A case-filing pre-trial system for an intelligent governance platform

Abstract

The present invention discloses a case filing pre-trial system for an intelligent governance platform, which relates to the technical field of data processing, and solves the problems of slow keyword extraction speed of case texts, high error rate of typo replacement, and low efficiency of case filing pre-trial for case texts; the case filing pre-trial system includes a processing and conversion layer, a matching and filtering layer, a learning and training layer, and an identification and pre-trial layer; the keywords of case texts are extracted through the TextMap algorithm, and typos in the form of similar-shaped characters and / or homophonic characters in the case texts are converted into correct fonts through a recurrent memory network; the case texts that have completed matching and filtering are iteratively trained through the learning and training layer to construct a comprehensive data weight matrix; the keyword extraction speed of case texts is improved, the error rate of typo replacement is reduced, and the efficiency of case filing pre-trial for case texts is increased.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a case pre-trial system for an intelligent governance platform. Background Art

[0002] With the development of society and the progress of information technology, grass-roots governance is facing more and more challenges. How to improve the efficiency of grass-roots 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 application models of information technologies. Reasonably predicting and judging whether grass-roots case-related information meets the case-filing conditions is crucial for improving the scientificity and accuracy of government administrative decisions.

[0004] Patent No. CN2022106050526 discloses an intelligent grass-roots governance system, which relates to the field of computer technology and includes: an application layer, a business layer, a platform layer, and a support layer; the application layer includes: a problem reporting APP, a problem handling APP, and an analysis and decision-making APP; the platform layer includes: an information service platform, a problem handling platform, and an analysis and decision-making platform; the support layer includes: a data management module and a permission management module; the data management module includes: a workflow management module, a content management module, and a 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 handling problems.

[0005] Patent No. CN2023107391062 discloses a grass-roots intelligent governance management system and method, which includes an application terminal layer, a business application layer, a business support layer, a data resource layer, an infrastructure layer, and a security guarantee layer; the application terminal layer is used to open a unified service port for handling affairs and information display; the business application layer is used to provide access and scope for grass-roots governance applications; the business support layer is used to provide technical capabilities involved in each process in grass-roots governance applications; the data resource layer is used to collect, store, integrate, and call various types of data; the infrastructure layer is a government affairs cloud network; the security guarantee 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 capacity, and good service quality.

[0006] However, the above patents have great deficiencies in data processing, resulting in problems such as slow extraction speed of keywords in case texts, high error rate of typo replacement, and low efficiency of case pre-trial for case texts. Summary of the Invention

[0007] The object of the present invention is to provide a case filing pre-examination system for a smart governance platform, which can extract keywords 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 the efficiency of case filing pre-examination.

[0008] The present invention utilizes the following technical solutions:

[0009] A case filing pre-examination system for a smart governance platform includes a processing and conversion layer, a matching and filtering layer, a learning and training layer, and an identification and pre-examination layer;

[0010] The processing and conversion layer is used to preprocess case texts, and the preprocessing includes text cleaning, typo replacement, and information extraction;

[0011] The matching and filtering layer is used to perform category matching, length filtering, and vector compilation on the case texts that have completed preprocessing;

[0012] The learning and training layer is used to iteratively train the case texts that have completed matching and filtering to construct a comprehensive data weight matrix;

[0013] The identification and pre-examination layer is used to mask the messy information of the case texts and then import them into the comprehensive data weight matrix for text identification and screening.

[0014] Preferably, the processing and 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 sections in the case texts, and eliminates duplicate case texts according to the location where the event occurred; the typo replacement unit uses a preset recurrent memory network to convert similar-looking characters and / or homophonic characters in the case texts into correct fonts; the information extraction unit uses the TextMap algorithm to extract keywords from the case texts that have completed typo replacement, thereby removing the event occurrence time, event occurrence location, and explanatory statements.

[0015] Preferably, the working process of the TextMap algorithm is as follows:

[0016] a: Convert the case text into an undirected graph, use words as nodes of the undirected graph, and use statements as edges of the undirected graph;

[0017] b: Calculate the weight values and information entropy of statement edges and word nodes respectively according to the occurrence frequencies of statements and words in the case text;

[0018] c: Abstract the obtained weight values into a directed graph, and construct a transition probability to randomly walk in the directed graph according to the information entropy of word nodes and statement edges to generate a global sequence;

[0019] d: Hierarchically encode the global sequence and perform clustering local optimization on the hierarchical encoding to generate a locally optimized result;

[0020] e: Extract keywords from the case text according to the locally optimized result.

[0021] Preferably, the cyclic memory network first sorts the similar-looking characters to form a sequence label; at the same time, the similar-looking characters, homophonic characters, and sequence labels are respectively input into three network branches for feature extraction; among them, the similar-looking characters use the first network branch to obtain the depth feature value of the similar-looking characters; the first network branch includes 1 bidirectional LSTM layer, 4 hidden state layers, 2 fully connected layers, and a PRELU activation function; the homophonic characters use the second network branch to obtain the depth feature value of the homophonic characters; the second network branch includes 2 residual blocks, 5 max pooling layers, 2 fully connected layers, and a LeakyReLU activation function; the sequence label uses the third network branch to extract the depth feature value; the third network branch includes 3 Inception layers, 2 auxiliary classification layers, and a RELU activation function; then, the depth feature values output by the first network branch and the third network branch are respectively spliced and fused using a feature fusion layer, and the fused depth feature values are input into the first branch of the recognition transformer to obtain a form matrix; the first branch includes 1 average pooling layer, 2 fully connected layers, a dropout function, and an ELU activation function to obtain a form matrix; at the same time, the depth feature values output by the second network branch and the third network branch are spliced and fused using a feature fusion layer, and the fused depth feature values are input into the second branch of the recognition transformer to obtain a sound form matrix; the second branch includes 1 LSTM layer, 2 hidden state layers, 1 fully connected layer, and a MISH activation function to obtain a sound form matrix; finally, the transformation process of homophonic characters and similar-looking characters is completed by splicing the sound form matrix and the form matrix.

[0022] Preferably, the matching and filtering layer includes a category matching unit, a text filtering unit, a labeling and partitioning 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 case text after matching according to a character length within 200; the labeling and partitioning unit uses LabeImg software to label the case text according to the functions of each department and the keywords extracted from the case text after filtering and screening 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.

[0023] Preferably, 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; the feature extractor includes 4 3Χ3 convolutional layers with a stride of 2 and 1 batch normalization layer; then the function matrix and the keyword matrix are respectively input into two branches of the backbone network for multiple iterative trainings 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 3Χ3 convolutional layer with a stride of 1, 1 batch normalization layer and a LeakyRelu activation function; the second branch includes 2 Transformer encoding blocks, 1 batch normalization layer, a fusion attention mechanism and a MISH function; finally, the input network output of the test set verifies the comprehensive data weight matrix by inference, so as to optimize the comprehensive data weight matrix; the network output includes 3 normalization layers, 2 fully connected layers and a Softmax activation function.

[0024] Preferably, the identification pre-trial layer includes an information masking unit; the information masking unit uses a masked language network to replace the event occurrence time, event occurrence location and explanatory language in the case text.

[0025] Preferably, the working process of the masked language network is as follows:

[0026] S1: Use a token encoder to segment the sentences of the case text into individual words and encode the words to form a case text encoding;

[0027] S2: Query the taboo encoding according to the preset taboo encoding table for the encoded words; the taboo encoding table includes the event occurrence time, event occurrence location and explanatory language;

[0028] S3: Use a masker to perform bilinear interpolation masking and replacement on the word encodings that meet the taboo encoding.

[0029] Preferably, the identification pre-trial layer further includes a case-filing pre-trial unit; the case-filing pre-trial unit uses a causal language network to input the masked case text encoding into the comprehensive data weight matrix for case classification and filing, and outputs a case-filing pre-trial result.

[0030] Preferably, the working process of the causal language network is as follows:

[0031] A: Strengthen the features of the masked case text encoding through a spatial channel attention mechanism;

[0032] B: Use 4 fully connected layers to import the masked case text encoding with enhanced features into the comprehensive data weight matrix for case classification;

[0033] C: Classify and rank the pre-trial levels of the case texts that have been classified according to the case handling progress of each department.

[0034] The present invention extracts keywords from case texts through the TextMap algorithm, converts misspelled words in the form of similar-shaped characters and / or homophonic characters in the case texts into correct fonts through a recurrent memory network; iteratively trains the case texts that have completed matching and filtering through a learning and training layer to construct a comprehensive data weight matrix; improves the keyword extraction speed of case texts, reduces the error rate of misspelled word replacement, and improves the efficiency of case text filing and pre-trial. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the drawings in the following description are only embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0036] Figure 1 It is a schematic diagram of the case filing and pre-trial system for the intelligent governance platform. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The present invention will be described in detail below with reference to the drawings and embodiments:

[0038] As Figure 1 shown, a case filing and pre-trial system for an intelligent governance platform according to the present invention includes a processing and conversion layer, a matching and filtering layer, a learning and training layer, and an identification and pre-trial layer;

[0039] The processing and conversion layer is used to preprocess case texts, and the preprocessing includes text cleaning, misspelled word replacement, and information extraction;

[0040] In the present invention, the processing and conversion layer includes a text cleaning unit, a misspelled word replacement unit, and an information extraction unit; the text cleaning unit uses regular expressions to remove special characters, punctuation marks, and blank sections in the case text, and eliminates duplicate case texts according to the location where the event occurred; the misspelled word replacement unit uses a preset recurrent memory network to convert misspelled words in the form of similar-shaped characters and / or homophonic characters in the case text into correct fonts; the information extraction unit uses the TextMap algorithm to extract keywords from the case text after misspelled word replacement, thereby removing the event occurrence time, event occurrence location, and explanatory statements;

[0041] In this embodiment, the text cleaning unit mainly includes three steps: removing special symbols, removing redundant blanks, and converting traditional Chinese text to simplified Chinese. Removing special symbols and redundant blanks is to make the feature representation of short texts focus as much as possible on the features and semantics of the short texts themselves, reducing the impact of other symbols on the classification accuracy. It also includes traditional-simplified conversion. That is, converting traditional Chinese text to simplified Chinese is for the convenience of subsequent text vectorization representation. Because the text vectorization representation method used in this article calls an external word vector model, if the words used in the text are not in the vocabulary, the current words will use the initialized vector representation method, changing the semantics of the words themselves. And most traditional Chinese texts are relatively complex and usually exceed the scope of the vocabulary. Converting them to simplified Chinese will neither change their original semantics nor facilitate vectorization representation. Therefore, the step of converting traditional Chinese text to simplified Chinese is very necessary;

[0042] In this embodiment, the text cleaning unit and the information extraction unit combined with the case text have the following characteristics: 1. Specific beginnings and endings; 2. Non-keywords. Among them, non-keywords mainly include the time of event occurrence, the location of event occurrence, and explanatory sentences. Such non-keywords generally follow fixed rules, so all non-keywords can be directly deleted within the entire text range through regular expressions. Because usually these words have no special semantics and will increase the training difficulty and may even affect the experimental results; 3. The text templates used are short. In order to classify the cases from the keywords as soon as possible, the staff will extract the important information and shorten the large-scale case descriptions to within 150 words; 4. Stop words. Words such as 'period' and 'the following' appear too many times. Although they do not 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;

[0043] The matching and filtering layer is used to perform category matching, length filtering, and vector compilation on the preprocessed case text;

[0044] In the present invention, the matching and filtering layer includes a category matching unit, a text filtering unit, a labeling and partitioning 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 case text that has completed the matching according to a character length within 200. The labeling and partitioning 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 the filtering and screening 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;

[0045] In this embodiment, the main responsibility of the text filtering unit is text length filtering; text length filtering is because the main research object is short texts, and short texts are usually in the form of texts with no more than 200 characters; if the text is too long, it will exceed the scope of this study, so such texts are filtered;

[0046] The learning and training layer is used to iteratively train the case text that has completed matching and filtering to construct a comprehensive data weight matrix;

[0047] In the present invention, 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; the feature extractor includes 4 3×3 convolutional layers with a stride of 2 and 1 batch normalization layer; then the function matrix and the keyword matrix are respectively input into two branches of the backbone network for multiple iterative trainings 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 3×3 convolutional layer with a stride of 1, 1 batch normalization layer and a LeakyRelu activation function; the second branch includes 2 Transformer encoding blocks, 1 batch normalization layer, a fusion attention mechanism and a MISH function; finally, the input of the test set to the network outputter performs inference verification on the comprehensive data weight matrix, thereby optimizing the comprehensive data weight matrix; the network outputter includes 3 normalization layers, 2 fully connected layers and a Softmax activation function;

[0048] The recognition and pre-trial layer is used to mask the messy information of the case text and then import it into the comprehensive data weight matrix for text recognition and screening.

[0049] In the present invention, the recognition and pre-trial layer includes an information masking unit and a case-filing pre-trial unit; the information masking unit uses a masked language network to replace the event occurrence time, event occurrence location and explanatory language in the case text; the case-filing pre-trial unit uses a causal language network to encode the masked case text and input it into the comprehensive data weight matrix for classification and case-filing, and outputs the case-filing pre-trial result.

[0050] In this embodiment, the working process of the masked language network is as follows:

[0051] S1: Use a token encoder to segment the sentences of the case text into individual words and encode the words to form a case text encoding;

[0052] S2: Query the taboo encoding according to the pre-set taboo encoding table for the encoded words; the taboo encoding table includes the event occurrence time, event occurrence location and explanatory language;

[0053] S3: Mask and replace the words encoded with the taboo code using bilinear interpolation by a masker;

[0054] In this embodiment, the workflow of the causal language network is as follows:

[0055] A: Enhance the features of the masked case text encoding through the spatial channel attention mechanism;

[0056] B: Use four fully connected layers for the case text encoding with enhanced features and import it into the comprehensive data weight matrix for case classification;

[0057] C: Divide and sort the pre-trial levels of the classified case text according to the case handling progress of each department;

[0058] In the present invention, the regular expression, text matching algorithm, length filtering algorithm, LabeImg software, and global bag-of-words algorithm are all commonly used technologies in the art and will not be elaborated here.

[0059] Furthermore, the workflow of the TextMap algorithm is as follows:

[0060] a: Convert the case text into an undirected graph, use the words as the nodes of the undirected graph, and use the sentences as the edges of the undirected graph;

[0061] b: Calculate the weight values and information entropy of the sentence edges and word nodes respectively according to the occurrence frequencies of the sentences and words in the case text:

[0062] (1)

[0063] Among them, represents the sentence edge weight value, represents the sentence edge serial number, represents the th sentence edge weight value, represents the damping coefficient, represents all the sentence edge sequences in the case text, represents the current sentence edge serial number, represents all the sequences connecting the sentence edges;

[0064] (2)

[0065] Among them, represents the word node weight value, represents the word node serial number, represents the th word node weight value, represents the damping coefficient, represents all the word node sequences in the case text, Indicates the serial number of the current word node, Indicates the sequence of all connecting word nodes;

[0066] (3)

[0067] Among them, Indicates the information entropy of each statement edge, Indicates the Huff function, Indicates the probability set of each statement edge, Indicates the number of statements, Indicates the statement edge probability, Indicates the current statement edge probability;

[0068] (4)

[0069] Among them, Indicates the information entropy of each word node, Indicates the probability set of each word node, Indicates the number of word nodes, Indicates the word node probability, Indicates the current word node probability;

[0070] c: Abstract the obtained weight value into a directed graph, and construct the transition probability to randomly walk within the directed graph according to the information entropy of the word node and the statement edge to generate a global sequence:

[0071] (5)

[0072] Among them, Indicates the position within the directed graph, Indicates the current word node or statement edge, Indicates the transition probability, Indicates the shift symbol, Indicates the shifted word node or statement edge, Indicates the position of the current word node or statement edge within the directed graph, Indicates the position of the shifted word node or statement edge within the directed graph, Indicates the position of the current word node or statement edge after shifting within the directed graph;

[0073] (6)

[0074] Among them, Indicates the global sequence, Indicates the normalization function, Indicates the transition adjustment coefficient;

[0075] d: Hierarchically encode the global sequence and perform clustering local optimization on the hierarchical encoding to generate a local optimization result;

[0076] e: Extract keywords from the case text according to the local optimization result.

[0077] Furthermore, the cyclic memory network first sorts the similar-looking characters to form a sequence label; at the same time, the similar-looking characters, homophonic characters, and sequence labels are respectively input into three network branches for feature extraction; among them, the similar-looking characters use the first network branch to obtain the depth feature value of the similar-looking characters; the first network branch includes 1 bidirectional LSTM layer, 4 hidden state layers, 2 fully connected layers, and a PRELU activation function; the homophonic characters use the second network branch to obtain the depth feature value of the homophonic characters; the second network branch includes 2 residual blocks, 5 max pooling layers, 2 fully connected layers, and a LeakyReLU activation function; the sequence label uses the third network branch to extract the depth feature value; the third network branch includes 3 Inception layers, 2 auxiliary classification layers, and a RELU activation function; then, the depth feature values output by the first network branch and the third network branch are respectively spliced and fused using a feature fusion layer, and the fused depth feature values are input into the first branch of the recognition transformer to obtain a form matrix; the first branch includes 1 average pooling layer, 2 fully connected layers, a dropout function, and an ELU activation function to obtain a form matrix; at the same time, the depth feature values output by the second network branch and the third network branch are spliced and fused using a feature fusion layer, and the fused depth feature values are input into the second branch of the recognition transformer to obtain a phonetic form matrix; the second branch includes 1 LSTM layer, 2 hidden state layers, 1 fully connected layer, and a MISH activation function to obtain a phonetic form matrix; finally, the conversion process of homophonic characters and similar-looking characters is completed by splicing the phonetic form matrix and the form matrix;

[0078] In the present invention, the working principle of the cyclic memory network is as follows: First, the input similar-looking characters and homophonic characters are arranged in a certain order to form a sequence label, which serves as the input for subsequent processing; the similar-looking characters, homophonic 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 the depth feature value; the second network branch uses methods such as residual blocks, max pooling layers, and fully connected layers to obtain the depth feature value of the homophonic characters; the third network branch uses methods such as Inception layers and auxiliary classification layers to extract the depth feature value of the sequence label;

[0079] Then, the depth feature values output by the first network branch and the third network branch are concatenated through a feature fusion layer to obtain a form matrix; the depth feature values output by the second network branch and the third network branch are concatenated through a feature fusion layer to obtain a sound form matrix; meanwhile, the form matrix and the sound form matrix are updated and combined in conjunction with text error correction rules;

[0080] Finally, a certain algorithm, such as an average pooling layer, a fully connected layer, etc., is used to process the depth feature values to obtain the final conversion results corresponding to homophonic words and similar-shaped words;

[0081] In this embodiment, the working principle of the text error correction rule is as follows: The first step is error detection: By using the Jieba Chinese word segmenter to segment words and break characters, since there are misspelled words in the sentence, the word segmenting and character breaking 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, at the 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, it is determined that the probability of this character being a suspected misspelled word is high; at the word granularity: the words that are not in the dictionary after word segmentation are likely to be suspected misspelled words; the second step is error correction; by traversing all suspected error positions and using a homophonic dictionary and a similar-shaped dictionary to replace the words at the error positions, and then calculating the sentence perplexity through the language model, comparing and sorting all candidate set results to obtain the optimal corrected characters.

[0082] In this embodiment, when performing iterative training on the learning and training layer, the optimal comprehensive data weight matrix is obtained by dynamically correcting the learning rate:

[0083] (7)

[0084] Among them, represents the training rate, represents the training time, represents the training rate at the current moment, represents the first network momentum coefficient, represents the weight gradient, represents the second network momentum coefficient;

[0085] (8)

[0086] Among them, represents the adaptive parameter, represents the adaptive parameter at the current moment, represents the average coefficient of the first network momentum and the second network momentum, represents the adaptive parameter at the previous moment;

[0087] (9)

[0088] Among them, represents the learning rate, represents the learning rate at the current moment, represents the decay coefficient, represents the initial learning rate;

[0089] (10)

[0090] Among them, 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.

[0091] Embodiment:

[0092] 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 sections in the case texts, and eliminates duplicate case texts according to the location where the event occurred; the typo replacement unit uses a preset recurrent memory network to first sort the similar-looking characters to form sequence tags; at the same time, the similar-looking characters, homophonic characters, and sequence tags are respectively input into three network branches for feature extraction; among them, the similar-looking characters use the first network branch to obtain the deep feature values of the similar-looking characters; the first network branch includes 1 bidirectional LSTM layer, 4 hidden state layers, 2 fully connected layers, and a PRELU activation function; the homophonic characters use the second network branch to obtain the deep feature values of the homophonic characters; the second network branch includes 2 residual blocks, 5 max pooling layers, 2 fully connected layers, and a LeakyReLU activation function; the sequence tags use the third network branch to extract the deep feature values; the third network branch includes 3 Inception layers, 2 auxiliary classification layers, and a RELU activation function; then the deep feature values output by the first network branch and the third network branch use the feature fusion layer to splice the deep feature values, and then the fused deep feature values are sequentially input into 1 average pooling layer, 2 fully connected layers, a dropout function, and an ELU activation function to obtain a form matrix; at the same time, the deep feature values output by the second network branch and the third network branch use the feature fusion layer to splice the deep feature values, and then the fused deep feature values are sequentially input into 1 LSTM layer, 2 hidden state layers, 1 fully connected layer, and a MISH activation function to obtain a phonetic matrix; finally, the conversion process of homophonic characters and similar-looking characters is completed by splicing the phonetic matrix and the form matrix;

[0093] The information extraction unit uses the TextMap algorithm to convert the case text with misspelling replaced into an undirected graph, taking words as the nodes of the undirected graph and sentences as the edges of the undirected graph; according to the occurrence frequencies 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 probabilities are constructed to randomly walk within the directed graph according to the information entropy of word nodes and sentence edges to generate a global sequence; the global sequence is hierarchically encoded, and the hierarchical encoding is clustered and locally optimized to generate a locally optimized result; keywords in the case text are extracted according to the locally optimized result, thereby removing the event occurrence time, event occurrence location, and explanatory sentences.

[0094] The category matching unit of the matching and filtering 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 case text after matching according to a character length within 200; the annotation and 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 after filtering and screening 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.

[0095] Then, 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; the feature extractor includes 4 3×3 convolutional layers with a stride of 2 and 1 batch normalization layer; then the function matrix and the keyword matrix are respectively input into two branches of the backbone network for multiple iterative trainings 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 3×3 convolutional layer with a stride of 1, 1 batch normalization layer, and a LeakyRelu activation function; the second branch includes 2 Transformer encoding blocks, 1 batch normalization layer, a fusion attention mechanism, and a MISH function; finally, the input of the test set to the network outputter performs inference verification on the comprehensive data weight matrix, thereby optimizing the comprehensive data weight matrix; the network outputter includes 3 normalization layers, 2 fully connected layers, and a Softmax activation function.

[0096] Finally, the information masking unit in the pre-trial layer is used to utilize the masked language network. The statements of the case text are segmented into individual words by the token encoder, and the words are encoded to form the case text encoding. The encoded words are queried for taboo encoding according to the pre-set taboo encoding table. The taboo encoding table includes the event occurrence time, the event occurrence location, and the explanatory language. The words encoded with taboo encoding are masked and replaced by bilinear interpolation using the masker. The case text encoding after masking is feature-enhanced by the case-filing pre-trial unit using the causal language network through the spatial channel attention mechanism. Then, the case text encoding after feature enhancement is imported into the comprehensive data weight matrix for case classification using 4 fully connected layers. Immediately afterwards, the classified case text is divided and sorted according to the case-handling progress of each department to complete the case classification and case-filing pre-trial task.

Claims

1. A case-filing pre-trial system for an intelligent governance platform, characterized in that: It includes a processing conversion layer, a matching and filtering layer, a learning and training layer, and an identification and preliminary review layer; The processing conversion layer is used to preprocess the case text, and the preprocessing includes text cleaning, typo replacement, and information extraction; The matching and filtering layer is used to perform category matching, length filtering, and vector compilation on the case text that has completed preprocessing; The learning and training layer is used to iteratively train the case text that has completed matching and filtering to construct a comprehensive data weight matrix; The identification and preliminary review layer is used to mask the messy information of the case text and then import it into the comprehensive data weight matrix for text identification and screening; The identification and preliminary review layer includes an information masking unit; The information masking unit uses a masked language network to replace the event occurrence time, event occurrence location, and explanatory language in the case text; The identification and preliminary review layer also includes a case-filing preliminary review unit; the case-filing preliminary review unit uses a causal language network to encode and input the masked case text into the comprehensive data weight matrix for case-filing classification and outputs the case-filing preliminary review result; The processing conversion layer includes a text cleaning unit, a typo replacement unit, and an information extraction unit; the typo replacement unit uses a preset recurrent memory network to replace the similar-looking characters and / or homophonic characters in the case text with the correct fonts; the information extraction unit uses the TextMap algorithm to extract keywords from the case text after typo replacement, so as to remove the event occurrence time, event occurrence location, and explanatory statements; the text cleaning unit uses regular expressions to remove special characters, punctuation marks, and blank sections in the case text, and eliminates duplicate case texts according to the event occurrence location; Among them, the working process of the TextMap algorithm is as follows: a: Convert the case text into an undirected graph, use words as the nodes of the undirected graph, and use sentences as the edges of the undirected graph at the same time; b: Calculate the weight values and information entropy of the sentence edges and word nodes respectively according to the occurrence frequencies of sentences and words in the case text; c: Abstract the obtained weight values into a directed graph, and construct a transition probability to randomly walk in the directed graph according to the information entropy of the word nodes and sentence edges respectively to generate a global sequence; d: Hierarchically encode the global sequence and perform clustering local optimization on the hierarchical encoding to generate a local optimization result; e: Extract keywords from the case text according to the local optimization result.

2. The case-filing pre-trial system of the intelligent governance platform according to claim 1, characterized in that: The cyclic memory network first sorts the visually similar characters and phonetically similar characters respectively to form combined sequence tags; at the same time, the visually similar characters, phonetically similar characters and sequence tags are respectively input into three network branches for feature extraction; among them, the visually similar characters use the first network branch to obtain the depth feature values of the visually similar characters; the first network branch includes 1 bidirectional LSTM layer, 4 hidden state layers, 2 fully connected layers and a PRELU activation function; the phonetically similar characters use the second network branch to obtain the depth feature values of the phonetically similar characters; the second network branch includes 2 residual blocks, 5 max pooling layers, 2 fully connected layers and a LeakyReLU activation function; the sequence tags use the third network branch to extract the depth feature values; the third network branch includes 3 Inception layers, 2 auxiliary classification layers and a RELU activation function; then the depth feature values output by the first network branch and the third network branch are used to splice the depth feature values by the feature fusion layer, and then the fused depth feature values are successively input into 1 average pooling layer, 2 fully connected layers, a dropout function and an ELU activation function to obtain a form matrix; at the same time, the depth feature values output by the second network branch and the third network branch are used to splice the depth feature values by the feature fusion layer, and then the fused depth feature values are successively input into 1 LSTM layer, 2 hidden state layers, 1 fully connected layer and a MISH activation function to obtain a phonetic form matrix; finally, the conversion process of phonetically similar characters and visually similar characters is completed by splicing the phonetic form matrix and the form matrix.

3. The case-filing pre-trial system of the intelligent governance platform according to claim 1, characterized in that: The matching and 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 within 200 characters in length; 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 filtered and screened case text 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.

4. The case-filing pre-trial system of the intelligent governance platform according to claim 1, characterized in that: The learning and training layer first inputs the vector matrix of the training set into a 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 the keyword matrix are respectively input into two branches of the backbone network for multiple iterative trainings to obtain a comprehensive data weight matrix; finally, the input network output of the test set verifies the comprehensive data weight matrix by inference, so as to optimize the comprehensive data weight matrix.

5. The case-filing pre-trial system of the intelligent governance platform according to claim 1, characterized in that: The working process of the masked language network is as follows: S1: Use a token encoder to segment the sentences of the case text into individual words and encode the words to form a case text encoding; S2: Query the taboo encoding according to the pre-set taboo encoding table for the encoded words; the taboo encoding table includes the event occurrence time, the event occurrence location and the explanatory language; S3: Use a masker to perform bilinear interpolation masking and replacement on the word encodings that meet the taboo encoding.

6. The case-filing pre-trial system of the intelligent governance platform according to claim 1, characterized in that: The working process of the causal language network is as follows: A: The case text with masking completed is encoded and feature-enhanced through the spatial channel attention mechanism; B: The case text encoding with feature enhancement completed is imported into the comprehensive data weight matrix for case classification by using 4 fully connected layers; C: The classified case text is pre-reviewed, ranked, and sorted according to the case handling progress of each department.

Images