case classification; C: The classified case texts will be divided and sorted into pre-examination levels according to the case handling progress of each department.
[0018] The present invention extracts keywords from case texts through a TextMap algorithm, converts typos in the form of similar-looking and / or similar-sounding characters in case texts into correct fonts through a recurrent memory network; it iteratively trains case texts that have completed matching filtering through a learning training layer to construct a comprehensive data weight matrix; thereby increasing the keyword extraction speed of case texts, reducing the error rate of typos replacement, and improving the case filing pre-examination efficiency of case texts. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the drawings required for use in the embodiments or related technical descriptions are briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0020] Figure 1 Schematic diagram of the case pre-examination system of the smart governance platform. DETAILED DESCRIPTION
[0021] The present invention is described in detail below with reference to the accompanying drawings and embodiments: like Figure 1 As shown, the case filing pre-examination system of a smart governance platform described in the present invention includes a processing conversion layer, a matching filtering layer, a learning and training layer, and an identification pre-examination layer; The processing conversion layer is used to pre-process the case text, including text cleaning, typo replacement and information extraction; In the present invention, the processing conversion layer includes a text cleaning unit, a typo replacement unit and an information extraction unit; the text cleaning unit uses regular expressions to remove special characters, punctuation marks and blank chapters in the case text, and removes duplicate case texts according to the location of the event; the typo replacement unit uses a preset recurrent memory network to convert typos in the case text in the form of similar characters and / or similar-sounding characters into correct fonts; the information extraction unit uses a TextMap algorithm to extract keywords from the case text after the typo replacement is completed, thereby removing the time of the event, the location of the event and the explanatory sentences; In this embodiment, the text cleaning unit mainly includes three steps: removing special symbols, removing redundant spaces, and converting traditional Chinese to simplified Chinese. Removing special symbols and redundant spaces is to make the feature representation of short text focus only on the features and semantics of the vocabulary of the short text itself as much as possible, and reduce the influence of other symbols on the classification accuracy. It also includes traditional Chinese to simplified Chinese conversion, that is, converting traditional Chinese to simplified Chinese is to facilitate the subsequent text vectorization representation, because the text vectorization representation method used in this article is to call an external word vector model. If the vocabulary used in the text is not in the vocabulary, the current vocabulary will use the initialized vector representation method, changing the semantics of the vocabulary itself. Most traditional Chinese texts are relatively complex and usually exceed the scope of the vocabulary. Converting them to simplified Chinese will not change their own semantics, but also facilitate vectorization representation. Therefore, the step of converting traditional Chinese to simplified Chinese is very necessary. In this embodiment, the text cleaning unit and the information extraction unit are combined with the following characteristics of the case text: 1. Specific beginning and ending; 2. Non-keywords; Among them, non-keywords mainly include the time of the event, the place of the event and explanatory sentences. Such non-keywords are generally fixed rules, so all non-keywords can be directly deleted in the whole text through regular expressions; because these words usually have no special semantics, and will increase the difficulty of training, and may even affect the experimental results; 3. The text template used is relatively short; in order to enable the case to be classified from the keywords as quickly as possible, the staff will extract important information and shorten the long case description to less than 150 words; 4. Stop words; such as "period" and "following" appear too many times. Although they cannot provide any semantic information in the classification task, they occupy an important position in the sentence distribution and affect the case classification task; a stop word list is constructed according to the case data in this article, and the stop words are directly filtered out through the stop word list; The matching filter layer is used to perform category matching, length filtering and vector compilation on the pre-processed case text; In the present invention, the matching filtering layer includes a category matching unit, a text filtering unit, a labeling division unit and a vector compilation unit; the category matching unit uses a text matching algorithm to match the keywords extracted from the case text according to the functions of each department; the text filtering unit uses a length filtering algorithm to filter and screen the matched case text according to the character length within 200; the labeling division unit uses LabeImg software to label the case text according to the functions of each department and the keywords extracted from the case text that has completed filtering and screening, so as to form a training set and a test set; the vector compilation unit uses a global bag-of-words algorithm to compile the training set and the test set into vector matrices respectively; In this embodiment, the main responsibility of the text filtering unit is to filter the text length; the text length filtering is because the main research object is short text, and short text is usually a text form of no more than 200 characters; if the text is too long, it will exceed the scope of this article, so such text will be filtered; The learning and training layer is used to iteratively train the case texts that have completed matching and filtering to build a comprehensive data weight matrix; In the present invention, the learning training layer first inputs the vector matrix of the training set into the feature extractor for channel expansion to divide the vector matrix of the training set into a functional matrix and a keyword matrix; the feature extractor includes 4 3X3 convolutional layers with a step size of 2 and 1 batch standard layer; then the functional matrix and the keyword matrix are respectively input into two branches of the backbone network device for multiple iterative training to obtain a comprehensive data weight matrix; the first branch includes 1 CBL block, 2 residual blocks, 3 segmentation layers, 3 Inception blocks and Silu activation function; the CBL block includes 1 3X3 convolutional layer with a step size of 1, 1 batch standard layer and LeakyRelu activation function; the second branch includes 2 Transformer encoding blocks, 1 batch standard layer, fusion attention mechanism and MISH function; finally, the input network output device of the test set performs reasoning verification on the comprehensive data weight matrix, thereby optimizing the comprehensive data weight matrix; the network output device includes 3 normalization layers, 2 fully connected layers and Softmax activation function; The recognition pre-screening layer is used to mask the messy information in the case text and then import it into the comprehensive data weight matrix for text recognition and screening.
[0022] In the present invention, the identification pre-examination layer includes an information masking unit and a case filing pre-examination unit; the information masking unit uses a masking language network to replace the event time, event location and explanatory language in the case text; the case filing pre-examination unit uses a causal language network to input the masked case text encoding into a comprehensive data weight matrix for classification and filing, and outputs the case filing pre-examination result.
[0023] In this embodiment, the workflow of the mask language network is: S1: Use a token encoder to segment sentences of the case text into individual words, and encode the words to form a case text code; S2: query the coded words according to the pre-set taboo coding table; the taboo coding table includes the time of the event, the place of the event and the explanatory language; S3: Use a masker to perform bilinear interpolation masking and replacement on the word codes that meet the taboo code; In this embodiment, the workflow of the causal language network is: A: The masked case text encoding is enhanced through the spatial channel attention mechanism; B: Encode the case text after feature enhancement using 4 fully connected layers and import the comprehensive data weight matrix for case classification; C: Classify and sort the classified case documents according to the progress of each department in pre-trial level; In the present invention, regular expressions, text matching algorithms, length filtering algorithms, Labelmg software and global bag-of-words algorithms are all commonly used technologies in the art and will not be described in detail here.
[0024] Furthermore, the workflow of the TextMap algorithm is: a: Convert the case text into an undirected graph, and use the words as nodes and the sentences as edges of the undirected graph; b: Calculate the weights and information entropy of sentence edges and word nodes according to the frequency of occurrence of sentences and words in the case text: (1) in, Represents the sentence edge weight value, represents the statement edge number, Indicates The edge weight of each statement, represents the damping coefficient, represents all sentence edge sequences in the case text, Indicates the current statement edge number, represents the sequence of all cohesive sentence edges; (2) in, Represents the weight value of the word node, Indicates the word node number, Indicates The weight value of the word node, represents the damping coefficient, Represents all word node sequences in the case text, Indicates the current word node number, Represents the sequence of all cohesive word nodes; (3) in, represents the information entropy of each sentence edge, represents the Hough function, represents the probability set of each statement edge, Indicates the number of statements, represents the sentence edge probability, Represents the edge probability of the current sentence; (4) in, Represents the information entropy of each word node, represents the probability set of each word node, Indicates the number of word nodes, represents the probability of word node, Represents the current word node probability; c: Abstract the obtained weight values into a directed graph, and construct the transition probability to randomly walk in the directed graph according to the information entropy of word nodes and sentence edges to generate a global sequence: (5) in, represents the position in the directed graph, Indicates the current word node or sentence edge, represents the transition probability, Indicates the shift symbol, Indicates a shifted word node or sentence edge, Indicates the position of the current word node or sentence edge in the directed graph. Indicates the position of the shifted word node or sentence edge in the directed graph. Indicates the position of the current word node or sentence edge after shifting in the directed graph; (6) in, represents a global sequence, represents the normalization function, represents the transfer adjustment factor; d: hierarchically encode the global sequence and perform clustering local optimization on the hierarchical encoding to generate local optimization results; e: Extract keywords from the case text based on the local optimization results.
[0025] Furthermore, the recurrent memory network first sorts similar characters to form sequence labels; at the same time, the similar characters, similar sounding characters and sequence labels are respectively input into three network branches for feature extraction; among them, the similar characters use the first network branch to obtain the deep feature values of the similar characters; the first network branch includes 1 bidirectional LSTM layer, 4 hidden state layers, 2 fully connected layers and PRELU activation function; the similar sounding characters use the second network branch to obtain the deep feature values of the similar sounding characters; the second network branch includes 2 residual blocks, 5 maximum pooling layers, 2 fully connected layers and LeakyReLU activation function; the sequence labels use the third network branch to extract the deep feature values; the third network branch includes 3 Inception layers, 2 auxiliary classification layers and RELU activation function; then respectively The deep feature values output by the first network branch and the third network branch are concatenated and fused using the feature fusion layer, and then the fused deep feature values are input into the first branch of the recognition converter to obtain the form matrix; the first branch includes 1 average pooling layer, 2 fully connected layers, random inactivation function and ELU activation function to obtain the form matrix; at the same time, the deep feature values output by the second network branch and the third network branch are concatenated and fused using the feature fusion layer, and then the fused deep feature values are input into the second branch of the recognition converter to obtain the phonetic matrix; the second branch includes 1 LSTM layer, 2 hidden state layers, 1 fully connected layer and MISH activation function to obtain the phonetic matrix; finally, the conversion process of similar-sounding characters and similar-looking characters is completed by concatenating the phonetic matrix and the form matrix; In the present invention, the working principle of the recurrent memory network is as follows: first, the inputted similar characters and similar-sounding characters are arranged in a certain order to form a sequence label as the input for subsequent processing; the similar-sounding characters, similar-sounding characters and sequence labels are respectively input into three network branches for feature extraction. Among them, the first network branch uses multiple layers such as bidirectional LSTM, hidden state layer, and fully connected layer to extract deep feature values; the second network branch uses residual blocks, maximum pooling layers, and fully connected layers to obtain deep feature values of similar-sounding characters; the third network branch uses Inception layers, auxiliary classification layers, and other methods to extract deep feature values of sequence labels; Then, the deep feature values output by the first network branch and the third network branch are spliced through the feature fusion layer to obtain a form matrix; the deep feature values output by the second network branch and the third network branch are spliced through the feature fusion layer to obtain a phonetic matrix; and at the same time, the form matrix and the phonetic matrix are updated and combined in combination with the text error correction rules; Finally, certain algorithms, such as average pooling layer and fully connected layer, are used to process the deep feature values to obtain the final conversion results corresponding to the similar-sounding characters and similar-looking characters. In this embodiment, the working principle of the text error correction rule is as follows: the first step is error detection: the Jieba Chinese word segmenter is used to segment words and characters. Since the sentence contains typos, the segmentation results often have segmentation errors. Therefore, errors are detected from both the character granularity and the word granularity, and the suspected error results of these two granularities are integrated to form a candidate set of suspected error positions; among them, character granularity: if the likelihood probability value of a certain character detected by the language model perplexity (ppl) is lower than the average value of the sentence text, then the probability of judging that the character is a suspected typo is high; word granularity: after segmentation, the word that is not in the dictionary is likely to be a suspected typo; the second step is error correction; by traversing all suspected error positions, and using a sound-similar dictionary and a shape-similar dictionary to replace the word in the wrong position, and then calculating the sentence perplexity through the language model, all candidate set results are compared and sorted to obtain the optimal corrected word.
[0026] In this embodiment, when the learning and training layer is iteratively trained, the optimal comprehensive data weight matrix is obtained by dynamically correcting the learning rate: (7) in, represents the training rate, Indicates the training time, Indicates the training rate at the current moment, represents the first network momentum coefficient, represents the weight gradient, represents the second network momentum coefficient; (8) in, represents the adaptive parameter, represents the adaptive parameters at the current moment, represents the average coefficient of the first network momentum and the second network momentum, represents the adaptive parameters at the previous moment; (9) in, represents the learning rate, represents the learning rate at the current moment, represents the attenuation coefficient, represents the initial learning rate; (10) in, represents the comprehensive data weight matrix, Represents the comprehensive data weight matrix at the current moment; represents the updated comprehensive data weight matrix, represents a non-zero positive number, represents the regularization coefficient.
[0027] Example: When using the case pre-trial system to process a large number of case texts, the text cleaning unit of the processing conversion layer uses regular expressions to remove special characters, punctuation marks and blank chapters in the case text, and eliminates duplicate case texts according to the location of the incident; the typo replacement unit uses a preset recurrent memory network to first sort similar characters to form sequence labels; at the same time, the similar characters, similar sounding characters and sequence labels are respectively input into three network branches for feature extraction; among them, the similar characters use the first network branch to obtain the deep feature values of the similar characters; the first network branch includes 1 bidirectional LSTM layer, 4 hidden state layers, 2 fully connected layers and PRELU activation function; the similar sounding characters use the second network branch to obtain the deep feature values of the similar sounding characters; the second network branch includes 2 residual blocks, 5 maximum pooling layers, 2 fully connected layers and LeakyReLU activation function. Active function; the sequence label uses the third network branch to extract deep feature values; the third network branch includes 3 Inception layers, 2 auxiliary classification layers and RELU activation function; then the deep feature values output by the first network branch and the third network branch are spliced by the feature fusion layer, and then the fused deep feature values are sequentially input into 1 average pooling layer, 2 fully connected layers, random inactivation function and ELU activation function to obtain the form matrix; at the same time, the deep feature values output by the second network branch and the third network branch are spliced by the feature fusion layer, and then the fused deep feature values are sequentially input into 1 LSTM layer, 2 hidden state layers, 1 fully connected layer and MISH activation function to obtain the phonetic matrix; finally, the conversion process of similar-sounding characters and similar-shaped characters is completed by splicing the phonetic matrix and the form matrix; The information extraction unit uses the TextMap algorithm to convert the case text in which the typos have been replaced into an undirected graph, and uses words as nodes of the undirected graph, and uses sentences as edges of the undirected graph; according to the frequency of occurrence of sentences and words in the case text, the weight values and information entropy of sentence edges and word nodes are calculated respectively; the obtained weight values are abstracted into a directed graph, and the transition probability is constructed to randomly walk in the directed graph according to the information entropy of word nodes and sentence edges respectively, so as to generate a global sequence; the global sequence is hierarchically encoded, and the hierarchical encoding is clustered and locally optimized to generate a local optimization result; the keywords in the case text are extracted according to the local optimization result, so as to remove the time of the event, the place of the event and the explanatory sentences; The category matching unit of the matching filter layer uses the text matching algorithm to match the keywords extracted from the case text according to the functions of each department; the text filtering unit uses the length filtering algorithm to filter and screen the matched case text within 200 characters; the annotation division unit uses the LabeImg software to annotate the case text according to the functions of each department and the keywords extracted from the case text that has completed the filtering and screening, so as to form a training set and a test set; the vector compilation unit uses the global bag-of-words algorithm to compile the training set and the test set into vector matrices respectively; Then the learning training layer first inputs the vector matrix of the training set into the feature extractor for channel expansion to divide the vector matrix of the training set into a functional matrix and a keyword matrix; the feature extractor includes 4 3X3 convolutional layers with a step size of 2 and 1 batch standard layer; then the functional matrix and the keyword matrix are respectively input into the two branches of the backbone network for multiple iterative training to obtain the comprehensive data weight matrix; the first branch includes 1 CBL block, 2 residual blocks, 3 segmentation layers, 3 Inception blocks and Silu activation function; the CBL block includes 1 3X3 convolutional layer with a step size of 1, 1 batch standard layer and LeakyRelu activation function; the second branch includes 2 Transformer encoding blocks, 1 batch standard layer, fusion attention mechanism and MISH function; finally, the input network output of the test set is used to infer and verify the comprehensive data weight matrix, thereby optimizing the comprehensive data weight matrix; the network output includes 3 normalization layers, 2 fully connected layers and Softmax activation function; Finally, by identifying the information masking unit of the pre-trial layer and using the masked language network, the case text sentences are segmented into individual words through the token encoder, and the words are encoded to form the case text encoding; the encoded words are queried for taboo coding according to the pre-set taboo coding table; the taboo coding table includes the time of the event, the place of the event and the explanatory language; the word encoding that meets the taboo coding is masked and replaced by bilinear interpolation using a masker; the case text with masking is encoded by the causal language network through the case filing pre-trial unit, and the feature is enhanced through the spatial channel attention mechanism; then the case text encoding with feature enhancement is imported into the comprehensive data weight matrix for case classification using 4 fully connected layers; then the classified case text is divided and sorted for pre-trial level according to the case handling progress of each department to complete the case text classification and filing pre-trial task.
Claims
1. A case filing pre-examination system for a smart governance platform, characterized by: Includes processing conversion layer, matching filter layer, learning training layer and recognition pre-screening layer; The processing conversion layer is used to pre-process the case text, including text cleaning, typo replacement and information extraction; The matching filter layer is used to perform category matching, length filtering and vector compilation on the pre-processed case text; The learning and training layer is used to iteratively train the case texts that have completed matching and filtering to build a comprehensive data weight matrix; The recognition pre-screening layer is used to mask the messy information in the case text and then import it into the comprehensive data weight matrix for text recognition and screening.
2. The case filing pre-examination system of the smart governance platform according to claim 1 is characterized by: The processing conversion layer includes a text cleaning unit, a typo replacement unit and an information extraction unit; the text cleaning unit uses regular expressions to remove special characters, punctuation marks and blank chapters in the case text, and eliminates duplicate case texts according to the location of the event; the typo replacement unit uses a preset recurrent memory network to replace similar characters and / or similar-sounding characters in the case text with correct fonts; the information extraction unit uses a TextMap algorithm to extract keywords from the case text after the typo replacement is completed, thereby removing the time of the event, the location of the event and explanatory sentences.
3. The case filing pre-examination system of the smart governance platform according to claim 2 is characterized by: The workflow of the TextMap algorithm is: a: Convert the case text into an undirected graph, and use the words as nodes and the sentences as edges of the undirected graph; b: Calculate the weights and information entropy of sentence edges and word nodes according to the frequency of occurrence of sentences and words in the case text; c: Abstract the obtained weight values into a directed graph, and construct the transition probability to randomly walk in the directed graph according to the information entropy of word nodes and sentence edges to generate a global sequence; d: hierarchically encode the global sequence and perform clustering local optimization on the hierarchical encoding to generate local optimization results; e: Extract keywords from the case text based on the local optimization results.
4. The case filing pre-examination system of the smart governance platform according to claim 2 is characterized by: The recurrent memory network first sorts similar-looking characters and similar-sounding characters to form a combined sequence label; at the same time, the similar-looking characters, similar-sounding characters and sequence labels are respectively input into three network branches for feature extraction; wherein, the similar-looking characters use the first network branch to obtain the deep feature values of the similar-looking characters; the similar-sounding characters use the second network branch to obtain the deep feature values of the similar-sounding characters; the sequence labels use the third network branch to extract the deep feature values; then, the deep feature values output by the first network branch and the third network branch are respectively spliced and fused by the feature fusion layer, and then the fused deep feature values are input into the first branch of the recognition converter to obtain the form matrix; at the same time, the deep feature values output by the second network branch and the third network branch are spliced and fused by the feature fusion layer, and then the fused deep feature values are input into the second branch of the recognition converter to obtain the phonetic matrix; finally, the conversion process of phonetic-sounding characters and similar-looking characters is completed by splicing the phonetic matrix and the form matrix, and the phonetic matrix and the form matrix are regularly updated in combination with the text error correction rules.
5. The case filing pre-examination system of the smart governance platform according to claim 1 is characterized by: The matching filtering layer includes a category matching unit, a text filtering unit, a labeling division unit and a vector compilation unit; the category matching unit uses a text matching algorithm to match the keywords extracted from the case text according to the functions of each department; the text filtering unit uses a length filtering algorithm to filter and screen the matched case text according to the character length within 200; the labeling division unit uses LabeImg software to label the case text according to the functions of each department and the keywords extracted from the case text that has completed filtering and screening, so as to form a training set and a test set; the vector compilation unit uses a global bag-of-words algorithm to compile the training set and the test set into vector matrices respectively.
6. The case filing pre-examination system of the smart governance platform according to claim 1 is characterized by: The learning and training layer first inputs the vector matrix of the training set into the feature extractor for channel expansion to divide the vector matrix of the training set into a function matrix and a keyword matrix; Then the function matrix and keyword matrix are respectively input into the two branches of the backbone network for multiple iterative training to obtain the comprehensive data weight matrix; finally, the input network output of the test set is used to infer and verify the comprehensive data weight matrix, thereby optimizing the comprehensive data weight matrix.
7. The case filing pre-examination system of the smart governance platform according to claim 1 is characterized by: The identification pre-examination layer includes an information masking unit; the information masking unit uses a masking language network to replace the event time, event location and explanatory language in the case text.
8. The case filing pre-examination system of the smart governance platform according to claim 7 is characterized by: The workflow of the masked language network is: S1: Use a token encoder to segment sentences of the case text into individual words, and encode the words to form a case text code; S2: query the coded words according to the pre-set taboo coding table; the taboo coding table includes the time of the event, the place of the event and the explanatory language; S3: Use a masker to perform bilinear interpolation masking and replacement on the word codes that meet the taboo code.
9. The case filing pre-examination system of the smart governance platform according to claim 1 is characterized by: The identification pre-examination layer also includes a case filing pre-examination unit; the case filing pre-examination unit uses a causal language network to input the masked case text encoding into a comprehensive data weight matrix for classification and filing, and outputs the case filing pre-examination result.
10. The case filing pre-examination system of the smart governance platform according to claim 7 or 9 is characterized in that: The workflow of the causal language network is: A: The masked case text encoding is enhanced through the spatial channel attention mechanism; B: Encode the case text after feature enhancement using 4 fully connected layers and import the comprehensive data weight matrix for case classification; C: The classified case texts will be divided and sorted into pre-examination levels according to the case handling progress of each department.
Citation Information
Patent Citations
Short text classification method based on deep learning network
CN114357165A
Entity identification method and system, computer readable storage medium and terminal
CN114818681A
Legal document classification method based on prompt learning fusion keyword
CN115687626A
Semantic sentiment analysis method fusing in-depth features and time sequence models
US11194972B1
Keyword generation method and apparatus, and electronic device and computer storage medium
WO2022134759A1
Cited By
Auditing question and answer large model optimization method based on prompt fine tuning
CN120031019A
